| ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 |
| |
| ; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s |
| ; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=VI %s |
| ; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s |
| ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s |
| |
| define <16 x float> @bitcast_v16i32_to_v16f32(<16 x i32> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16i32_to_v16f32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB0_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; GCN-NEXT: .LBB0_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16i32_to_v16f32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB0_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15 |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13 |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11 |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9 |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7 |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5 |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3 |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1 |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0 |
| ; VI-NEXT: .LBB0_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16i32_to_v16f32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB0_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_u32_e32 v15, 3, v15 |
| ; GFX9-NEXT: v_add_u32_e32 v14, 3, v14 |
| ; GFX9-NEXT: v_add_u32_e32 v13, 3, v13 |
| ; GFX9-NEXT: v_add_u32_e32 v12, 3, v12 |
| ; GFX9-NEXT: v_add_u32_e32 v11, 3, v11 |
| ; GFX9-NEXT: v_add_u32_e32 v10, 3, v10 |
| ; GFX9-NEXT: v_add_u32_e32 v9, 3, v9 |
| ; GFX9-NEXT: v_add_u32_e32 v8, 3, v8 |
| ; GFX9-NEXT: v_add_u32_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_add_u32_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_add_u32_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_add_u32_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_add_u32_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_add_u32_e32 v2, 3, v2 |
| ; GFX9-NEXT: v_add_u32_e32 v1, 3, v1 |
| ; GFX9-NEXT: v_add_u32_e32 v0, 3, v0 |
| ; GFX9-NEXT: .LBB0_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16i32_to_v16f32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB0_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0 |
| ; GFX11-NEXT: .LBB0_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <16 x i32> %a, splat (i32 3) |
| %a2 = bitcast <16 x i32> %a1 to <16 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x i32> %a to <16 x float> |
| br label %end |
| |
| end: |
| %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x float> %phi |
| } |
| |
| define <16 x i32> @bitcast_v16f32_to_v16i32(<16 x float> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16f32_to_v16i32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB1_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; GCN-NEXT: .LBB1_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16f32_to_v16i32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB1_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; VI-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; VI-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; VI-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; VI-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; VI-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; VI-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; VI-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; VI-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; VI-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; VI-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; VI-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; VI-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; VI-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; VI-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; VI-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; VI-NEXT: .LBB1_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16f32_to_v16i32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB1_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; GFX9-NEXT: .LBB1_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16f32_to_v16i32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB1_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14 |
| ; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12 |
| ; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10 |
| ; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8 |
| ; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6 |
| ; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4 |
| ; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2 |
| ; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0 |
| ; GFX11-NEXT: .LBB1_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <16 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <16 x float> %a1 to <16 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x float> %a to <16 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x i32> %phi |
| } |
| |
| define <8 x i64> @bitcast_v16i32_to_v8i64(<16 x i32> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16i32_to_v8i64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB2_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; GCN-NEXT: .LBB2_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16i32_to_v8i64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB2_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15 |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13 |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11 |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9 |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7 |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5 |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3 |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1 |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0 |
| ; VI-NEXT: .LBB2_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16i32_to_v8i64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB2_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_u32_e32 v15, 3, v15 |
| ; GFX9-NEXT: v_add_u32_e32 v14, 3, v14 |
| ; GFX9-NEXT: v_add_u32_e32 v13, 3, v13 |
| ; GFX9-NEXT: v_add_u32_e32 v12, 3, v12 |
| ; GFX9-NEXT: v_add_u32_e32 v11, 3, v11 |
| ; GFX9-NEXT: v_add_u32_e32 v10, 3, v10 |
| ; GFX9-NEXT: v_add_u32_e32 v9, 3, v9 |
| ; GFX9-NEXT: v_add_u32_e32 v8, 3, v8 |
| ; GFX9-NEXT: v_add_u32_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_add_u32_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_add_u32_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_add_u32_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_add_u32_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_add_u32_e32 v2, 3, v2 |
| ; GFX9-NEXT: v_add_u32_e32 v1, 3, v1 |
| ; GFX9-NEXT: v_add_u32_e32 v0, 3, v0 |
| ; GFX9-NEXT: .LBB2_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16i32_to_v8i64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB2_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0 |
| ; GFX11-NEXT: .LBB2_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <16 x i32> %a, splat (i32 3) |
| %a2 = bitcast <16 x i32> %a1 to <8 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x i32> %a to <8 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x i64> %phi |
| } |
| |
| define <16 x i32> @bitcast_v8i64_to_v16i32(<8 x i64> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8i64_to_v16i32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB3_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; GCN-NEXT: .LBB3_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8i64_to_v16i32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB3_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0 |
| ; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; VI-NEXT: .LBB3_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8i64_to_v16i32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB3_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc |
| ; GFX9-NEXT: .LBB3_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8i64_to_v16i32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB3_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo |
| ; GFX11-NEXT: .LBB3_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <8 x i64> %a, splat (i64 3) |
| %a2 = bitcast <8 x i64> %a1 to <16 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x i64> %a to <16 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x i32> %phi |
| } |
| |
| define <8 x double> @bitcast_v16i32_to_v8f64(<16 x i32> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16i32_to_v8f64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB4_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; GCN-NEXT: .LBB4_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16i32_to_v8f64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB4_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15 |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13 |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11 |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9 |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7 |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5 |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3 |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1 |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0 |
| ; VI-NEXT: .LBB4_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16i32_to_v8f64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB4_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_u32_e32 v15, 3, v15 |
| ; GFX9-NEXT: v_add_u32_e32 v14, 3, v14 |
| ; GFX9-NEXT: v_add_u32_e32 v13, 3, v13 |
| ; GFX9-NEXT: v_add_u32_e32 v12, 3, v12 |
| ; GFX9-NEXT: v_add_u32_e32 v11, 3, v11 |
| ; GFX9-NEXT: v_add_u32_e32 v10, 3, v10 |
| ; GFX9-NEXT: v_add_u32_e32 v9, 3, v9 |
| ; GFX9-NEXT: v_add_u32_e32 v8, 3, v8 |
| ; GFX9-NEXT: v_add_u32_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_add_u32_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_add_u32_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_add_u32_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_add_u32_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_add_u32_e32 v2, 3, v2 |
| ; GFX9-NEXT: v_add_u32_e32 v1, 3, v1 |
| ; GFX9-NEXT: v_add_u32_e32 v0, 3, v0 |
| ; GFX9-NEXT: .LBB4_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16i32_to_v8f64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB4_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0 |
| ; GFX11-NEXT: .LBB4_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <16 x i32> %a, splat (i32 3) |
| %a2 = bitcast <16 x i32> %a1 to <8 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x i32> %a to <8 x double> |
| br label %end |
| |
| end: |
| %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x double> %phi |
| } |
| |
| define <16 x i32> @bitcast_v8f64_to_v16i32(<8 x double> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8f64_to_v16i32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB5_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GCN-NEXT: .LBB5_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8f64_to_v16i32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB5_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; VI-NEXT: .LBB5_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8f64_to_v16i32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB5_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX9-NEXT: .LBB5_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8f64_to_v16i32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB5_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX11-NEXT: .LBB5_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <8 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <8 x double> %a1 to <16 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x double> %a to <16 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x i32> %phi |
| } |
| |
| define <32 x i16> @bitcast_v16i32_to_v32i16(<16 x i32> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16i32_to_v32i16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v30, v15 |
| ; GCN-NEXT: v_mov_b32_e32 v28, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v26, v13 |
| ; GCN-NEXT: v_mov_b32_e32 v24, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v22, v11 |
| ; GCN-NEXT: v_mov_b32_e32 v20, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v18, v9 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v14, v7 |
| ; GCN-NEXT: v_mov_b32_e32 v12, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v10, v5 |
| ; GCN-NEXT: v_mov_b32_e32 v8, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v6, v3 |
| ; GCN-NEXT: v_mov_b32_e32 v4, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v2, v1 |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB6_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_alignbit_b32 v29, v30, v28, 16 |
| ; GCN-NEXT: v_alignbit_b32 v25, v26, v24, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v22, v20, 16 |
| ; GCN-NEXT: v_alignbit_b32 v17, v18, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v10, v8, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v26 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v18 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2 |
| ; GCN-NEXT: .LBB6_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB6_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20 |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v26 |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24 |
| ; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30 |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28 |
| ; GCN-NEXT: v_alignbit_b32 v29, v30, v28, 16 |
| ; GCN-NEXT: v_alignbit_b32 v25, v26, v24, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v22, v20, 16 |
| ; GCN-NEXT: v_alignbit_b32 v17, v18, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v10, v8, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v26 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v18 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2 |
| ; GCN-NEXT: .LBB6_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_mov_b32_e32 v16, v32 |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16i32_to_v32i16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB6_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15 |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13 |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11 |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9 |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7 |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5 |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3 |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1 |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0 |
| ; VI-NEXT: .LBB6_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16i32_to_v32i16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB6_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_u32_e32 v15, 3, v15 |
| ; GFX9-NEXT: v_add_u32_e32 v14, 3, v14 |
| ; GFX9-NEXT: v_add_u32_e32 v13, 3, v13 |
| ; GFX9-NEXT: v_add_u32_e32 v12, 3, v12 |
| ; GFX9-NEXT: v_add_u32_e32 v11, 3, v11 |
| ; GFX9-NEXT: v_add_u32_e32 v10, 3, v10 |
| ; GFX9-NEXT: v_add_u32_e32 v9, 3, v9 |
| ; GFX9-NEXT: v_add_u32_e32 v8, 3, v8 |
| ; GFX9-NEXT: v_add_u32_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_add_u32_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_add_u32_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_add_u32_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_add_u32_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_add_u32_e32 v2, 3, v2 |
| ; GFX9-NEXT: v_add_u32_e32 v1, 3, v1 |
| ; GFX9-NEXT: v_add_u32_e32 v0, 3, v0 |
| ; GFX9-NEXT: .LBB6_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16i32_to_v32i16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB6_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0 |
| ; GFX11-NEXT: .LBB6_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <16 x i32> %a, splat (i32 3) |
| %a2 = bitcast <16 x i32> %a1 to <32 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x i32> %a to <32 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x i16> %phi |
| } |
| |
| define <16 x i32> @bitcast_v32i16_to_v16i32(<32 x i16> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32i16_to_v16i32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v38, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v31, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v0 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB7_3 |
| ; GCN-NEXT: ; %bb.1: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB7_4 |
| ; GCN-NEXT: .LBB7_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; GCN-NEXT: .LBB7_3: ; %cmp.false |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v31 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v32 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v54 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v55 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v33 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v34 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v35 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v36 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v37 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v38 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v30 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v39 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v48 |
| ; GCN-NEXT: v_or_b32_e32 v4, v4, v49 |
| ; GCN-NEXT: v_or_b32_e32 v5, v5, v50 |
| ; GCN-NEXT: v_or_b32_e32 v6, v6, v51 |
| ; GCN-NEXT: v_or_b32_e32 v7, v7, v52 |
| ; GCN-NEXT: v_or_b32_e32 v8, v8, v17 |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v19 |
| ; GCN-NEXT: v_or_b32_e32 v10, v10, v21 |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v23 |
| ; GCN-NEXT: v_or_b32_e32 v12, v12, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v27 |
| ; GCN-NEXT: v_or_b32_e32 v14, v14, v29 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v53 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB7_2 |
| ; GCN-NEXT: .LBB7_4: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31 |
| ; GCN-NEXT: s_mov_b32 s6, 0x30000 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_or_b32_e32 v0, v54, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v55, v1 |
| ; GCN-NEXT: v_or_b32_e32 v2, v39, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v48, v3 |
| ; GCN-NEXT: v_or_b32_e32 v4, v49, v4 |
| ; GCN-NEXT: v_or_b32_e32 v5, v50, v5 |
| ; GCN-NEXT: v_or_b32_e32 v6, v51, v6 |
| ; GCN-NEXT: v_or_b32_e32 v7, v52, v7 |
| ; GCN-NEXT: v_or_b32_e32 v8, v17, v8 |
| ; GCN-NEXT: v_or_b32_e32 v9, v19, v9 |
| ; GCN-NEXT: v_or_b32_e32 v10, v21, v10 |
| ; GCN-NEXT: v_or_b32_e32 v11, v23, v11 |
| ; GCN-NEXT: v_or_b32_e32 v12, v25, v12 |
| ; GCN-NEXT: v_or_b32_e32 v13, v27, v13 |
| ; GCN-NEXT: v_or_b32_e32 v14, v29, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v53, v15 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v13 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v14 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x30000, v15 |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32i16_to_v16i32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB7_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v17, 3 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v15 |
| ; VI-NEXT: v_add_u16_sdwa v15, v15, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v14 |
| ; VI-NEXT: v_add_u16_sdwa v14, v14, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v14, v16, v14 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v13 |
| ; VI-NEXT: v_add_u16_sdwa v13, v13, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v13, v16, v13 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v12 |
| ; VI-NEXT: v_add_u16_sdwa v12, v12, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v12, v16, v12 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v11 |
| ; VI-NEXT: v_add_u16_sdwa v11, v11, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v11, v16, v11 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v10 |
| ; VI-NEXT: v_add_u16_sdwa v10, v10, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v10, v16, v10 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v9 |
| ; VI-NEXT: v_add_u16_sdwa v9, v9, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v9, v16, v9 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v8 |
| ; VI-NEXT: v_add_u16_sdwa v8, v8, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v8, v16, v8 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v7 |
| ; VI-NEXT: v_add_u16_sdwa v7, v7, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v7, v16, v7 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v6 |
| ; VI-NEXT: v_add_u16_sdwa v6, v6, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v6, v16, v6 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v5 |
| ; VI-NEXT: v_add_u16_sdwa v5, v5, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v5, v16, v5 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v4 |
| ; VI-NEXT: v_add_u16_sdwa v4, v4, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v4, v16, v4 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v3 |
| ; VI-NEXT: v_add_u16_sdwa v3, v3, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v3, v16, v3 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v2 |
| ; VI-NEXT: v_add_u16_sdwa v2, v2, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v2, v16, v2 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v1 |
| ; VI-NEXT: v_add_u16_sdwa v1, v1, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v1, v16, v1 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v0 |
| ; VI-NEXT: v_add_u16_sdwa v0, v0, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v0, v16, v0 |
| ; VI-NEXT: .LBB7_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32i16_to_v16i32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB7_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB7_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32i16_to_v16i32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB7_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: .LBB7_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <32 x i16> %a, splat (i16 3) |
| %a2 = bitcast <32 x i16> %a1 to <16 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x i16> %a to <16 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x i32> %phi |
| } |
| |
| define <32 x half> @bitcast_v16i32_to_v32f16(<16 x i32> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16i32_to_v32f16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_mov_b32_e32 v33, v15 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v13 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v11 |
| ; GCN-NEXT: v_mov_b32_e32 v38, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v39, v9 |
| ; GCN-NEXT: v_mov_b32_e32 v48, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v49, v7 |
| ; GCN-NEXT: v_mov_b32_e32 v50, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v51, v5 |
| ; GCN-NEXT: v_mov_b32_e32 v52, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v53, v3 |
| ; GCN-NEXT: v_mov_b32_e32 v54, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v55, v1 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v0 |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB8_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v34 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v35 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v36 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v37 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v38 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v39 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v48 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v49 |
| ; GCN-NEXT: s_waitcnt expcnt(6) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v50 |
| ; GCN-NEXT: s_waitcnt expcnt(5) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v51 |
| ; GCN-NEXT: s_waitcnt expcnt(4) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v52 |
| ; GCN-NEXT: s_waitcnt expcnt(3) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v53 |
| ; GCN-NEXT: s_waitcnt expcnt(2) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v54 |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v55 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v53 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v55 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v40 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v41 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v42 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v43 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v44 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v45 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v46 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v32 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: .LBB8_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB8_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v32 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v55 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v54 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v53 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v52 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v51 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v50 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v49 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v48 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v39 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v37 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v36 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v35 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v34 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v33 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v0 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v1 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v7 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v9 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v14 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v12 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v10 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v4 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v32 |
| ; GCN-NEXT: .LBB8_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16i32_to_v32f16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB8_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15 |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13 |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11 |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9 |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7 |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5 |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3 |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1 |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0 |
| ; VI-NEXT: .LBB8_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16i32_to_v32f16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB8_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_u32_e32 v15, 3, v15 |
| ; GFX9-NEXT: v_add_u32_e32 v14, 3, v14 |
| ; GFX9-NEXT: v_add_u32_e32 v13, 3, v13 |
| ; GFX9-NEXT: v_add_u32_e32 v12, 3, v12 |
| ; GFX9-NEXT: v_add_u32_e32 v11, 3, v11 |
| ; GFX9-NEXT: v_add_u32_e32 v10, 3, v10 |
| ; GFX9-NEXT: v_add_u32_e32 v9, 3, v9 |
| ; GFX9-NEXT: v_add_u32_e32 v8, 3, v8 |
| ; GFX9-NEXT: v_add_u32_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_add_u32_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_add_u32_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_add_u32_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_add_u32_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_add_u32_e32 v2, 3, v2 |
| ; GFX9-NEXT: v_add_u32_e32 v1, 3, v1 |
| ; GFX9-NEXT: v_add_u32_e32 v0, 3, v0 |
| ; GFX9-NEXT: .LBB8_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16i32_to_v32f16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB8_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0 |
| ; GFX11-NEXT: .LBB8_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <16 x i32> %a, splat (i32 3) |
| %a2 = bitcast <16 x i32> %a1 to <32 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x i32> %a to <32 x half> |
| br label %end |
| |
| end: |
| %phi = phi <32 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x half> %phi |
| } |
| |
| define <16 x i32> @bitcast_v32f16_to_v16i32(<32 x half> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32f16_to_v16i32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v45, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v44, v0 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v43, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v42, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v41, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v52, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v40, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v50, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v55, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v48, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v54, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v38, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v53, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v36, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v51, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v34, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v49, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v33, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v39, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v32, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v37, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v31, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v35, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v23, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v28 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v20, v46 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v30 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB9_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v43 |
| ; GCN-NEXT: v_or_b32_e32 v0, v44, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v42, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v40 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v20 |
| ; GCN-NEXT: v_or_b32_e32 v2, v52, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v50, v3 |
| ; GCN-NEXT: v_or_b32_e32 v4, v48, v4 |
| ; GCN-NEXT: v_or_b32_e32 v5, v38, v5 |
| ; GCN-NEXT: v_or_b32_e32 v6, v36, v6 |
| ; GCN-NEXT: v_or_b32_e32 v7, v34, v7 |
| ; GCN-NEXT: v_or_b32_e32 v8, v33, v8 |
| ; GCN-NEXT: v_or_b32_e32 v9, v32, v9 |
| ; GCN-NEXT: v_or_b32_e32 v10, v31, v10 |
| ; GCN-NEXT: v_or_b32_e32 v11, v21, v11 |
| ; GCN-NEXT: v_or_b32_e32 v12, v19, v12 |
| ; GCN-NEXT: v_or_b32_e32 v13, v18, v13 |
| ; GCN-NEXT: v_or_b32_e32 v14, v17, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: .LBB9_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB9_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v45 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v44 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v43 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v42 |
| ; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GCN-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v3, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v41 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v40 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v55 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v53 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v31 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GCN-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v5, v4 |
| ; GCN-NEXT: v_or_b32_e32 v4, v7, v6 |
| ; GCN-NEXT: v_or_b32_e32 v5, v9, v8 |
| ; GCN-NEXT: v_or_b32_e32 v6, v11, v10 |
| ; GCN-NEXT: v_or_b32_e32 v7, v13, v12 |
| ; GCN-NEXT: v_or_b32_e32 v8, v15, v14 |
| ; GCN-NEXT: v_or_b32_e32 v9, v26, v24 |
| ; GCN-NEXT: v_or_b32_e32 v10, v28, v27 |
| ; GCN-NEXT: v_or_b32_e32 v11, v21, v29 |
| ; GCN-NEXT: v_or_b32_e32 v12, v19, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v18, v23 |
| ; GCN-NEXT: v_or_b32_e32 v14, v17, v22 |
| ; GCN-NEXT: v_or_b32_e32 v15, v16, v20 |
| ; GCN-NEXT: .LBB9_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32f16_to_v16i32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB9_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v16, 0x200 |
| ; VI-NEXT: v_add_f16_sdwa v17, v15, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v15, 0x200, v15 |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v14, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v14, 0x200, v14 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v13, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v13, 0x200, v13 |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v12, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v12, 0x200, v12 |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v11, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v11, 0x200, v11 |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v10, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v10, 0x200, v10 |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v9, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v9, 0x200, v9 |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v8, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v8, 0x200, v8 |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v7, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v7, 0x200, v7 |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v6, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v6, 0x200, v6 |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v5, 0x200, v5 |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v4, 0x200, v4 |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, 0x200, v3 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v2, 0x200, v2 |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v1, 0x200, v1 |
| ; VI-NEXT: v_add_f16_sdwa v16, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v0, 0x200, v0 |
| ; VI-NEXT: v_or_b32_e32 v1, v1, v17 |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v16 |
| ; VI-NEXT: .LBB9_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32f16_to_v16i32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB9_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: s_movk_i32 s6, 0x200 |
| ; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB9_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32f16_to_v16i32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB9_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-NEXT: .LBB9_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <32 x half> %a1 to <16 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x half> %a to <16 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x i32> %phi |
| } |
| |
| define <32 x bfloat> @bitcast_v16i32_to_v32bf16(<16 x i32> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16i32_to_v32bf16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v55, v15 |
| ; GCN-NEXT: v_mov_b32_e32 v54, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v53, v13 |
| ; GCN-NEXT: v_mov_b32_e32 v52, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v51, v11 |
| ; GCN-NEXT: v_mov_b32_e32 v50, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v49, v9 |
| ; GCN-NEXT: v_mov_b32_e32 v48, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v39, v7 |
| ; GCN-NEXT: v_mov_b32_e32 v38, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v5 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v3 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v1 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v0 |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB10_3 |
| ; GCN-NEXT: ; %bb.1: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB10_4 |
| ; GCN-NEXT: .LBB10_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; GCN-NEXT: .LBB10_3: ; %cmp.false |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v55 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v54 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v53 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v52 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v51 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v50 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v49 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v48 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v48 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v39 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v38 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v36 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v36 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v35 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v34 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v34 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v33 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v32 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB10_2 |
| ; GCN-NEXT: .LBB10_4: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v32 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v33 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v34 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v35 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v36 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v37 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v39 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v48 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v49 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v50 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v51 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v52 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v53 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v54 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v55 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v15 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v14 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v13 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v12 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v11 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v10 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v9 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v7 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v6 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v5 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v4 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16i32_to_v32bf16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB10_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15 |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13 |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11 |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9 |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7 |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5 |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3 |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1 |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0 |
| ; VI-NEXT: .LBB10_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16i32_to_v32bf16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB10_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_u32_e32 v15, 3, v15 |
| ; GFX9-NEXT: v_add_u32_e32 v14, 3, v14 |
| ; GFX9-NEXT: v_add_u32_e32 v13, 3, v13 |
| ; GFX9-NEXT: v_add_u32_e32 v12, 3, v12 |
| ; GFX9-NEXT: v_add_u32_e32 v11, 3, v11 |
| ; GFX9-NEXT: v_add_u32_e32 v10, 3, v10 |
| ; GFX9-NEXT: v_add_u32_e32 v9, 3, v9 |
| ; GFX9-NEXT: v_add_u32_e32 v8, 3, v8 |
| ; GFX9-NEXT: v_add_u32_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_add_u32_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_add_u32_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_add_u32_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_add_u32_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_add_u32_e32 v2, 3, v2 |
| ; GFX9-NEXT: v_add_u32_e32 v1, 3, v1 |
| ; GFX9-NEXT: v_add_u32_e32 v0, 3, v0 |
| ; GFX9-NEXT: .LBB10_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16i32_to_v32bf16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB10_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0 |
| ; GFX11-NEXT: .LBB10_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <16 x i32> %a, splat (i32 3) |
| %a2 = bitcast <16 x i32> %a1 to <32 x bfloat> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x i32> %a to <32 x bfloat> |
| br label %end |
| |
| end: |
| %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x bfloat> %phi |
| } |
| |
| define <16 x i32> @bitcast_v32bf16_to_v16i32(<32 x bfloat> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32bf16_to_v16i32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v1 |
| ; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v0 |
| ; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v3 |
| ; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v2 |
| ; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v5 |
| ; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v4 |
| ; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v7 |
| ; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v6 |
| ; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v9 |
| ; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v8 |
| ; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v11 |
| ; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v10 |
| ; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v13 |
| ; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v12 |
| ; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v15 |
| ; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v14 |
| ; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v17 |
| ; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v16 |
| ; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v19 |
| ; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v18 |
| ; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v21 |
| ; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v20 |
| ; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v23 |
| ; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v22 |
| ; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25 |
| ; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v24 |
| ; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v27 |
| ; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v26 |
| ; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v29 |
| ; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v28 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47 |
| ; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v46 |
| ; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v30 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB11_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v44 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v42 |
| ; GCN-NEXT: v_alignbit_b32 v0, v0, v45, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v1, v43, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v41 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v40 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v55 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v54 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v52 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v50 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v48 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v38 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v35 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v25 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v23 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v20 |
| ; GCN-NEXT: v_alignbit_b32 v2, v2, v51, 16 |
| ; GCN-NEXT: v_alignbit_b32 v3, v3, v49, 16 |
| ; GCN-NEXT: v_alignbit_b32 v4, v4, v39, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v5, v37, 16 |
| ; GCN-NEXT: v_alignbit_b32 v6, v6, v36, 16 |
| ; GCN-NEXT: v_alignbit_b32 v7, v7, v34, 16 |
| ; GCN-NEXT: v_alignbit_b32 v8, v8, v33, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v9, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v10, v10, v31, 16 |
| ; GCN-NEXT: v_alignbit_b32 v11, v11, v21, 16 |
| ; GCN-NEXT: v_alignbit_b32 v12, v12, v19, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v13, v18, 16 |
| ; GCN-NEXT: v_alignbit_b32 v14, v14, v17, 16 |
| ; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: .LBB11_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB11_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v45 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v44 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v43 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v42 |
| ; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v51 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v41 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v49 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v40 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v39 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v55 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v37 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v54 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v36 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v53 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v34 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v52 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v33 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v50 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v32 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v48 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v31 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v38 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v35 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20 |
| ; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16 |
| ; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16 |
| ; GCN-NEXT: v_alignbit_b32 v4, v7, v6, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v9, v8, 16 |
| ; GCN-NEXT: v_alignbit_b32 v6, v11, v10, 16 |
| ; GCN-NEXT: v_alignbit_b32 v7, v13, v12, 16 |
| ; GCN-NEXT: v_alignbit_b32 v8, v15, v14, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v26, v24, 16 |
| ; GCN-NEXT: v_alignbit_b32 v10, v28, v27, 16 |
| ; GCN-NEXT: v_alignbit_b32 v11, v29, v21, 16 |
| ; GCN-NEXT: v_alignbit_b32 v12, v25, v19, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v23, v18, 16 |
| ; GCN-NEXT: v_alignbit_b32 v14, v22, v17, 16 |
| ; GCN-NEXT: v_alignbit_b32 v15, v20, v16, 16 |
| ; GCN-NEXT: .LBB11_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32bf16_to_v16i32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB11_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v15 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17 |
| ; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v15, 16, 1 |
| ; VI-NEXT: s_movk_i32 s6, 0x7fff |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v15 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v15 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15 |
| ; VI-NEXT: v_alignbit_b32 v15, v15, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v14 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v14, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v14 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v14 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14 |
| ; VI-NEXT: v_alignbit_b32 v14, v14, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v13 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v13, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v13 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v13 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13 |
| ; VI-NEXT: v_alignbit_b32 v13, v13, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v12 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v12, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v12 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v12 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12 |
| ; VI-NEXT: v_alignbit_b32 v12, v12, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v11 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v11, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v11 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v11 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11 |
| ; VI-NEXT: v_alignbit_b32 v11, v11, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v10 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v10, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v10 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v10 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10 |
| ; VI-NEXT: v_alignbit_b32 v10, v10, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v9 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v9, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v9 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v9 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9 |
| ; VI-NEXT: v_alignbit_b32 v9, v9, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v8, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v8 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v8 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8 |
| ; VI-NEXT: v_alignbit_b32 v8, v8, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v7 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v7, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v7 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v7 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7 |
| ; VI-NEXT: v_alignbit_b32 v7, v7, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v6 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v6, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v6 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v6 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6 |
| ; VI-NEXT: v_alignbit_b32 v6, v6, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v5 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v5, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v5 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v5 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5 |
| ; VI-NEXT: v_alignbit_b32 v5, v5, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v4 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v4, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v4 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v4 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4 |
| ; VI-NEXT: v_alignbit_b32 v4, v4, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v3 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v3, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v3 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v3 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; VI-NEXT: v_alignbit_b32 v3, v3, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v2 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v2, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v2 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v2 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 |
| ; VI-NEXT: v_alignbit_b32 v2, v2, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v1 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v1, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v1 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1 |
| ; VI-NEXT: v_alignbit_b32 v1, v1, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v0, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v0 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v0 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0 |
| ; VI-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 |
| ; VI-NEXT: v_alignbit_b32 v0, v0, v16, 16 |
| ; VI-NEXT: .LBB11_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32bf16_to_v16i32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB11_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: s_movk_i32 s6, 0x7fff |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v15, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v15, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc |
| ; GFX9-NEXT: s_mov_b32 s7, 0x7060302 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v14, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v14, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v14 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v13, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v13, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v13 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v12, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v12, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v12 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v11, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v11, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v11 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v11, v11, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v10, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v10, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v10 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v10, v10, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v9, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v9, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v9 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v9, v9, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v8, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v8, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v8 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v8, v8, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v7, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v7, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v7 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v7, v7, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v6, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v6, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v6 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v6, v6, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v5, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v5, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v5 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v5, v5, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v4, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v4, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v4 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v4, v4, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v3, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v3, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v3 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v3, v3, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v2, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v2, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v2 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v2, v2, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v1, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v1, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v1 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v1, v1, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v0, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v0, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v0 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v0, v0, v16, s7 |
| ; GFX9-NEXT: .LBB11_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32bf16_to_v16i32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB11_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16 |
| ; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v18, v16, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v16 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_bfe_u32 v23, v14, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v17, 0x7fff |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v16, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v18, v19, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v19, v23, v14, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v17 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX11-NEXT: v_bfe_u32 v20, v15, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v15 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v20, v20, v15, 0x7fff |
| ; GFX11-NEXT: v_dual_cndmask_b32 v15, v20, v22 :: v_dual_lshlrev_b32 v20, 16, v13 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_perm_b32 v15, v15, v16, 0x7060302 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v17, v21, v18 :: v_dual_add_f32 v18, 0x40c00000, v20 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v14 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v13 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v14, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v12 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_perm_b32 v14, v14, v17, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v17, v13, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_cndmask_b32 v16, v16, v20 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v13, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v11 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_perm_b32 v13, v13, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v21 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX11-NEXT: v_bfe_u32 v18, v12, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v12 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v12, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v12, v18, v22, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v12, v12, v17, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v17, v11, 16, 1 |
| ; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v11, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v19, v10, 16, 1 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v19, v19, v10, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v11, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v10 |
| ; GFX11-NEXT: v_perm_b32 v11, v11, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v10, v19, v22 :: v_dual_lshlrev_b32 v21, 16, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_lshlrev_b32 v19, 16, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX11-NEXT: v_perm_b32 v10, v10, v17, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_add_f32 v8, 0x40c00000, v8 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v18, v8, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v9, 0x40c00000, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v8, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v17, v9, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v9 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v9, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v9, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v8 |
| ; GFX11-NEXT: v_perm_b32 v9, v9, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v8, v18, v22, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v7 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v8, v8, v17, 0x7060302 |
| ; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v19, 0x40c00000, v21 |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GFX11-NEXT: v_bfe_u32 v19, v6, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v7, 0x40c00000, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_add3_u32 v19, v19, v6, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v17, v7, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v7 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v7, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v7, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v5 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v6 |
| ; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_cndmask_b32 v17, v17, v20 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v20, v18, 16, 1 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v6, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v4 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX11-NEXT: v_perm_b32 v7, v7, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v6, v6, v17, 0x7060302 |
| ; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v19 |
| ; GFX11-NEXT: v_add3_u32 v19, v20, v18, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GFX11-NEXT: v_bfe_u32 v22, v17, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v16, v5, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v5 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v5, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v5, v16, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v16, v22, v17, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: v_perm_b32 v5, v5, v18, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v18, v4, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v19, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v4 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v4, 0x7fff |
| ; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v20 :: v_dual_lshlrev_b32 v20, 16, v2 |
| ; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v4, v18, v19, vcc_lo |
| ; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: v_perm_b32 v4, v4, v16, 0x7060302 |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v17, 0x7fff |
| ; GFX11-NEXT: v_bfe_u32 v21, v3, 16, 1 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v18 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v3, 0x7fff |
| ; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v3 |
| ; GFX11-NEXT: v_add3_u32 v21, v22, v18, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v1 |
| ; GFX11-NEXT: v_bfe_u32 v24, v2, 16, 1 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v22 |
| ; GFX11-NEXT: v_add3_u32 v20, v24, v2, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_perm_b32 v3, v3, v17, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v18, v21, v23, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v2 |
| ; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v2, v20, v21 :: v_dual_lshlrev_b32 v23, 16, v0 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) |
| ; GFX11-NEXT: v_add3_u32 v21, v22, v19, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v19 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v23 |
| ; GFX11-NEXT: v_perm_b32 v2, v2, v18, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v19, v21, v22 :: v_dual_add_f32 v0, 0x40c00000, v0 |
| ; GFX11-NEXT: v_bfe_u32 v24, v20, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v20 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v21, v0, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v0 |
| ; GFX11-NEXT: v_add3_u32 v24, v24, v20, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v0, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v23, v1, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 |
| ; GFX11-NEXT: v_add3_u32 v22, v23, v1, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v1, v22, v23, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 |
| ; GFX11-NEXT: v_perm_b32 v1, v1, v19, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v20, v24, v25, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v0, v21, v26, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) |
| ; GFX11-NEXT: v_perm_b32 v0, v0, v20, 0x7060302 |
| ; GFX11-NEXT: .LBB11_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x bfloat> %a, splat (bfloat 0xR40C0) |
| %a2 = bitcast <32 x bfloat> %a1 to <16 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x bfloat> %a to <16 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x i32> %phi |
| } |
| |
| define <64 x i8> @bitcast_v16i32_to_v64i8(<16 x i32> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16i32_to_v64i8: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB12_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v21, v16, v15, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v20, v14, v13, 16 |
| ; GCN-NEXT: v_alignbit_b32 v28, v14, v13, 8 |
| ; GCN-NEXT: v_alignbit_b32 v24, v12, v11, 24 |
| ; GCN-NEXT: v_alignbit_b32 v25, v12, v11, 16 |
| ; GCN-NEXT: v_alignbit_b32 v34, v12, v11, 8 |
| ; GCN-NEXT: v_alignbit_b32 v30, v10, v9, 24 |
| ; GCN-NEXT: v_alignbit_b32 v32, v10, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v48, v10, v9, 8 |
| ; GCN-NEXT: v_alignbit_b32 v36, v8, v7, 24 |
| ; GCN-NEXT: v_alignbit_b32 v37, v8, v7, 16 |
| ; GCN-NEXT: v_alignbit_b32 v54, v8, v7, 8 |
| ; GCN-NEXT: v_alignbit_b32 v49, v6, v5, 24 |
| ; GCN-NEXT: v_alignbit_b32 v51, v6, v5, 16 |
| ; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 8 |
| ; GCN-NEXT: v_alignbit_b32 v40, v4, v3, 24 |
| ; GCN-NEXT: v_alignbit_b32 v41, v4, v3, 16 |
| ; GCN-NEXT: v_alignbit_b32 v58, v4, v3, 8 |
| ; GCN-NEXT: v_alignbit_b32 v44, v2, v1, 24 |
| ; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 16 |
| ; GCN-NEXT: v_alignbit_b32 v60, v2, v1, 8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v50, 8, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v52, 24, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v42, 24, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v56, 24, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v61, 24, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2 |
| ; GCN-NEXT: .LBB12_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB12_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15 |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v21, v16, v15, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v20, v14, v13, 16 |
| ; GCN-NEXT: v_alignbit_b32 v28, v14, v13, 8 |
| ; GCN-NEXT: v_alignbit_b32 v24, v12, v11, 24 |
| ; GCN-NEXT: v_alignbit_b32 v25, v12, v11, 16 |
| ; GCN-NEXT: v_alignbit_b32 v34, v12, v11, 8 |
| ; GCN-NEXT: v_alignbit_b32 v30, v10, v9, 24 |
| ; GCN-NEXT: v_alignbit_b32 v32, v10, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v48, v10, v9, 8 |
| ; GCN-NEXT: v_alignbit_b32 v36, v8, v7, 24 |
| ; GCN-NEXT: v_alignbit_b32 v37, v8, v7, 16 |
| ; GCN-NEXT: v_alignbit_b32 v54, v8, v7, 8 |
| ; GCN-NEXT: v_alignbit_b32 v49, v6, v5, 24 |
| ; GCN-NEXT: v_alignbit_b32 v51, v6, v5, 16 |
| ; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 8 |
| ; GCN-NEXT: v_alignbit_b32 v40, v4, v3, 24 |
| ; GCN-NEXT: v_alignbit_b32 v41, v4, v3, 16 |
| ; GCN-NEXT: v_alignbit_b32 v58, v4, v3, 8 |
| ; GCN-NEXT: v_alignbit_b32 v44, v2, v1, 24 |
| ; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 16 |
| ; GCN-NEXT: v_alignbit_b32 v60, v2, v1, 8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v50, 8, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v52, 24, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v42, 24, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v56, 24, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v61, 24, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2 |
| ; GCN-NEXT: .LBB12_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 8, v60 |
| ; GCN-NEXT: s_waitcnt expcnt(3) |
| ; GCN-NEXT: v_or_b32_e32 v60, v1, v18 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v17 |
| ; GCN-NEXT: v_or_b32_e32 v17, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v58 |
| ; GCN-NEXT: v_or_b32_e32 v58, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v63 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v63, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v46 |
| ; GCN-NEXT: v_or_b32_e32 v46, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v44 |
| ; GCN-NEXT: v_and_b32_e32 v45, 0xff, v62 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v61, 24, v61 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v59 |
| ; GCN-NEXT: v_or_b32_e32 v44, v1, v2 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0 |
| ; GCN-NEXT: v_and_b32_e32 v41, 0xff, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v40, 24, v40 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v54 |
| ; GCN-NEXT: v_or_b32_e32 v7, v2, v3 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0 |
| ; GCN-NEXT: v_and_b32_e32 v57, 0xff, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v56, 24, v56 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v47 |
| ; GCN-NEXT: v_or_b32_e32 v54, v3, v4 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 12, v0 |
| ; GCN-NEXT: v_and_b32_e32 v47, 0xff, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v49 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v48 |
| ; GCN-NEXT: v_or_b32_e32 v9, v4, v5 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 16, v0 |
| ; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 24, v42 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v55 |
| ; GCN-NEXT: v_or_b32_e32 v48, v5, v6 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 20, v0 |
| ; GCN-NEXT: v_and_b32_e32 v55, 0xff, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v36 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 8, v34 |
| ; GCN-NEXT: v_or_b32_e32 v34, v6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0 |
| ; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v52 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xff, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 8, v50 |
| ; GCN-NEXT: v_or_b32_e32 v11, v8, v10 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 28, v0 |
| ; GCN-NEXT: v_and_b32_e32 v50, 0xff, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 24, v30 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 8, v28 |
| ; GCN-NEXT: v_or_b32_e32 v13, v10, v12 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 32, v0 |
| ; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v38 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v35 |
| ; GCN-NEXT: v_or_b32_e32 v28, v12, v14 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 36, v0 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 24, v24 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 8, v21 |
| ; GCN-NEXT: v_or_b32_e32 v21, v14, v15 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 40, v0 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xff, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 24, v31 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v29 |
| ; GCN-NEXT: v_or_b32_e32 v29, v15, v16 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 44, v0 |
| ; GCN-NEXT: v_and_b32_e32 v52, 0xff, v20 |
| ; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18 |
| ; GCN-NEXT: v_or_b32_e32 v31, v19, v16 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 48, v0 |
| ; GCN-NEXT: v_and_b32_e32 v42, 0xff, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 24, v26 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v45 |
| ; GCN-NEXT: v_or_b32_e32 v27, v61, v18 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 52, v0 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v45, 0xff, v19 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v41 |
| ; GCN-NEXT: v_or_b32_e32 v40, v40, v19 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v41, 24, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v57 |
| ; GCN-NEXT: v_or_b32_e32 v56, v56, v22 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 60, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v60 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v58 |
| ; GCN-NEXT: v_and_b32_e32 v59, 0xffff, v63 |
| ; GCN-NEXT: v_and_b32_e32 v46, 0xffff, v46 |
| ; GCN-NEXT: v_or_b32_e32 v49, v49, v47 |
| ; GCN-NEXT: v_and_b32_e32 v44, 0xffff, v44 |
| ; GCN-NEXT: v_or_b32_e32 v51, v51, v43 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GCN-NEXT: v_or_b32_e32 v36, v36, v55 |
| ; GCN-NEXT: v_and_b32_e32 v54, 0xffff, v54 |
| ; GCN-NEXT: v_or_b32_e32 v37, v37, v53 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_or_b32_e32 v30, v30, v50 |
| ; GCN-NEXT: v_and_b32_e32 v48, 0xffff, v48 |
| ; GCN-NEXT: v_or_b32_e32 v32, v32, v39 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xffff, v34 |
| ; GCN-NEXT: v_or_b32_e32 v24, v24, v35 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_or_b32_e32 v25, v25, v38 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_or_b32_e32 v20, v20, v52 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28 |
| ; GCN-NEXT: v_or_b32_e32 v26, v26, v42 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21 |
| ; GCN-NEXT: v_or_b32_e32 v33, v33, v45 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_or_b32_e32 v23, v41, v23 |
| ; GCN-NEXT: v_or_b32_e32 v31, v57, v31 |
| ; GCN-NEXT: v_or_b32_e32 v17, v17, v27 |
| ; GCN-NEXT: v_or_b32_e32 v27, v58, v40 |
| ; GCN-NEXT: v_or_b32_e32 v35, v59, v56 |
| ; GCN-NEXT: v_or_b32_e32 v38, v46, v49 |
| ; GCN-NEXT: v_or_b32_e32 v39, v44, v51 |
| ; GCN-NEXT: v_or_b32_e32 v7, v7, v36 |
| ; GCN-NEXT: v_or_b32_e32 v36, v54, v37 |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v30 |
| ; GCN-NEXT: v_or_b32_e32 v30, v48, v32 |
| ; GCN-NEXT: v_or_b32_e32 v24, v34, v24 |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v20 |
| ; GCN-NEXT: v_or_b32_e32 v20, v28, v26 |
| ; GCN-NEXT: v_or_b32_e32 v21, v21, v33 |
| ; GCN-NEXT: v_or_b32_e32 v23, v29, v23 |
| ; GCN-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v17, v1, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v27, v2, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v35, v3, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v38, v4, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v39, v5, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v7, v6, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v36, v8, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v9, v10, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v30, v12, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v24, v14, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v11, v15, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v13, v16, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v20, v18, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v21, v19, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v23, v22, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16i32_to_v64i8: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr49 |
| ; VI-NEXT: ; implicit-def: $vgpr48 |
| ; VI-NEXT: ; implicit-def: $vgpr39 |
| ; VI-NEXT: ; implicit-def: $vgpr38 |
| ; VI-NEXT: ; implicit-def: $vgpr37 |
| ; VI-NEXT: ; implicit-def: $vgpr36 |
| ; VI-NEXT: ; implicit-def: $vgpr35 |
| ; VI-NEXT: ; implicit-def: $vgpr34 |
| ; VI-NEXT: ; implicit-def: $vgpr33 |
| ; VI-NEXT: ; implicit-def: $vgpr32 |
| ; VI-NEXT: ; implicit-def: $vgpr31 |
| ; VI-NEXT: ; implicit-def: $vgpr30 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB12_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v16 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8] |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v16 |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; VI-NEXT: v_mov_b32_e32 v26, v22 |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4] |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2] |
| ; VI-NEXT: .LBB12_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB12_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16 |
| ; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15 |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13 |
| ; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11 |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9 |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3 |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1 |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7 |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4] |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5 |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8] |
| ; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2] |
| ; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 8, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: .LBB12_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v29 |
| ; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v22 |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v22, v27, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v28 |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v62 |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61 |
| ; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v59 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v57 |
| ; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v56 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v46 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44 |
| ; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v43 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v24 |
| ; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v55 |
| ; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v54 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v52 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v50 |
| ; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v49 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; VI-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37 |
| ; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; VI-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32 |
| ; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; VI-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v26 |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 60, v0 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16i32_to_v64i8: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB12_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v16 |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_mov_b32_e32 v26, v23 |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: .LBB12_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB12_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: v_add_u32_e32 v16, 3, v16 |
| ; GFX9-NEXT: v_add_u32_e32 v15, 3, v15 |
| ; GFX9-NEXT: v_add_u32_e32 v14, 3, v14 |
| ; GFX9-NEXT: v_add_u32_e32 v13, 3, v13 |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_add_u32_e32 v12, 3, v12 |
| ; GFX9-NEXT: v_add_u32_e32 v11, 3, v11 |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_add_u32_e32 v10, 3, v10 |
| ; GFX9-NEXT: v_add_u32_e32 v9, 3, v9 |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_add_u32_e32 v8, 3, v8 |
| ; GFX9-NEXT: v_add_u32_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_add_u32_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_add_u32_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_add_u32_e32 v2, 3, v2 |
| ; GFX9-NEXT: v_add_u32_e32 v1, 3, v1 |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_add_u32_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_add_u32_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: .LBB12_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v29 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v23 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v23, v27, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28 |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v22 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v60 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v59 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v57 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v56 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v46 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v21 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v55 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v50 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v49 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56 |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v26 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60 |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16i32_to_v64i8: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB12_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB12_2: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB12_4 |
| ; GFX11-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5 |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3 |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1 |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB12_4: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v96, 0xff, v96 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v86, 0xff, v86 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v87 |
| ; GFX11-NEXT: v_or_b32_e32 v24, v96, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v25 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v86, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84 |
| ; GFX11-NEXT: v_and_b32_e32 v83, 0xff, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v82 |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v80, 8, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v71, 8, v71 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v24 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v81 |
| ; GFX11-NEXT: v_and_b32_e32 v70, 0xff, v70 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v83, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v25, v80 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v71 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v70, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v22 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v66 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v65 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v55 |
| ; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v23, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v25, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v52 |
| ; GFX11-NEXT: v_lshlrev_b16 v20, 8, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v50 |
| ; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v48, 8, v48 |
| ; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v39 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v23, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v25, v49 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v48 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v39, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v38 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v37 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v35 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v34 |
| ; GFX11-NEXT: v_lshlrev_b16 v18, 8, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_lshlrev_b16 v69, 8, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v68, 0xff, v68 |
| ; GFX11-NEXT: v_lshlrev_b16 v67, 8, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v23, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v32 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v30 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v29 |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v26, 8, v26 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v69 |
| ; GFX11-NEXT: v_or_b32_e32 v67, v68, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v23, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v25, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v67 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; GFX11-NEXT: s_clause 0x3 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off |
| ; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <16 x i32> %a, splat (i32 3) |
| %a2 = bitcast <16 x i32> %a1 to <64 x i8> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x i32> %a to <64 x i8> |
| br label %end |
| |
| end: |
| %phi = phi <64 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <64 x i8> %phi |
| } |
| |
| define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v64i8_to_v16i32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_mov_b32_e32 v38, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v31, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104 |
| ; GCN-NEXT: s_waitcnt expcnt(3) |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:100 |
| ; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:96 |
| ; GCN-NEXT: s_waitcnt expcnt(2) |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:92 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88 |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:84 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76 |
| ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:72 |
| ; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:68 |
| ; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64 |
| ; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:60 |
| ; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56 |
| ; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:52 |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8 |
| ; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:132 |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:128 |
| ; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:124 |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:120 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v52 |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 |
| ; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:36 |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 |
| ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:28 |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:24 |
| ; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:20 |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v5 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v7 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v9 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v11 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v13 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v15 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v17 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v19 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v21 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v23 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v25 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v27 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v45 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 8, v44 |
| ; GCN-NEXT: s_waitcnt vmcnt(10) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v59, 24, v59 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v58 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v57, 24, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v58, 24, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v56, 8, v14 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v12 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v10 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v6 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v4 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v2 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:44 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:12 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116 |
| ; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0 |
| ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v46, 24, v46 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB13_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v42 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v41 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v40 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v55 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v35 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v36 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v37 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v38 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v50 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v36, 0xff, v54 |
| ; GCN-NEXT: v_and_b32_e32 v37, 0xff, v53 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xff, v52 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v49 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v48 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v39 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v63 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v62 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v61 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v60 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v21 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v32 |
| ; GCN-NEXT: v_or_b32_e32 v17, v33, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v34 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v20, v20, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v23, v24, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26 |
| ; GCN-NEXT: v_or_b32_e32 v26, v28, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v30 |
| ; GCN-NEXT: v_or_b32_e32 v27, v35, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_or_b32_e32 v29, v36, v44 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v37 |
| ; GCN-NEXT: v_or_b32_e32 v31, v38, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GCN-NEXT: v_or_b32_e32 v18, v18, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v7, v32, v7 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v6, v32, v6 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v5, v32, v5 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v32, v4 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v32, v21 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v22, v32, v22 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v24, v32, v24 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v28, v32, v28 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27 |
| ; GCN-NEXT: v_or_b32_e32 v25, v59, v25 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_or_b32_e32 v30, v57, v30 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31 |
| ; GCN-NEXT: v_or_b32_e32 v32, v58, v8 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v9 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v34, v8, v10 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xffff, v11 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v8, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v14, v8, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v16, v8, v16 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GCN-NEXT: v_or_b32_e32 v19, v46, v19 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v7 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v6 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v5 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; GCN-NEXT: v_or_b32_e32 v4, v17, v21 |
| ; GCN-NEXT: v_or_b32_e32 v5, v20, v22 |
| ; GCN-NEXT: v_or_b32_e32 v6, v23, v24 |
| ; GCN-NEXT: v_or_b32_e32 v7, v26, v28 |
| ; GCN-NEXT: v_or_b32_e32 v8, v27, v25 |
| ; GCN-NEXT: v_or_b32_e32 v9, v29, v30 |
| ; GCN-NEXT: v_or_b32_e32 v10, v31, v32 |
| ; GCN-NEXT: v_or_b32_e32 v11, v33, v34 |
| ; GCN-NEXT: v_or_b32_e32 v12, v35, v12 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; GCN-NEXT: v_or_b32_e32 v14, v15, v16 |
| ; GCN-NEXT: v_or_b32_e32 v15, v18, v19 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; kill: killed $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: .LBB13_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB13_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GCN-NEXT: v_or_b32_e32 v0, v42, v0 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GCN-NEXT: v_or_b32_e32 v1, v41, v1 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GCN-NEXT: v_or_b32_e32 v2, v40, v2 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GCN-NEXT: v_or_b32_e32 v3, v55, v3 |
| ; GCN-NEXT: s_movk_i32 s7, 0x300 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35 |
| ; GCN-NEXT: s_mov_b32 s6, 0x3000000 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v50 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v25 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v54 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v53 |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v52 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v49 |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v48 |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v39 |
| ; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v63 |
| ; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v62 |
| ; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v61 |
| ; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v60 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19 |
| ; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v51 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v7 |
| ; GCN-NEXT: v_or_b32_e32 v6, v17, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v9 |
| ; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v11 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v13 |
| ; GCN-NEXT: v_or_b32_e32 v13, v29, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15 |
| ; GCN-NEXT: v_or_b32_e32 v15, v27, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18 |
| ; GCN-NEXT: v_or_b32_e32 v17, v44, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v22 |
| ; GCN-NEXT: v_or_b32_e32 v20, v45, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v23 |
| ; GCN-NEXT: v_or_b32_e32 v23, v56, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v25, v25, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v30 |
| ; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v27, v27, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v32 |
| ; GCN-NEXT: v_or_b32_e32 v29, v43, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GCN-NEXT: v_or_b32_e32 v19, v47, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v34 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v31, v31, v36 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v8, v32, v8 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v5, v32, v5 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v32, v4 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v7, v32, v7 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v10, v32, v10 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v32, v12 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v14, v32, v14 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15 |
| ; GCN-NEXT: v_or_b32_e32 v16, v59, v16 |
| ; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v17 |
| ; GCN-NEXT: v_or_b32_e32 v18, v57, v18 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20 |
| ; GCN-NEXT: v_or_b32_e32 v22, v58, v22 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, s7, v23 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v24, v32, v24 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v26, v32, v26 |
| ; GCN-NEXT: v_add_i32_e32 v27, vcc, s7, v27 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v28, v32, v28 |
| ; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v32, v21 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 0x300, v19 |
| ; GCN-NEXT: v_or_b32_e32 v30, v46, v30 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GCN-NEXT: v_or_b32_e32 v0, v31, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v8, v1 |
| ; GCN-NEXT: v_or_b32_e32 v2, v5, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; GCN-NEXT: v_or_b32_e32 v4, v7, v6 |
| ; GCN-NEXT: v_or_b32_e32 v5, v10, v9 |
| ; GCN-NEXT: v_or_b32_e32 v6, v12, v11 |
| ; GCN-NEXT: v_or_b32_e32 v7, v14, v13 |
| ; GCN-NEXT: v_or_b32_e32 v8, v16, v15 |
| ; GCN-NEXT: v_or_b32_e32 v9, v18, v17 |
| ; GCN-NEXT: v_or_b32_e32 v10, v22, v20 |
| ; GCN-NEXT: v_or_b32_e32 v11, v24, v23 |
| ; GCN-NEXT: v_or_b32_e32 v12, v26, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v28, v27 |
| ; GCN-NEXT: v_or_b32_e32 v14, v21, v29 |
| ; GCN-NEXT: v_or_b32_e32 v15, v30, v19 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x3000000, v14 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x3000000, v15 |
| ; GCN-NEXT: .LBB13_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v64i8_to_v16i32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132 |
| ; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72 |
| ; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80 |
| ; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88 |
| ; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96 |
| ; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104 |
| ; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112 |
| ; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120 |
| ; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128 |
| ; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20 |
| ; VI-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12 |
| ; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100 |
| ; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68 |
| ; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36 |
| ; VI-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4 |
| ; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v26, 8, v23 |
| ; VI-NEXT: v_lshlrev_b16_e32 v28, 8, v25 |
| ; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v27 |
| ; VI-NEXT: v_lshlrev_b16_e32 v62, 8, v29 |
| ; VI-NEXT: v_lshlrev_b16_e32 v18, 8, v17 |
| ; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v19 |
| ; VI-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116 |
| ; VI-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124 |
| ; VI-NEXT: v_lshlrev_b16_e32 v31, 8, v1 |
| ; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v3 |
| ; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v5 |
| ; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v7 |
| ; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v9 |
| ; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v11 |
| ; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v13 |
| ; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v15 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54 |
| ; VI-NEXT: v_lshlrev_b16_e32 v61, 8, v0 |
| ; VI-NEXT: v_lshlrev_b16_e32 v59, 8, v2 |
| ; VI-NEXT: v_lshlrev_b16_e32 v58, 8, v4 |
| ; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v6 |
| ; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v8 |
| ; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v10 |
| ; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v12 |
| ; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v14 |
| ; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v24 |
| ; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v30 |
| ; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v53 |
| ; VI-NEXT: s_waitcnt vmcnt(13) |
| ; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v42 |
| ; VI-NEXT: s_waitcnt vmcnt(12) |
| ; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v44 |
| ; VI-NEXT: s_waitcnt vmcnt(11) |
| ; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v47 |
| ; VI-NEXT: s_waitcnt vmcnt(10) |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v57 |
| ; VI-NEXT: s_waitcnt vmcnt(9) |
| ; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v60 |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_lshlrev_b16_e32 v57, 8, v63 |
| ; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108 |
| ; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92 |
| ; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84 |
| ; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76 |
| ; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60 |
| ; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44 |
| ; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB13_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr30 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr23 |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(7) |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr31 |
| ; VI-NEXT: ; implicit-def: $vgpr32 |
| ; VI-NEXT: ; implicit-def: $vgpr33 |
| ; VI-NEXT: ; implicit-def: $vgpr34 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr38 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr35 |
| ; VI-NEXT: ; implicit-def: $vgpr36 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr37 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr49 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr48 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr39 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: .LBB13_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB13_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: v_mov_b32_e32 v15, 0x300 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v51 |
| ; VI-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v9, v9, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v24 |
| ; VI-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v10, v10, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v30 |
| ; VI-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v11, v11, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v60 |
| ; VI-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v12, v12, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v47 |
| ; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v13, v13, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v42 |
| ; VI-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v14, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v17, 3, v17 |
| ; VI-NEXT: v_or_b32_sdwa v17, v57, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: s_waitcnt vmcnt(7) |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v1 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v0, 0x300, v0 |
| ; VI-NEXT: v_add_u16_sdwa v1, v1, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; VI-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; VI-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v4, v4, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; VI-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v5, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; VI-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v6, v6, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; VI-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v7, v7, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v53 |
| ; VI-NEXT: v_or_b32_sdwa v16, v23, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v16, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v8 |
| ; VI-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v8, v8, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: v_add_u16_sdwa v15, v17, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v3 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; VI-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v4, 0x300, v4 |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; VI-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v5, 0x300, v5 |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; VI-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v6, 0x300, v6 |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; VI-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v7, 0x300, v7 |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v52 |
| ; VI-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v8, 0x300, v8 |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v50 |
| ; VI-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v9, 0x300, v9 |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v49 |
| ; VI-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v10, 0x300, v10 |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v63 |
| ; VI-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v11, 0x300, v11 |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v48 |
| ; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v12, 0x300, v12 |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v44 |
| ; VI-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v13, 0x300, v13 |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v39 |
| ; VI-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v14, 0x300, v14 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v16 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v19 |
| ; VI-NEXT: v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v16, 0x300, v16 |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: .LBB13_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v64i8_to_v16i32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132 |
| ; GFX9-NEXT: s_nop 0 |
| ; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64 |
| ; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72 |
| ; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80 |
| ; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88 |
| ; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96 |
| ; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104 |
| ; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112 |
| ; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120 |
| ; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128 |
| ; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20 |
| ; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12 |
| ; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100 |
| ; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68 |
| ; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36 |
| ; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v21 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v26, 8, v23 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v28, 8, v25 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v27 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v62, 8, v29 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v18, 8, v17 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v19 |
| ; GFX9-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116 |
| ; GFX9-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v31, 8, v1 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v3 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v5 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v7 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v9 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v11 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v13 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v15 |
| ; GFX9-NEXT: s_waitcnt vmcnt(25) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54 |
| ; GFX9-NEXT: s_waitcnt vmcnt(24) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v61, 8, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(23) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v59, 8, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(22) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v58, 8, v4 |
| ; GFX9-NEXT: s_waitcnt vmcnt(21) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(20) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v8 |
| ; GFX9-NEXT: s_waitcnt vmcnt(19) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v10 |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v12 |
| ; GFX9-NEXT: s_waitcnt vmcnt(17) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v14 |
| ; GFX9-NEXT: s_waitcnt vmcnt(16) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v24 |
| ; GFX9-NEXT: s_waitcnt vmcnt(15) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v30 |
| ; GFX9-NEXT: s_waitcnt vmcnt(14) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v53 |
| ; GFX9-NEXT: s_waitcnt vmcnt(13) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v42 |
| ; GFX9-NEXT: s_waitcnt vmcnt(12) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v44 |
| ; GFX9-NEXT: s_waitcnt vmcnt(11) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v47 |
| ; GFX9-NEXT: s_waitcnt vmcnt(10) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v57 |
| ; GFX9-NEXT: s_waitcnt vmcnt(9) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v60 |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v57, 8, v63 |
| ; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108 |
| ; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92 |
| ; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84 |
| ; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76 |
| ; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60 |
| ; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52 |
| ; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44 |
| ; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB13_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: .LBB13_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB13_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_movk_i32 s6, 0x300 |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_add_u16_e32 v9, 3, v51 |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_add_u16_e32 v10, 3, v24 |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v11, 3, v30 |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v12, 3, v60 |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v13, 3, v47 |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v14, 3, v42 |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v15, 3, v53 |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_add_u16_e32 v1, 3, v1 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GFX9-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v16, 3, v17 |
| ; GFX9-NEXT: v_or_b32_sdwa v16, v57, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_add_u16_e32 v8, 3, v8 |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3 |
| ; GFX9-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4 |
| ; GFX9-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5 |
| ; GFX9-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6 |
| ; GFX9-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7 |
| ; GFX9-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; GFX9-NEXT: v_add_u16_e32 v8, 3, v52 |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8 |
| ; GFX9-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; GFX9-NEXT: v_add_u16_e32 v9, 3, v50 |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9 |
| ; GFX9-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; GFX9-NEXT: v_add_u16_e32 v10, 3, v49 |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10 |
| ; GFX9-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; GFX9-NEXT: v_add_u16_e32 v11, 3, v63 |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11 |
| ; GFX9-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GFX9-NEXT: v_add_u16_e32 v12, 3, v48 |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12 |
| ; GFX9-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; GFX9-NEXT: v_add_u16_e32 v13, 3, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13 |
| ; GFX9-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; GFX9-NEXT: v_add_u16_e32 v14, 3, v39 |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14 |
| ; GFX9-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; GFX9-NEXT: v_add_u16_e32 v15, 3, v19 |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15 |
| ; GFX9-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; GFX9-NEXT: .LBB13_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v64i8_to_v16i32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12 |
| ; GFX11-NEXT: v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8 |
| ; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4 |
| ; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0 |
| ; GFX11-NEXT: s_clause 0x1f |
| ; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:128 |
| ; GFX11-NEXT: scratch_load_u16 v39, off, s32 offset:124 |
| ; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:120 |
| ; GFX11-NEXT: scratch_load_u16 v48, off, s32 offset:116 |
| ; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:112 |
| ; GFX11-NEXT: scratch_load_u16 v49, off, s32 offset:108 |
| ; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:104 |
| ; GFX11-NEXT: scratch_load_u16 v50, off, s32 offset:100 |
| ; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:96 |
| ; GFX11-NEXT: scratch_load_u16 v51, off, s32 offset:92 |
| ; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:88 |
| ; GFX11-NEXT: scratch_load_b32 v12, off, s32 offset:132 |
| ; GFX11-NEXT: scratch_load_u16 v14, off, s32 |
| ; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:8 |
| ; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:16 |
| ; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:24 |
| ; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:32 |
| ; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:40 |
| ; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:48 |
| ; GFX11-NEXT: scratch_load_u16 v128, off, s32 offset:56 |
| ; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:64 |
| ; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:72 |
| ; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:80 |
| ; GFX11-NEXT: scratch_load_u16 v52, off, s32 offset:84 |
| ; GFX11-NEXT: scratch_load_u16 v53, off, s32 offset:76 |
| ; GFX11-NEXT: scratch_load_u16 v54, off, s32 offset:68 |
| ; GFX11-NEXT: scratch_load_u16 v55, off, s32 offset:60 |
| ; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:52 |
| ; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:44 |
| ; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:36 |
| ; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:28 |
| ; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:20 |
| ; GFX11-NEXT: s_clause 0x1 |
| ; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:12 |
| ; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:4 |
| ; GFX11-NEXT: v_lshlrev_b16 v118, 8, v1 |
| ; GFX11-NEXT: v_lshlrev_b16 v119, 8, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v117, 8, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v114, 8, v7 |
| ; GFX11-NEXT: v_lshlrev_b16 v115, 8, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v116, 8, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v113, 8, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v100, 8, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v101, 8, v17 |
| ; GFX11-NEXT: v_lshlrev_b16 v102, 8, v19 |
| ; GFX11-NEXT: v_lshlrev_b16 v103, 8, v21 |
| ; GFX11-NEXT: v_lshlrev_b16 v112, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v87, 8, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v96, 8, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v97, 8, v29 |
| ; GFX11-NEXT: s_waitcnt vmcnt(33) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(31) |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v2 |
| ; GFX11-NEXT: s_waitcnt vmcnt(29) |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v4 |
| ; GFX11-NEXT: s_waitcnt vmcnt(27) |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v6 |
| ; GFX11-NEXT: s_waitcnt vmcnt(25) |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v8 |
| ; GFX11-NEXT: s_waitcnt vmcnt(22) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12 |
| ; GFX11-NEXT: s_waitcnt vmcnt(21) |
| ; GFX11-NEXT: v_lshlrev_b16 v98, 8, v14 |
| ; GFX11-NEXT: s_waitcnt vmcnt(20) |
| ; GFX11-NEXT: v_lshlrev_b16 v99, 8, v65 |
| ; GFX11-NEXT: s_waitcnt vmcnt(19) |
| ; GFX11-NEXT: v_lshlrev_b16 v81, 8, v66 |
| ; GFX11-NEXT: s_waitcnt vmcnt(18) |
| ; GFX11-NEXT: v_lshlrev_b16 v82, 8, v67 |
| ; GFX11-NEXT: s_waitcnt vmcnt(17) |
| ; GFX11-NEXT: v_lshlrev_b16 v83, 8, v83 |
| ; GFX11-NEXT: s_waitcnt vmcnt(16) |
| ; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84 |
| ; GFX11-NEXT: s_waitcnt vmcnt(15) |
| ; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85 |
| ; GFX11-NEXT: s_waitcnt vmcnt(14) |
| ; GFX11-NEXT: v_lshlrev_b16 v27, 8, v128 |
| ; GFX11-NEXT: s_waitcnt vmcnt(13) |
| ; GFX11-NEXT: v_lshlrev_b16 v29, 8, v129 |
| ; GFX11-NEXT: s_waitcnt vmcnt(12) |
| ; GFX11-NEXT: v_lshlrev_b16 v65, 8, v130 |
| ; GFX11-NEXT: s_waitcnt vmcnt(11) |
| ; GFX11-NEXT: v_lshlrev_b16 v66, 8, v131 |
| ; GFX11-NEXT: v_lshlrev_b16 v67, 8, v10 |
| ; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB13_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB13_4 |
| ; GFX11-NEXT: .LBB13_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB13_3: ; %cmp.false |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v37 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v118 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v119 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v35 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v117 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v6, v113 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v114 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v115 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v116 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v101 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v38 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v102 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v103 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v112 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v100 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v1 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v4, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v6, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v7, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v9, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v30 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v86 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v71 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v70 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v68 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v87 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v96 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v97 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v98 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v99 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v81 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v82 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v8, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v10, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v12, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v14, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v55 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v53 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v52 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v50 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v49 |
| ; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v48 |
| ; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v39 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v27 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v29 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v65 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v66 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v18, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v22, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v24, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v13, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v15, v16 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v17, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v19, v20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr16 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr118 |
| ; GFX11-NEXT: ; implicit-def: $vgpr119 |
| ; GFX11-NEXT: ; implicit-def: $vgpr117 |
| ; GFX11-NEXT: ; implicit-def: $vgpr114 |
| ; GFX11-NEXT: ; implicit-def: $vgpr115 |
| ; GFX11-NEXT: ; implicit-def: $vgpr116 |
| ; GFX11-NEXT: ; implicit-def: $vgpr113 |
| ; GFX11-NEXT: ; implicit-def: $vgpr100 |
| ; GFX11-NEXT: ; implicit-def: $vgpr101 |
| ; GFX11-NEXT: ; implicit-def: $vgpr102 |
| ; GFX11-NEXT: ; implicit-def: $vgpr103 |
| ; GFX11-NEXT: ; implicit-def: $vgpr112 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr97 |
| ; GFX11-NEXT: ; implicit-def: $vgpr98 |
| ; GFX11-NEXT: ; implicit-def: $vgpr99 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB13_2 |
| ; GFX11-NEXT: .LBB13_4: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u16 v0, v31, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v32, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v33, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v34, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v35, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v5, v36, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v6, v37, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v118, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v119, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v117, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v114, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v115, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v116, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v113, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v7, v16, 3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v5 |
| ; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v6, v38, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v8, v18, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v9, v20, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v10, v22, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v101, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v100, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v102, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v103, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v112, v10 |
| ; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7 |
| ; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8 |
| ; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9 |
| ; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v3, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v5, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v7, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v9, v10 |
| ; GFX11-NEXT: v_add_nc_u16 v6, v24, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v7, v26, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v8, v28, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v9, v30, 3 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_add_nc_u16 v10, v86, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v11, v80, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v12, v71, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v13, v70, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v14, v69, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v15, v68, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v87, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v96, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v97, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v98, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v99, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v81, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v82, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v83, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v84, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v85, v15 |
| ; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7 |
| ; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8 |
| ; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9 |
| ; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10 |
| ; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11 |
| ; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12 |
| ; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13 |
| ; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14 |
| ; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v8, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v10, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v12, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v14, v15 |
| ; GFX11-NEXT: v_add_nc_u16 v11, v64, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v12, v55, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v13, v54, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v14, v53, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v15, v52, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v16, v51, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v18, v50, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v20, v49, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v22, v48, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v24, v39, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v27, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v29, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v65, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v66, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v67, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v19, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v21, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v23, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v25, v24 |
| ; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11 |
| ; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12 |
| ; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13 |
| ; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14 |
| ; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15 |
| ; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16 |
| ; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17 |
| ; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18 |
| ; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19 |
| ; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v13, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v15, v16 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v17, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v19, v20 |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <64 x i8> %a, splat (i8 3) |
| %a2 = bitcast <64 x i8> %a1 to <16 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <64 x i8> %a to <16 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <16 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x i32> %phi |
| } |
| |
| define <8 x i64> @bitcast_v16f32_to_v8i64(<16 x float> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16f32_to_v8i64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB14_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; GCN-NEXT: .LBB14_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16f32_to_v8i64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB14_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; VI-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; VI-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; VI-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; VI-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; VI-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; VI-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; VI-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; VI-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; VI-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; VI-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; VI-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; VI-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; VI-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; VI-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; VI-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; VI-NEXT: .LBB14_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16f32_to_v8i64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB14_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; GFX9-NEXT: .LBB14_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16f32_to_v8i64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB14_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14 |
| ; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12 |
| ; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10 |
| ; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8 |
| ; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6 |
| ; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4 |
| ; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2 |
| ; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0 |
| ; GFX11-NEXT: .LBB14_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <16 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <16 x float> %a1 to <8 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x float> %a to <8 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x i64> %phi |
| } |
| |
| define <16 x float> @bitcast_v8i64_to_v16f32(<8 x i64> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8i64_to_v16f32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB15_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; GCN-NEXT: .LBB15_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8i64_to_v16f32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB15_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0 |
| ; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; VI-NEXT: .LBB15_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8i64_to_v16f32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB15_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc |
| ; GFX9-NEXT: .LBB15_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8i64_to_v16f32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB15_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo |
| ; GFX11-NEXT: .LBB15_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <8 x i64> %a, splat (i64 3) |
| %a2 = bitcast <8 x i64> %a1 to <16 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x i64> %a to <16 x float> |
| br label %end |
| |
| end: |
| %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x float> %phi |
| } |
| |
| define <8 x double> @bitcast_v16f32_to_v8f64(<16 x float> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16f32_to_v8f64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB16_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; GCN-NEXT: .LBB16_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16f32_to_v8f64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB16_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; VI-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; VI-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; VI-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; VI-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; VI-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; VI-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; VI-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; VI-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; VI-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; VI-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; VI-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; VI-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; VI-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; VI-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; VI-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; VI-NEXT: .LBB16_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16f32_to_v8f64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB16_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; GFX9-NEXT: .LBB16_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16f32_to_v8f64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB16_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14 |
| ; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12 |
| ; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10 |
| ; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8 |
| ; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6 |
| ; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4 |
| ; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2 |
| ; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0 |
| ; GFX11-NEXT: .LBB16_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <16 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <16 x float> %a1 to <8 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x float> %a to <8 x double> |
| br label %end |
| |
| end: |
| %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x double> %phi |
| } |
| |
| define <16 x float> @bitcast_v8f64_to_v16f32(<8 x double> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8f64_to_v16f32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB17_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GCN-NEXT: .LBB17_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8f64_to_v16f32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB17_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; VI-NEXT: .LBB17_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8f64_to_v16f32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB17_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX9-NEXT: .LBB17_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8f64_to_v16f32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB17_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX11-NEXT: .LBB17_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <8 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <8 x double> %a1 to <16 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x double> %a to <16 x float> |
| br label %end |
| |
| end: |
| %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x float> %phi |
| } |
| |
| define <32 x i16> @bitcast_v16f32_to_v32i16(<16 x float> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16f32_to_v32i16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v30, v15 |
| ; GCN-NEXT: v_mov_b32_e32 v28, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v26, v13 |
| ; GCN-NEXT: v_mov_b32_e32 v24, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v22, v11 |
| ; GCN-NEXT: v_mov_b32_e32 v20, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v18, v9 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v14, v7 |
| ; GCN-NEXT: v_mov_b32_e32 v12, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v10, v5 |
| ; GCN-NEXT: v_mov_b32_e32 v8, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v6, v3 |
| ; GCN-NEXT: v_mov_b32_e32 v4, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v2, v1 |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB18_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_alignbit_b32 v29, v30, v28, 16 |
| ; GCN-NEXT: v_alignbit_b32 v25, v26, v24, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v22, v20, 16 |
| ; GCN-NEXT: v_alignbit_b32 v17, v18, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v10, v8, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v26 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v18 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2 |
| ; GCN-NEXT: .LBB18_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB18_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; GCN-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; GCN-NEXT: v_add_f32_e32 v32, 1.0, v32 |
| ; GCN-NEXT: v_add_f32_e32 v22, 1.0, v22 |
| ; GCN-NEXT: v_add_f32_e32 v20, 1.0, v20 |
| ; GCN-NEXT: v_add_f32_e32 v26, 1.0, v26 |
| ; GCN-NEXT: v_add_f32_e32 v24, 1.0, v24 |
| ; GCN-NEXT: v_add_f32_e32 v30, 1.0, v30 |
| ; GCN-NEXT: v_add_f32_e32 v28, 1.0, v28 |
| ; GCN-NEXT: v_alignbit_b32 v29, v30, v28, 16 |
| ; GCN-NEXT: v_alignbit_b32 v25, v26, v24, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v22, v20, 16 |
| ; GCN-NEXT: v_alignbit_b32 v17, v18, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v10, v8, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v26 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v18 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2 |
| ; GCN-NEXT: .LBB18_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_mov_b32_e32 v16, v32 |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16f32_to_v32i16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB18_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; VI-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; VI-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; VI-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; VI-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; VI-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; VI-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; VI-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; VI-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; VI-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; VI-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; VI-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; VI-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; VI-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; VI-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; VI-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; VI-NEXT: .LBB18_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16f32_to_v32i16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB18_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; GFX9-NEXT: .LBB18_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16f32_to_v32i16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB18_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14 |
| ; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12 |
| ; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10 |
| ; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8 |
| ; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6 |
| ; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4 |
| ; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2 |
| ; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0 |
| ; GFX11-NEXT: .LBB18_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <16 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <16 x float> %a1 to <32 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x float> %a to <32 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x i16> %phi |
| } |
| |
| define <16 x float> @bitcast_v32i16_to_v16f32(<32 x i16> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32i16_to_v16f32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v38, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v31, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v0 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB19_3 |
| ; GCN-NEXT: ; %bb.1: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB19_4 |
| ; GCN-NEXT: .LBB19_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; GCN-NEXT: .LBB19_3: ; %cmp.false |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v31 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v32 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v54 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v55 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v33 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v34 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v35 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v36 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v37 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v38 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v30 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v39 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v48 |
| ; GCN-NEXT: v_or_b32_e32 v4, v4, v49 |
| ; GCN-NEXT: v_or_b32_e32 v5, v5, v50 |
| ; GCN-NEXT: v_or_b32_e32 v6, v6, v51 |
| ; GCN-NEXT: v_or_b32_e32 v7, v7, v52 |
| ; GCN-NEXT: v_or_b32_e32 v8, v8, v17 |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v19 |
| ; GCN-NEXT: v_or_b32_e32 v10, v10, v21 |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v23 |
| ; GCN-NEXT: v_or_b32_e32 v12, v12, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v27 |
| ; GCN-NEXT: v_or_b32_e32 v14, v14, v29 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v53 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB19_2 |
| ; GCN-NEXT: .LBB19_4: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31 |
| ; GCN-NEXT: s_mov_b32 s6, 0x30000 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_or_b32_e32 v0, v54, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v55, v1 |
| ; GCN-NEXT: v_or_b32_e32 v2, v39, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v48, v3 |
| ; GCN-NEXT: v_or_b32_e32 v4, v49, v4 |
| ; GCN-NEXT: v_or_b32_e32 v5, v50, v5 |
| ; GCN-NEXT: v_or_b32_e32 v6, v51, v6 |
| ; GCN-NEXT: v_or_b32_e32 v7, v52, v7 |
| ; GCN-NEXT: v_or_b32_e32 v8, v17, v8 |
| ; GCN-NEXT: v_or_b32_e32 v9, v19, v9 |
| ; GCN-NEXT: v_or_b32_e32 v10, v21, v10 |
| ; GCN-NEXT: v_or_b32_e32 v11, v23, v11 |
| ; GCN-NEXT: v_or_b32_e32 v12, v25, v12 |
| ; GCN-NEXT: v_or_b32_e32 v13, v27, v13 |
| ; GCN-NEXT: v_or_b32_e32 v14, v29, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v53, v15 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v13 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v14 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x30000, v15 |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32i16_to_v16f32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB19_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v17, 3 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v15 |
| ; VI-NEXT: v_add_u16_sdwa v15, v15, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v14 |
| ; VI-NEXT: v_add_u16_sdwa v14, v14, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v14, v16, v14 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v13 |
| ; VI-NEXT: v_add_u16_sdwa v13, v13, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v13, v16, v13 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v12 |
| ; VI-NEXT: v_add_u16_sdwa v12, v12, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v12, v16, v12 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v11 |
| ; VI-NEXT: v_add_u16_sdwa v11, v11, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v11, v16, v11 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v10 |
| ; VI-NEXT: v_add_u16_sdwa v10, v10, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v10, v16, v10 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v9 |
| ; VI-NEXT: v_add_u16_sdwa v9, v9, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v9, v16, v9 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v8 |
| ; VI-NEXT: v_add_u16_sdwa v8, v8, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v8, v16, v8 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v7 |
| ; VI-NEXT: v_add_u16_sdwa v7, v7, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v7, v16, v7 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v6 |
| ; VI-NEXT: v_add_u16_sdwa v6, v6, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v6, v16, v6 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v5 |
| ; VI-NEXT: v_add_u16_sdwa v5, v5, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v5, v16, v5 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v4 |
| ; VI-NEXT: v_add_u16_sdwa v4, v4, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v4, v16, v4 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v3 |
| ; VI-NEXT: v_add_u16_sdwa v3, v3, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v3, v16, v3 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v2 |
| ; VI-NEXT: v_add_u16_sdwa v2, v2, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v2, v16, v2 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v1 |
| ; VI-NEXT: v_add_u16_sdwa v1, v1, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v1, v16, v1 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v0 |
| ; VI-NEXT: v_add_u16_sdwa v0, v0, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v0, v16, v0 |
| ; VI-NEXT: .LBB19_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32i16_to_v16f32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB19_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB19_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32i16_to_v16f32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB19_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: .LBB19_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <32 x i16> %a, splat (i16 3) |
| %a2 = bitcast <32 x i16> %a1 to <16 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x i16> %a to <16 x float> |
| br label %end |
| |
| end: |
| %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x float> %phi |
| } |
| |
| define <32 x half> @bitcast_v16f32_to_v32f16(<16 x float> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16f32_to_v32f16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_mov_b32_e32 v33, v15 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v13 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v11 |
| ; GCN-NEXT: v_mov_b32_e32 v38, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v39, v9 |
| ; GCN-NEXT: v_mov_b32_e32 v48, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v49, v7 |
| ; GCN-NEXT: v_mov_b32_e32 v50, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v51, v5 |
| ; GCN-NEXT: v_mov_b32_e32 v52, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v53, v3 |
| ; GCN-NEXT: v_mov_b32_e32 v54, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v55, v1 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v0 |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB20_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v34 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v35 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v36 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v37 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v38 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v39 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v48 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v49 |
| ; GCN-NEXT: s_waitcnt expcnt(6) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v50 |
| ; GCN-NEXT: s_waitcnt expcnt(5) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v51 |
| ; GCN-NEXT: s_waitcnt expcnt(4) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v52 |
| ; GCN-NEXT: s_waitcnt expcnt(3) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v53 |
| ; GCN-NEXT: s_waitcnt expcnt(2) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v54 |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v55 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v53 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v55 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v40 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v41 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v42 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v43 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v44 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v45 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v46 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v32 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: .LBB20_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB20_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_f32_e32 v0, 1.0, v32 |
| ; GCN-NEXT: v_add_f32_e32 v1, 1.0, v55 |
| ; GCN-NEXT: v_add_f32_e32 v2, 1.0, v54 |
| ; GCN-NEXT: v_add_f32_e32 v3, 1.0, v53 |
| ; GCN-NEXT: v_add_f32_e32 v4, 1.0, v52 |
| ; GCN-NEXT: v_add_f32_e32 v5, 1.0, v51 |
| ; GCN-NEXT: v_add_f32_e32 v6, 1.0, v50 |
| ; GCN-NEXT: v_add_f32_e32 v7, 1.0, v49 |
| ; GCN-NEXT: v_add_f32_e32 v8, 1.0, v48 |
| ; GCN-NEXT: v_add_f32_e32 v9, 1.0, v39 |
| ; GCN-NEXT: v_add_f32_e32 v10, 1.0, v38 |
| ; GCN-NEXT: v_add_f32_e32 v11, 1.0, v37 |
| ; GCN-NEXT: v_add_f32_e32 v12, 1.0, v36 |
| ; GCN-NEXT: v_add_f32_e32 v13, 1.0, v35 |
| ; GCN-NEXT: v_add_f32_e32 v14, 1.0, v34 |
| ; GCN-NEXT: v_add_f32_e32 v15, 1.0, v33 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v0 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v1 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v7 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v9 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v14 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v12 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v10 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v4 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v32 |
| ; GCN-NEXT: .LBB20_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16f32_to_v32f16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB20_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; VI-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; VI-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; VI-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; VI-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; VI-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; VI-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; VI-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; VI-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; VI-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; VI-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; VI-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; VI-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; VI-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; VI-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; VI-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; VI-NEXT: .LBB20_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16f32_to_v32f16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB20_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; GFX9-NEXT: .LBB20_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16f32_to_v32f16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB20_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14 |
| ; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12 |
| ; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10 |
| ; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8 |
| ; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6 |
| ; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4 |
| ; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2 |
| ; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0 |
| ; GFX11-NEXT: .LBB20_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <16 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <16 x float> %a1 to <32 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x float> %a to <32 x half> |
| br label %end |
| |
| end: |
| %phi = phi <32 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x half> %phi |
| } |
| |
| define <16 x float> @bitcast_v32f16_to_v16f32(<32 x half> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32f16_to_v16f32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v45, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v44, v0 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v43, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v42, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v41, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v52, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v40, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v50, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v55, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v48, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v54, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v38, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v53, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v36, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v51, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v34, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v49, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v33, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v39, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v32, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v37, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v31, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v35, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v23, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v28 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v20, v46 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v30 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB21_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v43 |
| ; GCN-NEXT: v_or_b32_e32 v0, v44, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v42, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v40 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v20 |
| ; GCN-NEXT: v_or_b32_e32 v2, v52, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v50, v3 |
| ; GCN-NEXT: v_or_b32_e32 v4, v48, v4 |
| ; GCN-NEXT: v_or_b32_e32 v5, v38, v5 |
| ; GCN-NEXT: v_or_b32_e32 v6, v36, v6 |
| ; GCN-NEXT: v_or_b32_e32 v7, v34, v7 |
| ; GCN-NEXT: v_or_b32_e32 v8, v33, v8 |
| ; GCN-NEXT: v_or_b32_e32 v9, v32, v9 |
| ; GCN-NEXT: v_or_b32_e32 v10, v31, v10 |
| ; GCN-NEXT: v_or_b32_e32 v11, v21, v11 |
| ; GCN-NEXT: v_or_b32_e32 v12, v19, v12 |
| ; GCN-NEXT: v_or_b32_e32 v13, v18, v13 |
| ; GCN-NEXT: v_or_b32_e32 v14, v17, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: .LBB21_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB21_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v45 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v44 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v43 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v42 |
| ; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GCN-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v3, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v41 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v40 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v55 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v53 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v31 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GCN-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v5, v4 |
| ; GCN-NEXT: v_or_b32_e32 v4, v7, v6 |
| ; GCN-NEXT: v_or_b32_e32 v5, v9, v8 |
| ; GCN-NEXT: v_or_b32_e32 v6, v11, v10 |
| ; GCN-NEXT: v_or_b32_e32 v7, v13, v12 |
| ; GCN-NEXT: v_or_b32_e32 v8, v15, v14 |
| ; GCN-NEXT: v_or_b32_e32 v9, v26, v24 |
| ; GCN-NEXT: v_or_b32_e32 v10, v28, v27 |
| ; GCN-NEXT: v_or_b32_e32 v11, v21, v29 |
| ; GCN-NEXT: v_or_b32_e32 v12, v19, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v18, v23 |
| ; GCN-NEXT: v_or_b32_e32 v14, v17, v22 |
| ; GCN-NEXT: v_or_b32_e32 v15, v16, v20 |
| ; GCN-NEXT: .LBB21_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32f16_to_v16f32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB21_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v16, 0x200 |
| ; VI-NEXT: v_add_f16_sdwa v17, v15, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v15, 0x200, v15 |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v14, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v14, 0x200, v14 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v13, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v13, 0x200, v13 |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v12, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v12, 0x200, v12 |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v11, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v11, 0x200, v11 |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v10, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v10, 0x200, v10 |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v9, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v9, 0x200, v9 |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v8, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v8, 0x200, v8 |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v7, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v7, 0x200, v7 |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v6, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v6, 0x200, v6 |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v5, 0x200, v5 |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v4, 0x200, v4 |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, 0x200, v3 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v2, 0x200, v2 |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v1, 0x200, v1 |
| ; VI-NEXT: v_add_f16_sdwa v16, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v0, 0x200, v0 |
| ; VI-NEXT: v_or_b32_e32 v1, v1, v17 |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v16 |
| ; VI-NEXT: .LBB21_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32f16_to_v16f32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB21_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: s_movk_i32 s6, 0x200 |
| ; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB21_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32f16_to_v16f32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB21_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-NEXT: .LBB21_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <32 x half> %a1 to <16 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x half> %a to <16 x float> |
| br label %end |
| |
| end: |
| %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x float> %phi |
| } |
| |
| define <32 x bfloat> @bitcast_v16f32_to_v32bf16(<16 x float> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16f32_to_v32bf16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v55, v15 |
| ; GCN-NEXT: v_mov_b32_e32 v54, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v53, v13 |
| ; GCN-NEXT: v_mov_b32_e32 v52, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v51, v11 |
| ; GCN-NEXT: v_mov_b32_e32 v50, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v49, v9 |
| ; GCN-NEXT: v_mov_b32_e32 v48, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v39, v7 |
| ; GCN-NEXT: v_mov_b32_e32 v38, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v5 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v3 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v1 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v0 |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB22_3 |
| ; GCN-NEXT: ; %bb.1: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB22_4 |
| ; GCN-NEXT: .LBB22_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; GCN-NEXT: .LBB22_3: ; %cmp.false |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v55 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v54 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v53 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v52 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v51 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v50 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v49 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v48 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v48 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v39 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v38 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v36 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v36 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v35 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v34 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v34 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v33 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v32 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB22_2 |
| ; GCN-NEXT: .LBB22_4: ; %cmp.true |
| ; GCN-NEXT: v_add_f32_e32 v0, 1.0, v32 |
| ; GCN-NEXT: v_add_f32_e32 v1, 1.0, v33 |
| ; GCN-NEXT: v_add_f32_e32 v2, 1.0, v34 |
| ; GCN-NEXT: v_add_f32_e32 v3, 1.0, v35 |
| ; GCN-NEXT: v_add_f32_e32 v4, 1.0, v36 |
| ; GCN-NEXT: v_add_f32_e32 v5, 1.0, v37 |
| ; GCN-NEXT: v_add_f32_e32 v6, 1.0, v38 |
| ; GCN-NEXT: v_add_f32_e32 v7, 1.0, v39 |
| ; GCN-NEXT: v_add_f32_e32 v8, 1.0, v48 |
| ; GCN-NEXT: v_add_f32_e32 v9, 1.0, v49 |
| ; GCN-NEXT: v_add_f32_e32 v10, 1.0, v50 |
| ; GCN-NEXT: v_add_f32_e32 v11, 1.0, v51 |
| ; GCN-NEXT: v_add_f32_e32 v12, 1.0, v52 |
| ; GCN-NEXT: v_add_f32_e32 v13, 1.0, v53 |
| ; GCN-NEXT: v_add_f32_e32 v14, 1.0, v54 |
| ; GCN-NEXT: v_add_f32_e32 v15, 1.0, v55 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v15 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v14 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v13 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v12 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v11 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v10 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v9 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v7 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v6 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v5 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v4 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16f32_to_v32bf16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB22_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; VI-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; VI-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; VI-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; VI-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; VI-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; VI-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; VI-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; VI-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; VI-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; VI-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; VI-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; VI-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; VI-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; VI-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; VI-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; VI-NEXT: .LBB22_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16f32_to_v32bf16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB22_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; GFX9-NEXT: .LBB22_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16f32_to_v32bf16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB22_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14 |
| ; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12 |
| ; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10 |
| ; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8 |
| ; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6 |
| ; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4 |
| ; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2 |
| ; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0 |
| ; GFX11-NEXT: .LBB22_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <16 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <16 x float> %a1 to <32 x bfloat> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x float> %a to <32 x bfloat> |
| br label %end |
| |
| end: |
| %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x bfloat> %phi |
| } |
| |
| define <16 x float> @bitcast_v32bf16_to_v16f32(<32 x bfloat> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32bf16_to_v16f32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v1 |
| ; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v0 |
| ; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v3 |
| ; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v2 |
| ; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v5 |
| ; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v4 |
| ; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v7 |
| ; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v6 |
| ; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v9 |
| ; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v8 |
| ; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v11 |
| ; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v10 |
| ; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v13 |
| ; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v12 |
| ; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v15 |
| ; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v14 |
| ; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v17 |
| ; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v16 |
| ; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v19 |
| ; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v18 |
| ; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v21 |
| ; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v20 |
| ; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v23 |
| ; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v22 |
| ; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25 |
| ; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v24 |
| ; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v27 |
| ; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v26 |
| ; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v29 |
| ; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v28 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47 |
| ; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v46 |
| ; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v30 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB23_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v44 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v42 |
| ; GCN-NEXT: v_alignbit_b32 v0, v0, v45, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v1, v43, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v41 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v40 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v55 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v54 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v52 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v50 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v48 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v38 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v35 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v25 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v23 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v20 |
| ; GCN-NEXT: v_alignbit_b32 v2, v2, v51, 16 |
| ; GCN-NEXT: v_alignbit_b32 v3, v3, v49, 16 |
| ; GCN-NEXT: v_alignbit_b32 v4, v4, v39, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v5, v37, 16 |
| ; GCN-NEXT: v_alignbit_b32 v6, v6, v36, 16 |
| ; GCN-NEXT: v_alignbit_b32 v7, v7, v34, 16 |
| ; GCN-NEXT: v_alignbit_b32 v8, v8, v33, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v9, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v10, v10, v31, 16 |
| ; GCN-NEXT: v_alignbit_b32 v11, v11, v21, 16 |
| ; GCN-NEXT: v_alignbit_b32 v12, v12, v19, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v13, v18, 16 |
| ; GCN-NEXT: v_alignbit_b32 v14, v14, v17, 16 |
| ; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: .LBB23_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB23_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v45 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v44 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v43 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v42 |
| ; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v51 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v41 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v49 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v40 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v39 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v55 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v37 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v54 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v36 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v53 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v34 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v52 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v33 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v50 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v32 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v48 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v31 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v38 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v35 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20 |
| ; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16 |
| ; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16 |
| ; GCN-NEXT: v_alignbit_b32 v4, v7, v6, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v9, v8, 16 |
| ; GCN-NEXT: v_alignbit_b32 v6, v11, v10, 16 |
| ; GCN-NEXT: v_alignbit_b32 v7, v13, v12, 16 |
| ; GCN-NEXT: v_alignbit_b32 v8, v15, v14, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v26, v24, 16 |
| ; GCN-NEXT: v_alignbit_b32 v10, v28, v27, 16 |
| ; GCN-NEXT: v_alignbit_b32 v11, v29, v21, 16 |
| ; GCN-NEXT: v_alignbit_b32 v12, v25, v19, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v23, v18, 16 |
| ; GCN-NEXT: v_alignbit_b32 v14, v22, v17, 16 |
| ; GCN-NEXT: v_alignbit_b32 v15, v20, v16, 16 |
| ; GCN-NEXT: .LBB23_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32bf16_to_v16f32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB23_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v15 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17 |
| ; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v15, 16, 1 |
| ; VI-NEXT: s_movk_i32 s6, 0x7fff |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v15 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v15 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15 |
| ; VI-NEXT: v_alignbit_b32 v15, v15, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v14 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v14, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v14 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v14 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14 |
| ; VI-NEXT: v_alignbit_b32 v14, v14, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v13 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v13, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v13 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v13 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13 |
| ; VI-NEXT: v_alignbit_b32 v13, v13, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v12 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v12, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v12 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v12 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12 |
| ; VI-NEXT: v_alignbit_b32 v12, v12, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v11 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v11, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v11 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v11 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11 |
| ; VI-NEXT: v_alignbit_b32 v11, v11, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v10 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v10, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v10 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v10 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10 |
| ; VI-NEXT: v_alignbit_b32 v10, v10, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v9 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v9, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v9 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v9 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9 |
| ; VI-NEXT: v_alignbit_b32 v9, v9, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v8, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v8 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v8 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8 |
| ; VI-NEXT: v_alignbit_b32 v8, v8, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v7 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v7, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v7 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v7 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7 |
| ; VI-NEXT: v_alignbit_b32 v7, v7, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v6 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v6, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v6 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v6 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6 |
| ; VI-NEXT: v_alignbit_b32 v6, v6, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v5 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v5, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v5 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v5 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5 |
| ; VI-NEXT: v_alignbit_b32 v5, v5, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v4 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v4, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v4 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v4 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4 |
| ; VI-NEXT: v_alignbit_b32 v4, v4, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v3 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v3, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v3 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v3 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; VI-NEXT: v_alignbit_b32 v3, v3, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v2 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v2, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v2 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v2 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 |
| ; VI-NEXT: v_alignbit_b32 v2, v2, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v1 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v1, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v1 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1 |
| ; VI-NEXT: v_alignbit_b32 v1, v1, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v0, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v0 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v0 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0 |
| ; VI-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 |
| ; VI-NEXT: v_alignbit_b32 v0, v0, v16, 16 |
| ; VI-NEXT: .LBB23_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32bf16_to_v16f32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB23_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: s_movk_i32 s6, 0x7fff |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v15, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v15, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc |
| ; GFX9-NEXT: s_mov_b32 s7, 0x7060302 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v14, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v14, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v14 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v13, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v13, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v13 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v12, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v12, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v12 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v11, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v11, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v11 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v11, v11, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v10, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v10, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v10 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v10, v10, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v9, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v9, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v9 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v9, v9, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v8, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v8, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v8 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v8, v8, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v7, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v7, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v7 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v7, v7, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v6, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v6, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v6 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v6, v6, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v5, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v5, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v5 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v5, v5, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v4, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v4, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v4 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v4, v4, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v3, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v3, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v3 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v3, v3, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v2, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v2, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v2 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v2, v2, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v1, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v1, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v1 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v1, v1, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v0, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v0, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v0 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v0, v0, v16, s7 |
| ; GFX9-NEXT: .LBB23_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32bf16_to_v16f32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB23_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16 |
| ; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v18, v16, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v16 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_bfe_u32 v23, v14, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v17, 0x7fff |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v16, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v18, v19, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v19, v23, v14, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v17 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX11-NEXT: v_bfe_u32 v20, v15, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v15 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v20, v20, v15, 0x7fff |
| ; GFX11-NEXT: v_dual_cndmask_b32 v15, v20, v22 :: v_dual_lshlrev_b32 v20, 16, v13 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_perm_b32 v15, v15, v16, 0x7060302 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v17, v21, v18 :: v_dual_add_f32 v18, 0x40c00000, v20 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v14 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v13 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v14, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v12 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_perm_b32 v14, v14, v17, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v17, v13, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_cndmask_b32 v16, v16, v20 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v13, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v11 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_perm_b32 v13, v13, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v21 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX11-NEXT: v_bfe_u32 v18, v12, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v12 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v12, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v12, v18, v22, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v12, v12, v17, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v17, v11, 16, 1 |
| ; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v11, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v19, v10, 16, 1 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v19, v19, v10, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v11, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v10 |
| ; GFX11-NEXT: v_perm_b32 v11, v11, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v10, v19, v22 :: v_dual_lshlrev_b32 v21, 16, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_lshlrev_b32 v19, 16, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX11-NEXT: v_perm_b32 v10, v10, v17, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_add_f32 v8, 0x40c00000, v8 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v18, v8, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v9, 0x40c00000, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v8, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v17, v9, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v9 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v9, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v9, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v8 |
| ; GFX11-NEXT: v_perm_b32 v9, v9, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v8, v18, v22, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v7 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v8, v8, v17, 0x7060302 |
| ; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v19, 0x40c00000, v21 |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GFX11-NEXT: v_bfe_u32 v19, v6, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v7, 0x40c00000, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_add3_u32 v19, v19, v6, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v17, v7, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v7 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v7, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v7, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v5 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v6 |
| ; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_cndmask_b32 v17, v17, v20 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v20, v18, 16, 1 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v6, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v4 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX11-NEXT: v_perm_b32 v7, v7, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v6, v6, v17, 0x7060302 |
| ; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v19 |
| ; GFX11-NEXT: v_add3_u32 v19, v20, v18, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GFX11-NEXT: v_bfe_u32 v22, v17, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v16, v5, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v5 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v5, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v5, v16, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v16, v22, v17, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: v_perm_b32 v5, v5, v18, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v18, v4, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v19, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v4 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v4, 0x7fff |
| ; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v20 :: v_dual_lshlrev_b32 v20, 16, v2 |
| ; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v4, v18, v19, vcc_lo |
| ; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: v_perm_b32 v4, v4, v16, 0x7060302 |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v17, 0x7fff |
| ; GFX11-NEXT: v_bfe_u32 v21, v3, 16, 1 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v18 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v3, 0x7fff |
| ; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v3 |
| ; GFX11-NEXT: v_add3_u32 v21, v22, v18, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v1 |
| ; GFX11-NEXT: v_bfe_u32 v24, v2, 16, 1 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v22 |
| ; GFX11-NEXT: v_add3_u32 v20, v24, v2, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_perm_b32 v3, v3, v17, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v18, v21, v23, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v2 |
| ; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v2, v20, v21 :: v_dual_lshlrev_b32 v23, 16, v0 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) |
| ; GFX11-NEXT: v_add3_u32 v21, v22, v19, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v19 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v23 |
| ; GFX11-NEXT: v_perm_b32 v2, v2, v18, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v19, v21, v22 :: v_dual_add_f32 v0, 0x40c00000, v0 |
| ; GFX11-NEXT: v_bfe_u32 v24, v20, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v20 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v21, v0, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v0 |
| ; GFX11-NEXT: v_add3_u32 v24, v24, v20, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v0, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v23, v1, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 |
| ; GFX11-NEXT: v_add3_u32 v22, v23, v1, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v1, v22, v23, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 |
| ; GFX11-NEXT: v_perm_b32 v1, v1, v19, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v20, v24, v25, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v0, v21, v26, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) |
| ; GFX11-NEXT: v_perm_b32 v0, v0, v20, 0x7060302 |
| ; GFX11-NEXT: .LBB23_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x bfloat> %a, splat (bfloat 0xR40C0) |
| %a2 = bitcast <32 x bfloat> %a1 to <16 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x bfloat> %a to <16 x float> |
| br label %end |
| |
| end: |
| %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x float> %phi |
| } |
| |
| define <64 x i8> @bitcast_v16f32_to_v64i8(<16 x float> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v16f32_to_v64i8: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB24_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v21, v16, v15, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v20, v14, v13, 16 |
| ; GCN-NEXT: v_alignbit_b32 v28, v14, v13, 8 |
| ; GCN-NEXT: v_alignbit_b32 v24, v12, v11, 24 |
| ; GCN-NEXT: v_alignbit_b32 v25, v12, v11, 16 |
| ; GCN-NEXT: v_alignbit_b32 v34, v12, v11, 8 |
| ; GCN-NEXT: v_alignbit_b32 v30, v10, v9, 24 |
| ; GCN-NEXT: v_alignbit_b32 v32, v10, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v48, v10, v9, 8 |
| ; GCN-NEXT: v_alignbit_b32 v36, v8, v7, 24 |
| ; GCN-NEXT: v_alignbit_b32 v37, v8, v7, 16 |
| ; GCN-NEXT: v_alignbit_b32 v54, v8, v7, 8 |
| ; GCN-NEXT: v_alignbit_b32 v49, v6, v5, 24 |
| ; GCN-NEXT: v_alignbit_b32 v51, v6, v5, 16 |
| ; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 8 |
| ; GCN-NEXT: v_alignbit_b32 v40, v4, v3, 24 |
| ; GCN-NEXT: v_alignbit_b32 v41, v4, v3, 16 |
| ; GCN-NEXT: v_alignbit_b32 v58, v4, v3, 8 |
| ; GCN-NEXT: v_alignbit_b32 v44, v2, v1, 24 |
| ; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 16 |
| ; GCN-NEXT: v_alignbit_b32 v60, v2, v1, 8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v50, 8, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v52, 24, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v42, 24, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v56, 24, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v61, 24, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2 |
| ; GCN-NEXT: .LBB24_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB24_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; GCN-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; GCN-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; GCN-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; GCN-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; GCN-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; GCN-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; GCN-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; GCN-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; GCN-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; GCN-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; GCN-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; GCN-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; GCN-NEXT: v_add_f32_e32 v16, 1.0, v16 |
| ; GCN-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v21, v16, v15, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v20, v14, v13, 16 |
| ; GCN-NEXT: v_alignbit_b32 v28, v14, v13, 8 |
| ; GCN-NEXT: v_alignbit_b32 v24, v12, v11, 24 |
| ; GCN-NEXT: v_alignbit_b32 v25, v12, v11, 16 |
| ; GCN-NEXT: v_alignbit_b32 v34, v12, v11, 8 |
| ; GCN-NEXT: v_alignbit_b32 v30, v10, v9, 24 |
| ; GCN-NEXT: v_alignbit_b32 v32, v10, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v48, v10, v9, 8 |
| ; GCN-NEXT: v_alignbit_b32 v36, v8, v7, 24 |
| ; GCN-NEXT: v_alignbit_b32 v37, v8, v7, 16 |
| ; GCN-NEXT: v_alignbit_b32 v54, v8, v7, 8 |
| ; GCN-NEXT: v_alignbit_b32 v49, v6, v5, 24 |
| ; GCN-NEXT: v_alignbit_b32 v51, v6, v5, 16 |
| ; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 8 |
| ; GCN-NEXT: v_alignbit_b32 v40, v4, v3, 24 |
| ; GCN-NEXT: v_alignbit_b32 v41, v4, v3, 16 |
| ; GCN-NEXT: v_alignbit_b32 v58, v4, v3, 8 |
| ; GCN-NEXT: v_alignbit_b32 v44, v2, v1, 24 |
| ; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 16 |
| ; GCN-NEXT: v_alignbit_b32 v60, v2, v1, 8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 24, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v50, 8, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v52, 24, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v42, 24, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v56, 24, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v61, 24, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2 |
| ; GCN-NEXT: .LBB24_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 8, v60 |
| ; GCN-NEXT: s_waitcnt expcnt(3) |
| ; GCN-NEXT: v_or_b32_e32 v60, v1, v18 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v17 |
| ; GCN-NEXT: v_or_b32_e32 v17, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v58 |
| ; GCN-NEXT: v_or_b32_e32 v58, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v63 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v63, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v46 |
| ; GCN-NEXT: v_or_b32_e32 v46, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v44 |
| ; GCN-NEXT: v_and_b32_e32 v45, 0xff, v62 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v61, 24, v61 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v59 |
| ; GCN-NEXT: v_or_b32_e32 v44, v1, v2 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0 |
| ; GCN-NEXT: v_and_b32_e32 v41, 0xff, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v40, 24, v40 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v54 |
| ; GCN-NEXT: v_or_b32_e32 v7, v2, v3 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0 |
| ; GCN-NEXT: v_and_b32_e32 v57, 0xff, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v56, 24, v56 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v47 |
| ; GCN-NEXT: v_or_b32_e32 v54, v3, v4 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 12, v0 |
| ; GCN-NEXT: v_and_b32_e32 v47, 0xff, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v49 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v48 |
| ; GCN-NEXT: v_or_b32_e32 v9, v4, v5 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 16, v0 |
| ; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 24, v42 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v55 |
| ; GCN-NEXT: v_or_b32_e32 v48, v5, v6 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 20, v0 |
| ; GCN-NEXT: v_and_b32_e32 v55, 0xff, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v36 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 8, v34 |
| ; GCN-NEXT: v_or_b32_e32 v34, v6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0 |
| ; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v52 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xff, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 8, v50 |
| ; GCN-NEXT: v_or_b32_e32 v11, v8, v10 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 28, v0 |
| ; GCN-NEXT: v_and_b32_e32 v50, 0xff, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 24, v30 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 8, v28 |
| ; GCN-NEXT: v_or_b32_e32 v13, v10, v12 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 32, v0 |
| ; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v38 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v35 |
| ; GCN-NEXT: v_or_b32_e32 v28, v12, v14 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 36, v0 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 24, v24 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 8, v21 |
| ; GCN-NEXT: v_or_b32_e32 v21, v14, v15 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 40, v0 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xff, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 24, v31 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v29 |
| ; GCN-NEXT: v_or_b32_e32 v29, v15, v16 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 44, v0 |
| ; GCN-NEXT: v_and_b32_e32 v52, 0xff, v20 |
| ; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18 |
| ; GCN-NEXT: v_or_b32_e32 v31, v19, v16 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 48, v0 |
| ; GCN-NEXT: v_and_b32_e32 v42, 0xff, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 24, v26 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v45 |
| ; GCN-NEXT: v_or_b32_e32 v27, v61, v18 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 52, v0 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v45, 0xff, v19 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v41 |
| ; GCN-NEXT: v_or_b32_e32 v40, v40, v19 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v41, 24, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v57 |
| ; GCN-NEXT: v_or_b32_e32 v56, v56, v22 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 60, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v60 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v58 |
| ; GCN-NEXT: v_and_b32_e32 v59, 0xffff, v63 |
| ; GCN-NEXT: v_and_b32_e32 v46, 0xffff, v46 |
| ; GCN-NEXT: v_or_b32_e32 v49, v49, v47 |
| ; GCN-NEXT: v_and_b32_e32 v44, 0xffff, v44 |
| ; GCN-NEXT: v_or_b32_e32 v51, v51, v43 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GCN-NEXT: v_or_b32_e32 v36, v36, v55 |
| ; GCN-NEXT: v_and_b32_e32 v54, 0xffff, v54 |
| ; GCN-NEXT: v_or_b32_e32 v37, v37, v53 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_or_b32_e32 v30, v30, v50 |
| ; GCN-NEXT: v_and_b32_e32 v48, 0xffff, v48 |
| ; GCN-NEXT: v_or_b32_e32 v32, v32, v39 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xffff, v34 |
| ; GCN-NEXT: v_or_b32_e32 v24, v24, v35 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_or_b32_e32 v25, v25, v38 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_or_b32_e32 v20, v20, v52 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28 |
| ; GCN-NEXT: v_or_b32_e32 v26, v26, v42 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21 |
| ; GCN-NEXT: v_or_b32_e32 v33, v33, v45 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_or_b32_e32 v23, v41, v23 |
| ; GCN-NEXT: v_or_b32_e32 v31, v57, v31 |
| ; GCN-NEXT: v_or_b32_e32 v17, v17, v27 |
| ; GCN-NEXT: v_or_b32_e32 v27, v58, v40 |
| ; GCN-NEXT: v_or_b32_e32 v35, v59, v56 |
| ; GCN-NEXT: v_or_b32_e32 v38, v46, v49 |
| ; GCN-NEXT: v_or_b32_e32 v39, v44, v51 |
| ; GCN-NEXT: v_or_b32_e32 v7, v7, v36 |
| ; GCN-NEXT: v_or_b32_e32 v36, v54, v37 |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v30 |
| ; GCN-NEXT: v_or_b32_e32 v30, v48, v32 |
| ; GCN-NEXT: v_or_b32_e32 v24, v34, v24 |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v20 |
| ; GCN-NEXT: v_or_b32_e32 v20, v28, v26 |
| ; GCN-NEXT: v_or_b32_e32 v21, v21, v33 |
| ; GCN-NEXT: v_or_b32_e32 v23, v29, v23 |
| ; GCN-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v17, v1, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v27, v2, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v35, v3, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v38, v4, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v39, v5, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v7, v6, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v36, v8, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v9, v10, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v30, v12, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v24, v14, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v11, v15, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v13, v16, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v20, v18, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v21, v19, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v23, v22, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v16f32_to_v64i8: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr49 |
| ; VI-NEXT: ; implicit-def: $vgpr48 |
| ; VI-NEXT: ; implicit-def: $vgpr39 |
| ; VI-NEXT: ; implicit-def: $vgpr38 |
| ; VI-NEXT: ; implicit-def: $vgpr37 |
| ; VI-NEXT: ; implicit-def: $vgpr36 |
| ; VI-NEXT: ; implicit-def: $vgpr35 |
| ; VI-NEXT: ; implicit-def: $vgpr34 |
| ; VI-NEXT: ; implicit-def: $vgpr33 |
| ; VI-NEXT: ; implicit-def: $vgpr32 |
| ; VI-NEXT: ; implicit-def: $vgpr31 |
| ; VI-NEXT: ; implicit-def: $vgpr30 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB24_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v16 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8] |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v16 |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; VI-NEXT: v_mov_b32_e32 v26, v22 |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4] |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2] |
| ; VI-NEXT: .LBB24_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB24_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e32 v16, 1.0, v16 |
| ; VI-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; VI-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; VI-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; VI-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; VI-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; VI-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; VI-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; VI-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; VI-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; VI-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; VI-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4] |
| ; VI-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; VI-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8] |
| ; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2] |
| ; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 8, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: .LBB24_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v29 |
| ; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v22 |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v22, v27, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v28 |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v62 |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61 |
| ; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v59 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v57 |
| ; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v56 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v46 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44 |
| ; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v43 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v24 |
| ; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v55 |
| ; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v54 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v52 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v50 |
| ; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v49 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; VI-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37 |
| ; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; VI-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32 |
| ; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; VI-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v26 |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 60, v0 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v16f32_to_v64i8: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB24_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v16 |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_mov_b32_e32 v26, v23 |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: .LBB24_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB24_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: .LBB24_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v29 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v23 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v23, v27, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28 |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v22 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v60 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v59 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v57 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v56 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v46 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v21 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v55 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v50 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v49 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56 |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v26 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60 |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v16f32_to_v64i8: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB24_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB24_2: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB24_4 |
| ; GFX11-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-NEXT: v_dual_add_f32 v8, 1.0, v8 :: v_dual_add_f32 v15, 1.0, v15 |
| ; GFX11-NEXT: v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v7, 1.0, v7 |
| ; GFX11-NEXT: v_dual_add_f32 v10, 1.0, v10 :: v_dual_add_f32 v13, 1.0, v13 |
| ; GFX11-NEXT: v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v9, 1.0, v9 |
| ; GFX11-NEXT: v_dual_add_f32 v12, 1.0, v12 :: v_dual_add_f32 v11, 1.0, v11 |
| ; GFX11-NEXT: v_dual_add_f32 v6, 1.0, v6 :: v_dual_add_f32 v5, 1.0, v5 |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_dual_add_f32 v4, 1.0, v4 :: v_dual_add_f32 v3, 1.0, v3 |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_dual_add_f32 v2, 1.0, v2 :: v_dual_add_f32 v1, 1.0, v1 |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB24_4: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v96, 0xff, v96 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v86, 0xff, v86 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v87 |
| ; GFX11-NEXT: v_or_b32_e32 v24, v96, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v25 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v86, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84 |
| ; GFX11-NEXT: v_and_b32_e32 v83, 0xff, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v82 |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v80, 8, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v71, 8, v71 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v24 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v81 |
| ; GFX11-NEXT: v_and_b32_e32 v70, 0xff, v70 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v83, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v25, v80 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v71 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v70, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v22 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v66 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v65 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v55 |
| ; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v23, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v25, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v52 |
| ; GFX11-NEXT: v_lshlrev_b16 v20, 8, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v50 |
| ; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v48, 8, v48 |
| ; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v39 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v23, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v25, v49 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v48 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v39, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v38 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v37 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v35 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v34 |
| ; GFX11-NEXT: v_lshlrev_b16 v18, 8, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_lshlrev_b16 v69, 8, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v68, 0xff, v68 |
| ; GFX11-NEXT: v_lshlrev_b16 v67, 8, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v23, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v32 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v30 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v29 |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v26, 8, v26 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v69 |
| ; GFX11-NEXT: v_or_b32_e32 v67, v68, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v23, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v25, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v67 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; GFX11-NEXT: s_clause 0x3 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off |
| ; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <16 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <16 x float> %a1 to <64 x i8> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <16 x float> %a to <64 x i8> |
| br label %end |
| |
| end: |
| %phi = phi <64 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <64 x i8> %phi |
| } |
| |
| define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v64i8_to_v16f32: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_mov_b32_e32 v38, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v31, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104 |
| ; GCN-NEXT: s_waitcnt expcnt(3) |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:100 |
| ; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:96 |
| ; GCN-NEXT: s_waitcnt expcnt(2) |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:92 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88 |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:84 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76 |
| ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:72 |
| ; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:68 |
| ; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64 |
| ; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:60 |
| ; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56 |
| ; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:52 |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8 |
| ; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:132 |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:128 |
| ; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:124 |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:120 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v52 |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 |
| ; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:36 |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 |
| ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:28 |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:24 |
| ; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:20 |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v5 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v7 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v9 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v11 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v13 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v15 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v17 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v19 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v21 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v23 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v25 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v27 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v45 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 8, v44 |
| ; GCN-NEXT: s_waitcnt vmcnt(10) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v59, 24, v59 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v58 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v57, 24, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v58, 24, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v56, 8, v14 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v12 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v10 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v6 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v4 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v2 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:44 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:12 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116 |
| ; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0 |
| ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v46, 24, v46 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB25_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v42 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v41 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v40 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v55 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v35 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v36 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v37 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v38 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v50 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v36, 0xff, v54 |
| ; GCN-NEXT: v_and_b32_e32 v37, 0xff, v53 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xff, v52 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v49 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v48 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v39 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v63 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v62 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v61 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v60 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v21 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v32 |
| ; GCN-NEXT: v_or_b32_e32 v17, v33, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v34 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v20, v20, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v23, v24, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26 |
| ; GCN-NEXT: v_or_b32_e32 v26, v28, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v30 |
| ; GCN-NEXT: v_or_b32_e32 v27, v35, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_or_b32_e32 v29, v36, v44 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v37 |
| ; GCN-NEXT: v_or_b32_e32 v31, v38, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GCN-NEXT: v_or_b32_e32 v18, v18, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v7, v32, v7 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v6, v32, v6 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v5, v32, v5 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v32, v4 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v32, v21 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v22, v32, v22 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v24, v32, v24 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v28, v32, v28 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27 |
| ; GCN-NEXT: v_or_b32_e32 v25, v59, v25 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_or_b32_e32 v30, v57, v30 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31 |
| ; GCN-NEXT: v_or_b32_e32 v32, v58, v8 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v9 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v34, v8, v10 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xffff, v11 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v8, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v14, v8, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v16, v8, v16 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GCN-NEXT: v_or_b32_e32 v19, v46, v19 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v7 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v6 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v5 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; GCN-NEXT: v_or_b32_e32 v4, v17, v21 |
| ; GCN-NEXT: v_or_b32_e32 v5, v20, v22 |
| ; GCN-NEXT: v_or_b32_e32 v6, v23, v24 |
| ; GCN-NEXT: v_or_b32_e32 v7, v26, v28 |
| ; GCN-NEXT: v_or_b32_e32 v8, v27, v25 |
| ; GCN-NEXT: v_or_b32_e32 v9, v29, v30 |
| ; GCN-NEXT: v_or_b32_e32 v10, v31, v32 |
| ; GCN-NEXT: v_or_b32_e32 v11, v33, v34 |
| ; GCN-NEXT: v_or_b32_e32 v12, v35, v12 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; GCN-NEXT: v_or_b32_e32 v14, v15, v16 |
| ; GCN-NEXT: v_or_b32_e32 v15, v18, v19 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; kill: killed $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: .LBB25_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB25_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GCN-NEXT: v_or_b32_e32 v0, v42, v0 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GCN-NEXT: v_or_b32_e32 v1, v41, v1 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GCN-NEXT: v_or_b32_e32 v2, v40, v2 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GCN-NEXT: v_or_b32_e32 v3, v55, v3 |
| ; GCN-NEXT: s_movk_i32 s7, 0x300 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35 |
| ; GCN-NEXT: s_mov_b32 s6, 0x3000000 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v50 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v25 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v54 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v53 |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v52 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v49 |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v48 |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v39 |
| ; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v63 |
| ; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v62 |
| ; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v61 |
| ; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v60 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19 |
| ; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v51 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v7 |
| ; GCN-NEXT: v_or_b32_e32 v6, v17, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v9 |
| ; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v11 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v13 |
| ; GCN-NEXT: v_or_b32_e32 v13, v29, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15 |
| ; GCN-NEXT: v_or_b32_e32 v15, v27, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18 |
| ; GCN-NEXT: v_or_b32_e32 v17, v44, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v22 |
| ; GCN-NEXT: v_or_b32_e32 v20, v45, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v23 |
| ; GCN-NEXT: v_or_b32_e32 v23, v56, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v25, v25, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v30 |
| ; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v27, v27, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v32 |
| ; GCN-NEXT: v_or_b32_e32 v29, v43, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GCN-NEXT: v_or_b32_e32 v19, v47, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v34 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v31, v31, v36 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v8, v32, v8 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v5, v32, v5 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v32, v4 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v7, v32, v7 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v10, v32, v10 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v32, v12 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v14, v32, v14 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15 |
| ; GCN-NEXT: v_or_b32_e32 v16, v59, v16 |
| ; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v17 |
| ; GCN-NEXT: v_or_b32_e32 v18, v57, v18 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20 |
| ; GCN-NEXT: v_or_b32_e32 v22, v58, v22 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, s7, v23 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v24, v32, v24 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v26, v32, v26 |
| ; GCN-NEXT: v_add_i32_e32 v27, vcc, s7, v27 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v28, v32, v28 |
| ; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v32, v21 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 0x300, v19 |
| ; GCN-NEXT: v_or_b32_e32 v30, v46, v30 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GCN-NEXT: v_or_b32_e32 v0, v31, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v8, v1 |
| ; GCN-NEXT: v_or_b32_e32 v2, v5, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; GCN-NEXT: v_or_b32_e32 v4, v7, v6 |
| ; GCN-NEXT: v_or_b32_e32 v5, v10, v9 |
| ; GCN-NEXT: v_or_b32_e32 v6, v12, v11 |
| ; GCN-NEXT: v_or_b32_e32 v7, v14, v13 |
| ; GCN-NEXT: v_or_b32_e32 v8, v16, v15 |
| ; GCN-NEXT: v_or_b32_e32 v9, v18, v17 |
| ; GCN-NEXT: v_or_b32_e32 v10, v22, v20 |
| ; GCN-NEXT: v_or_b32_e32 v11, v24, v23 |
| ; GCN-NEXT: v_or_b32_e32 v12, v26, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v28, v27 |
| ; GCN-NEXT: v_or_b32_e32 v14, v21, v29 |
| ; GCN-NEXT: v_or_b32_e32 v15, v30, v19 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x3000000, v14 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x3000000, v15 |
| ; GCN-NEXT: .LBB25_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v64i8_to_v16f32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132 |
| ; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72 |
| ; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80 |
| ; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88 |
| ; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96 |
| ; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104 |
| ; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112 |
| ; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120 |
| ; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128 |
| ; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20 |
| ; VI-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12 |
| ; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100 |
| ; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68 |
| ; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36 |
| ; VI-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4 |
| ; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v26, 8, v23 |
| ; VI-NEXT: v_lshlrev_b16_e32 v28, 8, v25 |
| ; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v27 |
| ; VI-NEXT: v_lshlrev_b16_e32 v62, 8, v29 |
| ; VI-NEXT: v_lshlrev_b16_e32 v18, 8, v17 |
| ; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v19 |
| ; VI-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116 |
| ; VI-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124 |
| ; VI-NEXT: v_lshlrev_b16_e32 v31, 8, v1 |
| ; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v3 |
| ; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v5 |
| ; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v7 |
| ; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v9 |
| ; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v11 |
| ; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v13 |
| ; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v15 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54 |
| ; VI-NEXT: v_lshlrev_b16_e32 v61, 8, v0 |
| ; VI-NEXT: v_lshlrev_b16_e32 v59, 8, v2 |
| ; VI-NEXT: v_lshlrev_b16_e32 v58, 8, v4 |
| ; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v6 |
| ; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v8 |
| ; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v10 |
| ; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v12 |
| ; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v14 |
| ; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v24 |
| ; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v30 |
| ; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v53 |
| ; VI-NEXT: s_waitcnt vmcnt(13) |
| ; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v42 |
| ; VI-NEXT: s_waitcnt vmcnt(12) |
| ; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v44 |
| ; VI-NEXT: s_waitcnt vmcnt(11) |
| ; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v47 |
| ; VI-NEXT: s_waitcnt vmcnt(10) |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v57 |
| ; VI-NEXT: s_waitcnt vmcnt(9) |
| ; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v60 |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_lshlrev_b16_e32 v57, 8, v63 |
| ; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108 |
| ; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92 |
| ; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84 |
| ; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76 |
| ; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60 |
| ; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44 |
| ; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB25_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr30 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr23 |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(7) |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr31 |
| ; VI-NEXT: ; implicit-def: $vgpr32 |
| ; VI-NEXT: ; implicit-def: $vgpr33 |
| ; VI-NEXT: ; implicit-def: $vgpr34 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr38 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr35 |
| ; VI-NEXT: ; implicit-def: $vgpr36 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr37 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr49 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr48 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr39 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: .LBB25_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB25_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: v_mov_b32_e32 v15, 0x300 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v51 |
| ; VI-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v9, v9, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v24 |
| ; VI-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v10, v10, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v30 |
| ; VI-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v11, v11, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v60 |
| ; VI-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v12, v12, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v47 |
| ; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v13, v13, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v42 |
| ; VI-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v14, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v17, 3, v17 |
| ; VI-NEXT: v_or_b32_sdwa v17, v57, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: s_waitcnt vmcnt(7) |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v1 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v0, 0x300, v0 |
| ; VI-NEXT: v_add_u16_sdwa v1, v1, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; VI-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; VI-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v4, v4, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; VI-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v5, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; VI-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v6, v6, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; VI-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v7, v7, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v53 |
| ; VI-NEXT: v_or_b32_sdwa v16, v23, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v16, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v8 |
| ; VI-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v8, v8, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: v_add_u16_sdwa v15, v17, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v3 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; VI-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v4, 0x300, v4 |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; VI-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v5, 0x300, v5 |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; VI-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v6, 0x300, v6 |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; VI-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v7, 0x300, v7 |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v52 |
| ; VI-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v8, 0x300, v8 |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v50 |
| ; VI-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v9, 0x300, v9 |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v49 |
| ; VI-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v10, 0x300, v10 |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v63 |
| ; VI-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v11, 0x300, v11 |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v48 |
| ; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v12, 0x300, v12 |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v44 |
| ; VI-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v13, 0x300, v13 |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v39 |
| ; VI-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v14, 0x300, v14 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v16 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v19 |
| ; VI-NEXT: v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v16, 0x300, v16 |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: .LBB25_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v64i8_to_v16f32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132 |
| ; GFX9-NEXT: s_nop 0 |
| ; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64 |
| ; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72 |
| ; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80 |
| ; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88 |
| ; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96 |
| ; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104 |
| ; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112 |
| ; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120 |
| ; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128 |
| ; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20 |
| ; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12 |
| ; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100 |
| ; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68 |
| ; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36 |
| ; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v21 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v26, 8, v23 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v28, 8, v25 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v27 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v62, 8, v29 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v18, 8, v17 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v19 |
| ; GFX9-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116 |
| ; GFX9-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v31, 8, v1 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v3 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v5 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v7 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v9 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v11 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v13 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v15 |
| ; GFX9-NEXT: s_waitcnt vmcnt(25) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54 |
| ; GFX9-NEXT: s_waitcnt vmcnt(24) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v61, 8, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(23) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v59, 8, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(22) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v58, 8, v4 |
| ; GFX9-NEXT: s_waitcnt vmcnt(21) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(20) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v8 |
| ; GFX9-NEXT: s_waitcnt vmcnt(19) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v10 |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v12 |
| ; GFX9-NEXT: s_waitcnt vmcnt(17) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v14 |
| ; GFX9-NEXT: s_waitcnt vmcnt(16) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v24 |
| ; GFX9-NEXT: s_waitcnt vmcnt(15) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v30 |
| ; GFX9-NEXT: s_waitcnt vmcnt(14) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v53 |
| ; GFX9-NEXT: s_waitcnt vmcnt(13) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v42 |
| ; GFX9-NEXT: s_waitcnt vmcnt(12) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v44 |
| ; GFX9-NEXT: s_waitcnt vmcnt(11) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v47 |
| ; GFX9-NEXT: s_waitcnt vmcnt(10) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v57 |
| ; GFX9-NEXT: s_waitcnt vmcnt(9) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v60 |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v57, 8, v63 |
| ; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108 |
| ; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92 |
| ; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84 |
| ; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76 |
| ; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60 |
| ; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52 |
| ; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44 |
| ; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB25_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: .LBB25_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB25_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_movk_i32 s6, 0x300 |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_add_u16_e32 v9, 3, v51 |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_add_u16_e32 v10, 3, v24 |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v11, 3, v30 |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v12, 3, v60 |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v13, 3, v47 |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v14, 3, v42 |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v15, 3, v53 |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_add_u16_e32 v1, 3, v1 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GFX9-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v16, 3, v17 |
| ; GFX9-NEXT: v_or_b32_sdwa v16, v57, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_add_u16_e32 v8, 3, v8 |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3 |
| ; GFX9-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4 |
| ; GFX9-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5 |
| ; GFX9-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6 |
| ; GFX9-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7 |
| ; GFX9-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; GFX9-NEXT: v_add_u16_e32 v8, 3, v52 |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8 |
| ; GFX9-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; GFX9-NEXT: v_add_u16_e32 v9, 3, v50 |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9 |
| ; GFX9-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; GFX9-NEXT: v_add_u16_e32 v10, 3, v49 |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10 |
| ; GFX9-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; GFX9-NEXT: v_add_u16_e32 v11, 3, v63 |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11 |
| ; GFX9-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GFX9-NEXT: v_add_u16_e32 v12, 3, v48 |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12 |
| ; GFX9-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; GFX9-NEXT: v_add_u16_e32 v13, 3, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13 |
| ; GFX9-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; GFX9-NEXT: v_add_u16_e32 v14, 3, v39 |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14 |
| ; GFX9-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; GFX9-NEXT: v_add_u16_e32 v15, 3, v19 |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15 |
| ; GFX9-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; GFX9-NEXT: .LBB25_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v64i8_to_v16f32: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12 |
| ; GFX11-NEXT: v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8 |
| ; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4 |
| ; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0 |
| ; GFX11-NEXT: s_clause 0x1f |
| ; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:128 |
| ; GFX11-NEXT: scratch_load_u16 v39, off, s32 offset:124 |
| ; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:120 |
| ; GFX11-NEXT: scratch_load_u16 v48, off, s32 offset:116 |
| ; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:112 |
| ; GFX11-NEXT: scratch_load_u16 v49, off, s32 offset:108 |
| ; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:104 |
| ; GFX11-NEXT: scratch_load_u16 v50, off, s32 offset:100 |
| ; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:96 |
| ; GFX11-NEXT: scratch_load_u16 v51, off, s32 offset:92 |
| ; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:88 |
| ; GFX11-NEXT: scratch_load_b32 v12, off, s32 offset:132 |
| ; GFX11-NEXT: scratch_load_u16 v14, off, s32 |
| ; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:8 |
| ; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:16 |
| ; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:24 |
| ; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:32 |
| ; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:40 |
| ; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:48 |
| ; GFX11-NEXT: scratch_load_u16 v128, off, s32 offset:56 |
| ; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:64 |
| ; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:72 |
| ; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:80 |
| ; GFX11-NEXT: scratch_load_u16 v52, off, s32 offset:84 |
| ; GFX11-NEXT: scratch_load_u16 v53, off, s32 offset:76 |
| ; GFX11-NEXT: scratch_load_u16 v54, off, s32 offset:68 |
| ; GFX11-NEXT: scratch_load_u16 v55, off, s32 offset:60 |
| ; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:52 |
| ; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:44 |
| ; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:36 |
| ; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:28 |
| ; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:20 |
| ; GFX11-NEXT: s_clause 0x1 |
| ; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:12 |
| ; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:4 |
| ; GFX11-NEXT: v_lshlrev_b16 v118, 8, v1 |
| ; GFX11-NEXT: v_lshlrev_b16 v119, 8, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v117, 8, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v114, 8, v7 |
| ; GFX11-NEXT: v_lshlrev_b16 v115, 8, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v116, 8, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v113, 8, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v100, 8, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v101, 8, v17 |
| ; GFX11-NEXT: v_lshlrev_b16 v102, 8, v19 |
| ; GFX11-NEXT: v_lshlrev_b16 v103, 8, v21 |
| ; GFX11-NEXT: v_lshlrev_b16 v112, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v87, 8, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v96, 8, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v97, 8, v29 |
| ; GFX11-NEXT: s_waitcnt vmcnt(33) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(31) |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v2 |
| ; GFX11-NEXT: s_waitcnt vmcnt(29) |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v4 |
| ; GFX11-NEXT: s_waitcnt vmcnt(27) |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v6 |
| ; GFX11-NEXT: s_waitcnt vmcnt(25) |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v8 |
| ; GFX11-NEXT: s_waitcnt vmcnt(22) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12 |
| ; GFX11-NEXT: s_waitcnt vmcnt(21) |
| ; GFX11-NEXT: v_lshlrev_b16 v98, 8, v14 |
| ; GFX11-NEXT: s_waitcnt vmcnt(20) |
| ; GFX11-NEXT: v_lshlrev_b16 v99, 8, v65 |
| ; GFX11-NEXT: s_waitcnt vmcnt(19) |
| ; GFX11-NEXT: v_lshlrev_b16 v81, 8, v66 |
| ; GFX11-NEXT: s_waitcnt vmcnt(18) |
| ; GFX11-NEXT: v_lshlrev_b16 v82, 8, v67 |
| ; GFX11-NEXT: s_waitcnt vmcnt(17) |
| ; GFX11-NEXT: v_lshlrev_b16 v83, 8, v83 |
| ; GFX11-NEXT: s_waitcnt vmcnt(16) |
| ; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84 |
| ; GFX11-NEXT: s_waitcnt vmcnt(15) |
| ; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85 |
| ; GFX11-NEXT: s_waitcnt vmcnt(14) |
| ; GFX11-NEXT: v_lshlrev_b16 v27, 8, v128 |
| ; GFX11-NEXT: s_waitcnt vmcnt(13) |
| ; GFX11-NEXT: v_lshlrev_b16 v29, 8, v129 |
| ; GFX11-NEXT: s_waitcnt vmcnt(12) |
| ; GFX11-NEXT: v_lshlrev_b16 v65, 8, v130 |
| ; GFX11-NEXT: s_waitcnt vmcnt(11) |
| ; GFX11-NEXT: v_lshlrev_b16 v66, 8, v131 |
| ; GFX11-NEXT: v_lshlrev_b16 v67, 8, v10 |
| ; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB25_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB25_4 |
| ; GFX11-NEXT: .LBB25_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB25_3: ; %cmp.false |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v37 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v118 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v119 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v35 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v117 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v6, v113 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v114 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v115 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v116 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v101 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v38 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v102 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v103 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v112 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v100 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v1 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v4, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v6, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v7, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v9, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v30 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v86 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v71 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v70 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v68 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v87 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v96 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v97 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v98 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v99 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v81 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v82 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v8, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v10, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v12, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v14, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v55 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v53 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v52 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v50 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v49 |
| ; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v48 |
| ; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v39 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v27 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v29 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v65 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v66 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v18, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v22, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v24, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v13, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v15, v16 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v17, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v19, v20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr16 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr118 |
| ; GFX11-NEXT: ; implicit-def: $vgpr119 |
| ; GFX11-NEXT: ; implicit-def: $vgpr117 |
| ; GFX11-NEXT: ; implicit-def: $vgpr114 |
| ; GFX11-NEXT: ; implicit-def: $vgpr115 |
| ; GFX11-NEXT: ; implicit-def: $vgpr116 |
| ; GFX11-NEXT: ; implicit-def: $vgpr113 |
| ; GFX11-NEXT: ; implicit-def: $vgpr100 |
| ; GFX11-NEXT: ; implicit-def: $vgpr101 |
| ; GFX11-NEXT: ; implicit-def: $vgpr102 |
| ; GFX11-NEXT: ; implicit-def: $vgpr103 |
| ; GFX11-NEXT: ; implicit-def: $vgpr112 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr97 |
| ; GFX11-NEXT: ; implicit-def: $vgpr98 |
| ; GFX11-NEXT: ; implicit-def: $vgpr99 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB25_2 |
| ; GFX11-NEXT: .LBB25_4: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u16 v0, v31, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v32, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v33, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v34, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v35, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v5, v36, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v6, v37, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v118, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v119, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v117, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v114, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v115, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v116, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v113, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v7, v16, 3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v5 |
| ; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v6, v38, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v8, v18, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v9, v20, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v10, v22, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v101, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v100, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v102, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v103, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v112, v10 |
| ; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7 |
| ; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8 |
| ; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9 |
| ; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v3, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v5, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v7, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v9, v10 |
| ; GFX11-NEXT: v_add_nc_u16 v6, v24, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v7, v26, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v8, v28, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v9, v30, 3 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_add_nc_u16 v10, v86, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v11, v80, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v12, v71, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v13, v70, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v14, v69, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v15, v68, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v87, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v96, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v97, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v98, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v99, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v81, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v82, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v83, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v84, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v85, v15 |
| ; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7 |
| ; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8 |
| ; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9 |
| ; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10 |
| ; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11 |
| ; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12 |
| ; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13 |
| ; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14 |
| ; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v8, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v10, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v12, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v14, v15 |
| ; GFX11-NEXT: v_add_nc_u16 v11, v64, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v12, v55, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v13, v54, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v14, v53, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v15, v52, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v16, v51, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v18, v50, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v20, v49, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v22, v48, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v24, v39, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v27, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v29, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v65, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v66, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v67, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v19, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v21, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v23, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v25, v24 |
| ; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11 |
| ; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12 |
| ; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13 |
| ; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14 |
| ; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15 |
| ; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16 |
| ; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17 |
| ; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18 |
| ; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19 |
| ; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v13, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v15, v16 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v17, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v19, v20 |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <64 x i8> %a, splat (i8 3) |
| %a2 = bitcast <64 x i8> %a1 to <16 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <64 x i8> %a to <16 x float> |
| br label %end |
| |
| end: |
| %phi = phi <16 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <16 x float> %phi |
| } |
| |
| define <8 x double> @bitcast_v8i64_to_v8f64(<8 x i64> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8i64_to_v8f64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB26_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; GCN-NEXT: .LBB26_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8i64_to_v8f64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB26_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0 |
| ; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; VI-NEXT: .LBB26_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8i64_to_v8f64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB26_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc |
| ; GFX9-NEXT: .LBB26_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8i64_to_v8f64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB26_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo |
| ; GFX11-NEXT: .LBB26_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <8 x i64> %a, splat (i64 3) |
| %a2 = bitcast <8 x i64> %a1 to <8 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x i64> %a to <8 x double> |
| br label %end |
| |
| end: |
| %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x double> %phi |
| } |
| |
| define <8 x i64> @bitcast_v8f64_to_v8i64(<8 x double> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8f64_to_v8i64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB27_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GCN-NEXT: .LBB27_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8f64_to_v8i64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB27_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; VI-NEXT: .LBB27_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8f64_to_v8i64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB27_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX9-NEXT: .LBB27_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8f64_to_v8i64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB27_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX11-NEXT: .LBB27_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <8 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <8 x double> %a1 to <8 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x double> %a to <8 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x i64> %phi |
| } |
| |
| define <32 x i16> @bitcast_v8i64_to_v32i16(<8 x i64> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8i64_to_v32i16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v30, v15 |
| ; GCN-NEXT: v_mov_b32_e32 v28, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v26, v13 |
| ; GCN-NEXT: v_mov_b32_e32 v24, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v22, v11 |
| ; GCN-NEXT: v_mov_b32_e32 v20, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v18, v9 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v14, v7 |
| ; GCN-NEXT: v_mov_b32_e32 v12, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v10, v5 |
| ; GCN-NEXT: v_mov_b32_e32 v8, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v6, v3 |
| ; GCN-NEXT: v_mov_b32_e32 v4, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v2, v1 |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB28_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_alignbit_b32 v29, v30, v28, 16 |
| ; GCN-NEXT: v_alignbit_b32 v25, v26, v24, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v22, v20, 16 |
| ; GCN-NEXT: v_alignbit_b32 v17, v18, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v10, v8, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v26 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v18 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2 |
| ; GCN-NEXT: .LBB28_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB28_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; GCN-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; GCN-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; GCN-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc |
| ; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v32 |
| ; GCN-NEXT: v_addc_u32_e32 v18, vcc, 0, v18, vcc |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20 |
| ; GCN-NEXT: v_addc_u32_e32 v22, vcc, 0, v22, vcc |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v24 |
| ; GCN-NEXT: v_addc_u32_e32 v26, vcc, 0, v26, vcc |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28 |
| ; GCN-NEXT: v_addc_u32_e32 v30, vcc, 0, v30, vcc |
| ; GCN-NEXT: v_alignbit_b32 v29, v30, v28, 16 |
| ; GCN-NEXT: v_alignbit_b32 v25, v26, v24, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v22, v20, 16 |
| ; GCN-NEXT: v_alignbit_b32 v17, v18, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v14, v12, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v10, v8, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v6, v4, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v0, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v30 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v26 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v18 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v2 |
| ; GCN-NEXT: .LBB28_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_mov_b32_e32 v16, v32 |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8i64_to_v32i16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB28_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0 |
| ; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; VI-NEXT: .LBB28_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8i64_to_v32i16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB28_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc |
| ; GFX9-NEXT: .LBB28_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8i64_to_v32i16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB28_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo |
| ; GFX11-NEXT: .LBB28_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <8 x i64> %a, splat (i64 3) |
| %a2 = bitcast <8 x i64> %a1 to <32 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x i64> %a to <32 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x i16> %phi |
| } |
| |
| define <8 x i64> @bitcast_v32i16_to_v8i64(<32 x i16> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32i16_to_v8i64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v38, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v31, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v0 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB29_3 |
| ; GCN-NEXT: ; %bb.1: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB29_4 |
| ; GCN-NEXT: .LBB29_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; GCN-NEXT: .LBB29_3: ; %cmp.false |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v31 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v32 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v54 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v55 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v33 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v34 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v35 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v36 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v37 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v38 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v30 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v39 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v48 |
| ; GCN-NEXT: v_or_b32_e32 v4, v4, v49 |
| ; GCN-NEXT: v_or_b32_e32 v5, v5, v50 |
| ; GCN-NEXT: v_or_b32_e32 v6, v6, v51 |
| ; GCN-NEXT: v_or_b32_e32 v7, v7, v52 |
| ; GCN-NEXT: v_or_b32_e32 v8, v8, v17 |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v19 |
| ; GCN-NEXT: v_or_b32_e32 v10, v10, v21 |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v23 |
| ; GCN-NEXT: v_or_b32_e32 v12, v12, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v27 |
| ; GCN-NEXT: v_or_b32_e32 v14, v14, v29 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v53 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB29_2 |
| ; GCN-NEXT: .LBB29_4: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31 |
| ; GCN-NEXT: s_mov_b32 s6, 0x30000 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_or_b32_e32 v0, v54, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v55, v1 |
| ; GCN-NEXT: v_or_b32_e32 v2, v39, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v48, v3 |
| ; GCN-NEXT: v_or_b32_e32 v4, v49, v4 |
| ; GCN-NEXT: v_or_b32_e32 v5, v50, v5 |
| ; GCN-NEXT: v_or_b32_e32 v6, v51, v6 |
| ; GCN-NEXT: v_or_b32_e32 v7, v52, v7 |
| ; GCN-NEXT: v_or_b32_e32 v8, v17, v8 |
| ; GCN-NEXT: v_or_b32_e32 v9, v19, v9 |
| ; GCN-NEXT: v_or_b32_e32 v10, v21, v10 |
| ; GCN-NEXT: v_or_b32_e32 v11, v23, v11 |
| ; GCN-NEXT: v_or_b32_e32 v12, v25, v12 |
| ; GCN-NEXT: v_or_b32_e32 v13, v27, v13 |
| ; GCN-NEXT: v_or_b32_e32 v14, v29, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v53, v15 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v13 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v14 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x30000, v15 |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32i16_to_v8i64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB29_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v17, 3 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v15 |
| ; VI-NEXT: v_add_u16_sdwa v15, v15, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v14 |
| ; VI-NEXT: v_add_u16_sdwa v14, v14, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v14, v16, v14 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v13 |
| ; VI-NEXT: v_add_u16_sdwa v13, v13, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v13, v16, v13 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v12 |
| ; VI-NEXT: v_add_u16_sdwa v12, v12, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v12, v16, v12 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v11 |
| ; VI-NEXT: v_add_u16_sdwa v11, v11, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v11, v16, v11 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v10 |
| ; VI-NEXT: v_add_u16_sdwa v10, v10, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v10, v16, v10 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v9 |
| ; VI-NEXT: v_add_u16_sdwa v9, v9, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v9, v16, v9 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v8 |
| ; VI-NEXT: v_add_u16_sdwa v8, v8, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v8, v16, v8 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v7 |
| ; VI-NEXT: v_add_u16_sdwa v7, v7, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v7, v16, v7 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v6 |
| ; VI-NEXT: v_add_u16_sdwa v6, v6, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v6, v16, v6 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v5 |
| ; VI-NEXT: v_add_u16_sdwa v5, v5, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v5, v16, v5 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v4 |
| ; VI-NEXT: v_add_u16_sdwa v4, v4, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v4, v16, v4 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v3 |
| ; VI-NEXT: v_add_u16_sdwa v3, v3, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v3, v16, v3 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v2 |
| ; VI-NEXT: v_add_u16_sdwa v2, v2, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v2, v16, v2 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v1 |
| ; VI-NEXT: v_add_u16_sdwa v1, v1, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v1, v16, v1 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v0 |
| ; VI-NEXT: v_add_u16_sdwa v0, v0, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v0, v16, v0 |
| ; VI-NEXT: .LBB29_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32i16_to_v8i64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB29_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB29_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32i16_to_v8i64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB29_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: .LBB29_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <32 x i16> %a, splat (i16 3) |
| %a2 = bitcast <32 x i16> %a1 to <8 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x i16> %a to <8 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x i64> %phi |
| } |
| |
| define <32 x half> @bitcast_v8i64_to_v32f16(<8 x i64> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8i64_to_v32f16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_mov_b32_e32 v34, v15 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v13 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v38, v11 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v48, v9 |
| ; GCN-NEXT: v_mov_b32_e32 v39, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v50, v7 |
| ; GCN-NEXT: v_mov_b32_e32 v49, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v52, v5 |
| ; GCN-NEXT: v_mov_b32_e32 v51, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v54, v3 |
| ; GCN-NEXT: v_mov_b32_e32 v53, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v55, v1 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v0 |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB30_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v33 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v36 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v35 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v38 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v37 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v48 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v39 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v50 |
| ; GCN-NEXT: s_waitcnt expcnt(6) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v49 |
| ; GCN-NEXT: s_waitcnt expcnt(5) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v52 |
| ; GCN-NEXT: s_waitcnt expcnt(4) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v51 |
| ; GCN-NEXT: s_waitcnt expcnt(3) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v54 |
| ; GCN-NEXT: s_waitcnt expcnt(2) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v44, 16, v53 |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v45, 16, v55 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v46, 16, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v53 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v55 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v40 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v41 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v42 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v43 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v44 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v45 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v46 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v32 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: .LBB30_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB30_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v32 |
| ; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v55, vcc |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v53 |
| ; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v54, vcc |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v51 |
| ; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v52, vcc |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v49 |
| ; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v50, vcc |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v39 |
| ; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v48, vcc |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v37 |
| ; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v38, vcc |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v35 |
| ; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v36, vcc |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v33 |
| ; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v34, vcc |
| ; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v0 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v1 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v7 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v9 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v14 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v12 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v10 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v6 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v4 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v32 |
| ; GCN-NEXT: .LBB30_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8i64_to_v32f16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB30_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0 |
| ; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; VI-NEXT: .LBB30_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8i64_to_v32f16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB30_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc |
| ; GFX9-NEXT: .LBB30_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8i64_to_v32f16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB30_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo |
| ; GFX11-NEXT: .LBB30_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <8 x i64> %a, splat (i64 3) |
| %a2 = bitcast <8 x i64> %a1 to <32 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x i64> %a to <32 x half> |
| br label %end |
| |
| end: |
| %phi = phi <32 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x half> %phi |
| } |
| |
| define <8 x i64> @bitcast_v32f16_to_v8i64(<32 x half> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32f16_to_v8i64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v45, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v44, v0 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v43, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v42, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v41, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v52, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v40, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v50, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v55, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v48, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v54, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v38, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v53, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v36, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v51, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v34, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v49, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v33, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v39, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v32, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v37, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v31, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v35, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v23, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v28 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v20, v46 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v30 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB31_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v43 |
| ; GCN-NEXT: v_or_b32_e32 v0, v44, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v42, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v40 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v20 |
| ; GCN-NEXT: v_or_b32_e32 v2, v52, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v50, v3 |
| ; GCN-NEXT: v_or_b32_e32 v4, v48, v4 |
| ; GCN-NEXT: v_or_b32_e32 v5, v38, v5 |
| ; GCN-NEXT: v_or_b32_e32 v6, v36, v6 |
| ; GCN-NEXT: v_or_b32_e32 v7, v34, v7 |
| ; GCN-NEXT: v_or_b32_e32 v8, v33, v8 |
| ; GCN-NEXT: v_or_b32_e32 v9, v32, v9 |
| ; GCN-NEXT: v_or_b32_e32 v10, v31, v10 |
| ; GCN-NEXT: v_or_b32_e32 v11, v21, v11 |
| ; GCN-NEXT: v_or_b32_e32 v12, v19, v12 |
| ; GCN-NEXT: v_or_b32_e32 v13, v18, v13 |
| ; GCN-NEXT: v_or_b32_e32 v14, v17, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: .LBB31_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB31_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v45 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v44 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v43 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v42 |
| ; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GCN-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v3, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v41 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v40 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v55 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v53 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v31 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GCN-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v5, v4 |
| ; GCN-NEXT: v_or_b32_e32 v4, v7, v6 |
| ; GCN-NEXT: v_or_b32_e32 v5, v9, v8 |
| ; GCN-NEXT: v_or_b32_e32 v6, v11, v10 |
| ; GCN-NEXT: v_or_b32_e32 v7, v13, v12 |
| ; GCN-NEXT: v_or_b32_e32 v8, v15, v14 |
| ; GCN-NEXT: v_or_b32_e32 v9, v26, v24 |
| ; GCN-NEXT: v_or_b32_e32 v10, v28, v27 |
| ; GCN-NEXT: v_or_b32_e32 v11, v21, v29 |
| ; GCN-NEXT: v_or_b32_e32 v12, v19, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v18, v23 |
| ; GCN-NEXT: v_or_b32_e32 v14, v17, v22 |
| ; GCN-NEXT: v_or_b32_e32 v15, v16, v20 |
| ; GCN-NEXT: .LBB31_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32f16_to_v8i64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB31_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v16, 0x200 |
| ; VI-NEXT: v_add_f16_sdwa v17, v15, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v15, 0x200, v15 |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v14, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v14, 0x200, v14 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v13, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v13, 0x200, v13 |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v12, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v12, 0x200, v12 |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v11, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v11, 0x200, v11 |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v10, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v10, 0x200, v10 |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v9, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v9, 0x200, v9 |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v8, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v8, 0x200, v8 |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v7, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v7, 0x200, v7 |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v6, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v6, 0x200, v6 |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v5, 0x200, v5 |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v4, 0x200, v4 |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, 0x200, v3 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v2, 0x200, v2 |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v1, 0x200, v1 |
| ; VI-NEXT: v_add_f16_sdwa v16, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v0, 0x200, v0 |
| ; VI-NEXT: v_or_b32_e32 v1, v1, v17 |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v16 |
| ; VI-NEXT: .LBB31_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32f16_to_v8i64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB31_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: s_movk_i32 s6, 0x200 |
| ; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB31_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32f16_to_v8i64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB31_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-NEXT: .LBB31_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <32 x half> %a1 to <8 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x half> %a to <8 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x i64> %phi |
| } |
| |
| define <32 x bfloat> @bitcast_v8i64_to_v32bf16(<8 x i64> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8i64_to_v32bf16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v55, v15 |
| ; GCN-NEXT: v_mov_b32_e32 v54, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v53, v13 |
| ; GCN-NEXT: v_mov_b32_e32 v52, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v51, v11 |
| ; GCN-NEXT: v_mov_b32_e32 v50, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v49, v9 |
| ; GCN-NEXT: v_mov_b32_e32 v48, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v39, v7 |
| ; GCN-NEXT: v_mov_b32_e32 v38, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v5 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v3 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v1 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v0 |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB32_3 |
| ; GCN-NEXT: ; %bb.1: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB32_4 |
| ; GCN-NEXT: .LBB32_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; GCN-NEXT: .LBB32_3: ; %cmp.false |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v55 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v54 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v53 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v52 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v51 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v50 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v49 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v48 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v48 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v39 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v38 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v36 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v36 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v35 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v34 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v34 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v33 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v32 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB32_2 |
| ; GCN-NEXT: .LBB32_4: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v32 |
| ; GCN-NEXT: v_addc_u32_e32 v1, vcc, 0, v33, vcc |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v34 |
| ; GCN-NEXT: v_addc_u32_e32 v3, vcc, 0, v35, vcc |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v36 |
| ; GCN-NEXT: v_addc_u32_e32 v5, vcc, 0, v37, vcc |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v38 |
| ; GCN-NEXT: v_addc_u32_e32 v7, vcc, 0, v39, vcc |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v48 |
| ; GCN-NEXT: v_addc_u32_e32 v9, vcc, 0, v49, vcc |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v50 |
| ; GCN-NEXT: v_addc_u32_e32 v11, vcc, 0, v51, vcc |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v52 |
| ; GCN-NEXT: v_addc_u32_e32 v13, vcc, 0, v53, vcc |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v54 |
| ; GCN-NEXT: v_addc_u32_e32 v15, vcc, 0, v55, vcc |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v15 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v14 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v13 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v12 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v11 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v10 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v9 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v7 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v6 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v5 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v4 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v3 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8i64_to_v32bf16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB32_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14 |
| ; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12 |
| ; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10 |
| ; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8 |
| ; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6 |
| ; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4 |
| ; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2 |
| ; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0 |
| ; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; VI-NEXT: .LBB32_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8i64_to_v32bf16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB32_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc |
| ; GFX9-NEXT: .LBB32_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8i64_to_v32bf16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB32_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo |
| ; GFX11-NEXT: .LBB32_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <8 x i64> %a, splat (i64 3) |
| %a2 = bitcast <8 x i64> %a1 to <32 x bfloat> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x i64> %a to <32 x bfloat> |
| br label %end |
| |
| end: |
| %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x bfloat> %phi |
| } |
| |
| define <8 x i64> @bitcast_v32bf16_to_v8i64(<32 x bfloat> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32bf16_to_v8i64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v1 |
| ; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v0 |
| ; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v3 |
| ; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v2 |
| ; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v5 |
| ; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v4 |
| ; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v7 |
| ; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v6 |
| ; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v9 |
| ; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v8 |
| ; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v11 |
| ; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v10 |
| ; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v13 |
| ; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v12 |
| ; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v15 |
| ; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v14 |
| ; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v17 |
| ; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v16 |
| ; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v19 |
| ; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v18 |
| ; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v21 |
| ; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v20 |
| ; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v23 |
| ; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v22 |
| ; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25 |
| ; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v24 |
| ; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v27 |
| ; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v26 |
| ; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v29 |
| ; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v28 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47 |
| ; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v46 |
| ; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v30 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB33_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v44 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v42 |
| ; GCN-NEXT: v_alignbit_b32 v0, v0, v45, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v1, v43, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v41 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v40 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v55 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v54 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v52 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v50 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v48 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v38 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v35 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v25 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v23 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v20 |
| ; GCN-NEXT: v_alignbit_b32 v2, v2, v51, 16 |
| ; GCN-NEXT: v_alignbit_b32 v3, v3, v49, 16 |
| ; GCN-NEXT: v_alignbit_b32 v4, v4, v39, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v5, v37, 16 |
| ; GCN-NEXT: v_alignbit_b32 v6, v6, v36, 16 |
| ; GCN-NEXT: v_alignbit_b32 v7, v7, v34, 16 |
| ; GCN-NEXT: v_alignbit_b32 v8, v8, v33, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v9, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v10, v10, v31, 16 |
| ; GCN-NEXT: v_alignbit_b32 v11, v11, v21, 16 |
| ; GCN-NEXT: v_alignbit_b32 v12, v12, v19, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v13, v18, 16 |
| ; GCN-NEXT: v_alignbit_b32 v14, v14, v17, 16 |
| ; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: .LBB33_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB33_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v45 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v44 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v43 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v42 |
| ; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v51 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v41 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v49 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v40 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v39 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v55 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v37 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v54 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v36 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v53 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v34 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v52 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v33 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v50 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v32 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v48 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v31 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v38 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v35 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20 |
| ; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16 |
| ; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16 |
| ; GCN-NEXT: v_alignbit_b32 v4, v7, v6, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v9, v8, 16 |
| ; GCN-NEXT: v_alignbit_b32 v6, v11, v10, 16 |
| ; GCN-NEXT: v_alignbit_b32 v7, v13, v12, 16 |
| ; GCN-NEXT: v_alignbit_b32 v8, v15, v14, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v26, v24, 16 |
| ; GCN-NEXT: v_alignbit_b32 v10, v28, v27, 16 |
| ; GCN-NEXT: v_alignbit_b32 v11, v29, v21, 16 |
| ; GCN-NEXT: v_alignbit_b32 v12, v25, v19, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v23, v18, 16 |
| ; GCN-NEXT: v_alignbit_b32 v14, v22, v17, 16 |
| ; GCN-NEXT: v_alignbit_b32 v15, v20, v16, 16 |
| ; GCN-NEXT: .LBB33_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32bf16_to_v8i64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB33_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v15 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17 |
| ; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v15, 16, 1 |
| ; VI-NEXT: s_movk_i32 s6, 0x7fff |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v15 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v15 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15 |
| ; VI-NEXT: v_alignbit_b32 v15, v15, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v14 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v14, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v14 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v14 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14 |
| ; VI-NEXT: v_alignbit_b32 v14, v14, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v13 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v13, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v13 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v13 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13 |
| ; VI-NEXT: v_alignbit_b32 v13, v13, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v12 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v12, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v12 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v12 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12 |
| ; VI-NEXT: v_alignbit_b32 v12, v12, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v11 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v11, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v11 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v11 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11 |
| ; VI-NEXT: v_alignbit_b32 v11, v11, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v10 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v10, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v10 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v10 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10 |
| ; VI-NEXT: v_alignbit_b32 v10, v10, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v9 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v9, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v9 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v9 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9 |
| ; VI-NEXT: v_alignbit_b32 v9, v9, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v8, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v8 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v8 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8 |
| ; VI-NEXT: v_alignbit_b32 v8, v8, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v7 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v7, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v7 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v7 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7 |
| ; VI-NEXT: v_alignbit_b32 v7, v7, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v6 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v6, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v6 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v6 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6 |
| ; VI-NEXT: v_alignbit_b32 v6, v6, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v5 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v5, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v5 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v5 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5 |
| ; VI-NEXT: v_alignbit_b32 v5, v5, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v4 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v4, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v4 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v4 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4 |
| ; VI-NEXT: v_alignbit_b32 v4, v4, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v3 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v3, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v3 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v3 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; VI-NEXT: v_alignbit_b32 v3, v3, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v2 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v2, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v2 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v2 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 |
| ; VI-NEXT: v_alignbit_b32 v2, v2, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v1 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v1, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v1 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1 |
| ; VI-NEXT: v_alignbit_b32 v1, v1, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v0, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v0 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v0 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0 |
| ; VI-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 |
| ; VI-NEXT: v_alignbit_b32 v0, v0, v16, 16 |
| ; VI-NEXT: .LBB33_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32bf16_to_v8i64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB33_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: s_movk_i32 s6, 0x7fff |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v15, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v15, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc |
| ; GFX9-NEXT: s_mov_b32 s7, 0x7060302 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v14, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v14, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v14 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v13, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v13, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v13 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v12, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v12, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v12 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v11, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v11, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v11 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v11, v11, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v10, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v10, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v10 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v10, v10, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v9, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v9, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v9 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v9, v9, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v8, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v8, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v8 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v8, v8, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v7, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v7, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v7 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v7, v7, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v6, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v6, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v6 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v6, v6, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v5, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v5, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v5 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v5, v5, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v4, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v4, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v4 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v4, v4, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v3, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v3, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v3 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v3, v3, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v2, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v2, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v2 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v2, v2, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v1, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v1, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v1 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v1, v1, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v0, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v0, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v0 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v0, v0, v16, s7 |
| ; GFX9-NEXT: .LBB33_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32bf16_to_v8i64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB33_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16 |
| ; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v18, v16, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v16 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_bfe_u32 v23, v14, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v17, 0x7fff |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v16, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v18, v19, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v19, v23, v14, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v17 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX11-NEXT: v_bfe_u32 v20, v15, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v15 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v20, v20, v15, 0x7fff |
| ; GFX11-NEXT: v_dual_cndmask_b32 v15, v20, v22 :: v_dual_lshlrev_b32 v20, 16, v13 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_perm_b32 v15, v15, v16, 0x7060302 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v17, v21, v18 :: v_dual_add_f32 v18, 0x40c00000, v20 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v14 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v13 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v14, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v12 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_perm_b32 v14, v14, v17, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v17, v13, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_cndmask_b32 v16, v16, v20 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v13, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v11 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_perm_b32 v13, v13, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v21 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX11-NEXT: v_bfe_u32 v18, v12, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v12 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v12, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v12, v18, v22, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v12, v12, v17, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v17, v11, 16, 1 |
| ; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v11, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v19, v10, 16, 1 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v19, v19, v10, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v11, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v10 |
| ; GFX11-NEXT: v_perm_b32 v11, v11, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v10, v19, v22 :: v_dual_lshlrev_b32 v21, 16, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_lshlrev_b32 v19, 16, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX11-NEXT: v_perm_b32 v10, v10, v17, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_add_f32 v8, 0x40c00000, v8 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v18, v8, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v9, 0x40c00000, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v8, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v17, v9, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v9 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v9, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v9, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v8 |
| ; GFX11-NEXT: v_perm_b32 v9, v9, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v8, v18, v22, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v7 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v8, v8, v17, 0x7060302 |
| ; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v19, 0x40c00000, v21 |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GFX11-NEXT: v_bfe_u32 v19, v6, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v7, 0x40c00000, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_add3_u32 v19, v19, v6, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v17, v7, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v7 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v7, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v7, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v5 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v6 |
| ; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_cndmask_b32 v17, v17, v20 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v20, v18, 16, 1 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v6, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v4 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX11-NEXT: v_perm_b32 v7, v7, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v6, v6, v17, 0x7060302 |
| ; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v19 |
| ; GFX11-NEXT: v_add3_u32 v19, v20, v18, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GFX11-NEXT: v_bfe_u32 v22, v17, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v16, v5, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v5 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v5, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v5, v16, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v16, v22, v17, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: v_perm_b32 v5, v5, v18, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v18, v4, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v19, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v4 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v4, 0x7fff |
| ; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v20 :: v_dual_lshlrev_b32 v20, 16, v2 |
| ; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v4, v18, v19, vcc_lo |
| ; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: v_perm_b32 v4, v4, v16, 0x7060302 |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v17, 0x7fff |
| ; GFX11-NEXT: v_bfe_u32 v21, v3, 16, 1 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v18 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v3, 0x7fff |
| ; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v3 |
| ; GFX11-NEXT: v_add3_u32 v21, v22, v18, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v1 |
| ; GFX11-NEXT: v_bfe_u32 v24, v2, 16, 1 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v22 |
| ; GFX11-NEXT: v_add3_u32 v20, v24, v2, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_perm_b32 v3, v3, v17, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v18, v21, v23, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v2 |
| ; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v2, v20, v21 :: v_dual_lshlrev_b32 v23, 16, v0 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) |
| ; GFX11-NEXT: v_add3_u32 v21, v22, v19, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v19 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v23 |
| ; GFX11-NEXT: v_perm_b32 v2, v2, v18, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v19, v21, v22 :: v_dual_add_f32 v0, 0x40c00000, v0 |
| ; GFX11-NEXT: v_bfe_u32 v24, v20, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v20 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v21, v0, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v0 |
| ; GFX11-NEXT: v_add3_u32 v24, v24, v20, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v0, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v23, v1, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 |
| ; GFX11-NEXT: v_add3_u32 v22, v23, v1, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v1, v22, v23, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 |
| ; GFX11-NEXT: v_perm_b32 v1, v1, v19, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v20, v24, v25, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v0, v21, v26, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) |
| ; GFX11-NEXT: v_perm_b32 v0, v0, v20, 0x7060302 |
| ; GFX11-NEXT: .LBB33_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x bfloat> %a, splat (bfloat 0xR40C0) |
| %a2 = bitcast <32 x bfloat> %a1 to <8 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x bfloat> %a to <8 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x i64> %phi |
| } |
| |
| define <64 x i8> @bitcast_v8i64_to_v64i8(<8 x i64> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8i64_to_v64i8: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB34_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v21, v16, v15, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v20, v14, v13, 16 |
| ; GCN-NEXT: v_alignbit_b32 v28, v14, v13, 8 |
| ; GCN-NEXT: v_alignbit_b32 v23, v12, v11, 24 |
| ; GCN-NEXT: v_alignbit_b32 v25, v12, v11, 16 |
| ; GCN-NEXT: v_alignbit_b32 v34, v12, v11, 8 |
| ; GCN-NEXT: v_alignbit_b32 v30, v10, v9, 24 |
| ; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v48, v10, v9, 8 |
| ; GCN-NEXT: v_alignbit_b32 v36, v8, v7, 24 |
| ; GCN-NEXT: v_alignbit_b32 v37, v8, v7, 16 |
| ; GCN-NEXT: v_alignbit_b32 v54, v8, v7, 8 |
| ; GCN-NEXT: v_alignbit_b32 v49, v6, v5, 24 |
| ; GCN-NEXT: v_alignbit_b32 v50, v6, v5, 16 |
| ; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 8 |
| ; GCN-NEXT: v_alignbit_b32 v40, v4, v3, 24 |
| ; GCN-NEXT: v_alignbit_b32 v41, v4, v3, 16 |
| ; GCN-NEXT: v_alignbit_b32 v58, v4, v3, 8 |
| ; GCN-NEXT: v_alignbit_b32 v44, v2, v1, 24 |
| ; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 16 |
| ; GCN-NEXT: v_alignbit_b32 v60, v2, v1, 8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v32, 24, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v51, 8, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v52, 24, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v42, 24, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v56, 24, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v61, 24, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2 |
| ; GCN-NEXT: .LBB34_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB34_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1 |
| ; GCN-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3 |
| ; GCN-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5 |
| ; GCN-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7 |
| ; GCN-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9 |
| ; GCN-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11 |
| ; GCN-NEXT: v_addc_u32_e32 v12, vcc, 0, v12, vcc |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13 |
| ; GCN-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15 |
| ; GCN-NEXT: v_addc_u32_e32 v16, vcc, 0, v16, vcc |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v21, v16, v15, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v20, v14, v13, 16 |
| ; GCN-NEXT: v_alignbit_b32 v28, v14, v13, 8 |
| ; GCN-NEXT: v_alignbit_b32 v23, v12, v11, 24 |
| ; GCN-NEXT: v_alignbit_b32 v25, v12, v11, 16 |
| ; GCN-NEXT: v_alignbit_b32 v34, v12, v11, 8 |
| ; GCN-NEXT: v_alignbit_b32 v30, v10, v9, 24 |
| ; GCN-NEXT: v_alignbit_b32 v31, v10, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v48, v10, v9, 8 |
| ; GCN-NEXT: v_alignbit_b32 v36, v8, v7, 24 |
| ; GCN-NEXT: v_alignbit_b32 v37, v8, v7, 16 |
| ; GCN-NEXT: v_alignbit_b32 v54, v8, v7, 8 |
| ; GCN-NEXT: v_alignbit_b32 v49, v6, v5, 24 |
| ; GCN-NEXT: v_alignbit_b32 v50, v6, v5, 16 |
| ; GCN-NEXT: v_alignbit_b32 v46, v6, v5, 8 |
| ; GCN-NEXT: v_alignbit_b32 v40, v4, v3, 24 |
| ; GCN-NEXT: v_alignbit_b32 v41, v4, v3, 16 |
| ; GCN-NEXT: v_alignbit_b32 v58, v4, v3, 8 |
| ; GCN-NEXT: v_alignbit_b32 v44, v2, v1, 24 |
| ; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 16 |
| ; GCN-NEXT: v_alignbit_b32 v60, v2, v1, 8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v22, 24, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 8, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v32, 24, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v51, 8, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 24, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v55, 8, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v52, 24, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v47, 8, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v42, 24, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v56, 24, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v57, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v61, 24, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2 |
| ; GCN-NEXT: .LBB34_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 8, v60 |
| ; GCN-NEXT: s_waitcnt expcnt(3) |
| ; GCN-NEXT: v_or_b32_e32 v60, v1, v18 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v17 |
| ; GCN-NEXT: v_or_b32_e32 v17, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v58 |
| ; GCN-NEXT: v_or_b32_e32 v58, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v63 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v63, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v46 |
| ; GCN-NEXT: v_or_b32_e32 v46, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 24, v44 |
| ; GCN-NEXT: v_and_b32_e32 v45, 0xff, v62 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v61, 24, v61 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v59 |
| ; GCN-NEXT: v_or_b32_e32 v44, v1, v2 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0 |
| ; GCN-NEXT: v_and_b32_e32 v41, 0xff, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v40, 24, v40 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v54 |
| ; GCN-NEXT: v_or_b32_e32 v7, v2, v3 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0 |
| ; GCN-NEXT: v_and_b32_e32 v57, 0xff, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v56, 24, v56 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v47 |
| ; GCN-NEXT: v_or_b32_e32 v54, v3, v4 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 12, v0 |
| ; GCN-NEXT: v_and_b32_e32 v47, 0xff, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v49 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v48 |
| ; GCN-NEXT: v_or_b32_e32 v9, v4, v5 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 16, v0 |
| ; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 24, v42 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v55 |
| ; GCN-NEXT: v_or_b32_e32 v48, v5, v6 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 20, v0 |
| ; GCN-NEXT: v_and_b32_e32 v55, 0xff, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v36 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 8, v34 |
| ; GCN-NEXT: v_or_b32_e32 v34, v6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0 |
| ; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v52 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xff, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 8, v51 |
| ; GCN-NEXT: v_or_b32_e32 v11, v8, v10 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 28, v0 |
| ; GCN-NEXT: v_and_b32_e32 v51, 0xff, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 24, v30 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 8, v28 |
| ; GCN-NEXT: v_or_b32_e32 v13, v10, v12 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 32, v0 |
| ; GCN-NEXT: v_and_b32_e32 v39, 0xff, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v31, 24, v38 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v35 |
| ; GCN-NEXT: v_or_b32_e32 v28, v12, v14 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 36, v0 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 24, v23 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 8, v21 |
| ; GCN-NEXT: v_or_b32_e32 v21, v14, v15 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 40, v0 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xff, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 24, v32 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v29 |
| ; GCN-NEXT: v_or_b32_e32 v29, v15, v16 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 44, v0 |
| ; GCN-NEXT: v_and_b32_e32 v52, 0xff, v20 |
| ; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18 |
| ; GCN-NEXT: v_or_b32_e32 v32, v19, v16 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 48, v0 |
| ; GCN-NEXT: v_and_b32_e32 v42, 0xff, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 24, v26 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v45 |
| ; GCN-NEXT: v_or_b32_e32 v27, v61, v18 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 52, v0 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v45, 0xff, v19 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v41 |
| ; GCN-NEXT: v_or_b32_e32 v40, v40, v19 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 56, v0 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v41, 24, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v57 |
| ; GCN-NEXT: v_or_b32_e32 v56, v56, v22 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 60, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v42 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GCN-NEXT: v_and_b32_e32 v57, 0xffff, v60 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v58 |
| ; GCN-NEXT: v_and_b32_e32 v59, 0xffff, v63 |
| ; GCN-NEXT: v_and_b32_e32 v46, 0xffff, v46 |
| ; GCN-NEXT: v_or_b32_e32 v49, v49, v47 |
| ; GCN-NEXT: v_and_b32_e32 v44, 0xffff, v44 |
| ; GCN-NEXT: v_or_b32_e32 v50, v50, v43 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GCN-NEXT: v_or_b32_e32 v36, v36, v55 |
| ; GCN-NEXT: v_and_b32_e32 v54, 0xffff, v54 |
| ; GCN-NEXT: v_or_b32_e32 v37, v37, v53 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_or_b32_e32 v30, v30, v51 |
| ; GCN-NEXT: v_and_b32_e32 v48, 0xffff, v48 |
| ; GCN-NEXT: v_or_b32_e32 v31, v31, v39 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xffff, v34 |
| ; GCN-NEXT: v_or_b32_e32 v23, v23, v35 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_or_b32_e32 v25, v25, v38 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_or_b32_e32 v20, v20, v52 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28 |
| ; GCN-NEXT: v_or_b32_e32 v26, v26, v42 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21 |
| ; GCN-NEXT: v_or_b32_e32 v33, v33, v45 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_or_b32_e32 v24, v41, v24 |
| ; GCN-NEXT: v_or_b32_e32 v32, v57, v32 |
| ; GCN-NEXT: v_or_b32_e32 v17, v17, v27 |
| ; GCN-NEXT: v_or_b32_e32 v27, v58, v40 |
| ; GCN-NEXT: v_or_b32_e32 v35, v59, v56 |
| ; GCN-NEXT: v_or_b32_e32 v38, v46, v49 |
| ; GCN-NEXT: v_or_b32_e32 v39, v44, v50 |
| ; GCN-NEXT: v_or_b32_e32 v7, v7, v36 |
| ; GCN-NEXT: v_or_b32_e32 v36, v54, v37 |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v30 |
| ; GCN-NEXT: v_or_b32_e32 v30, v48, v31 |
| ; GCN-NEXT: v_or_b32_e32 v23, v34, v23 |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v20 |
| ; GCN-NEXT: v_or_b32_e32 v20, v28, v26 |
| ; GCN-NEXT: v_or_b32_e32 v21, v21, v33 |
| ; GCN-NEXT: v_or_b32_e32 v24, v29, v24 |
| ; GCN-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v17, v1, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v27, v2, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v35, v3, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v38, v4, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v39, v5, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v7, v6, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v36, v8, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v9, v10, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v30, v12, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v23, v14, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v11, v15, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v13, v16, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v20, v18, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v21, v19, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v24, v22, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8i64_to_v64i8: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr49 |
| ; VI-NEXT: ; implicit-def: $vgpr48 |
| ; VI-NEXT: ; implicit-def: $vgpr39 |
| ; VI-NEXT: ; implicit-def: $vgpr38 |
| ; VI-NEXT: ; implicit-def: $vgpr37 |
| ; VI-NEXT: ; implicit-def: $vgpr36 |
| ; VI-NEXT: ; implicit-def: $vgpr35 |
| ; VI-NEXT: ; implicit-def: $vgpr34 |
| ; VI-NEXT: ; implicit-def: $vgpr33 |
| ; VI-NEXT: ; implicit-def: $vgpr32 |
| ; VI-NEXT: ; implicit-def: $vgpr31 |
| ; VI-NEXT: ; implicit-def: $vgpr30 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB34_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v16 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8] |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v16 |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; VI-NEXT: v_mov_b32_e32 v26, v22 |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4] |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2] |
| ; VI-NEXT: .LBB34_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB34_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1 |
| ; VI-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc |
| ; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3 |
| ; VI-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc |
| ; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5 |
| ; VI-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc |
| ; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7 |
| ; VI-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc |
| ; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9 |
| ; VI-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc |
| ; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11 |
| ; VI-NEXT: v_addc_u32_e32 v12, vcc, 0, v12, vcc |
| ; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13 |
| ; VI-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc |
| ; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15 |
| ; VI-NEXT: v_addc_u32_e32 v16, vcc, 0, v16, vcc |
| ; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4] |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8] |
| ; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2] |
| ; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 8, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: .LBB34_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v29 |
| ; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v22 |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v22, v27, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v28 |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v62 |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61 |
| ; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v59 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v57 |
| ; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v56 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v46 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44 |
| ; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v43 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v24 |
| ; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v55 |
| ; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v54 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v52 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v50 |
| ; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v49 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; VI-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37 |
| ; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; VI-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32 |
| ; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; VI-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v26 |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 60, v0 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8i64_to_v64i8: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB34_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v16 |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_mov_b32_e32 v26, v23 |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: .LBB34_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB34_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: v_add_co_u32_e32 v1, vcc, 3, v1 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v2, vcc, 0, v2, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v3, vcc, 3, v3 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v4, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v5, vcc, 3, v5 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v6, vcc, 0, v6, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v7, vcc, 3, v7 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v8, vcc, 0, v8, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v9, vcc, 3, v9 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v10, vcc, 0, v10, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v11, vcc, 3, v11 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v12, vcc, 0, v12, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v13, vcc, 3, v13 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v14, vcc, 0, v14, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v15, vcc, 3, v15 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v16, vcc, 0, v16, vcc |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: .LBB34_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v29 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v23 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v23, v27, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28 |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v22 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v60 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v59 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v57 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v56 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v46 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v21 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v55 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v50 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v49 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56 |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v26 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60 |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8i64_to_v64i8: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB34_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB34_2: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB34_4 |
| ; GFX11-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-NEXT: v_add_co_u32 v1, vcc_lo, v1, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v2, null, 0, v2, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, v3, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v4, null, 0, v4, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v5, vcc_lo, v5, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v6, null, 0, v6, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v9, vcc_lo, v9, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v10, null, 0, v10, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v11, vcc_lo, v11, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v12, null, 0, v12, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v13, vcc_lo, v13, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v14, null, 0, v14, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v15, vcc_lo, v15, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v16, null, 0, v16, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v7, vcc_lo, v7, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v8, null, 0, v8, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB34_4: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v96, 0xff, v96 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v86, 0xff, v86 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v87 |
| ; GFX11-NEXT: v_or_b32_e32 v24, v96, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v25 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v86, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84 |
| ; GFX11-NEXT: v_and_b32_e32 v83, 0xff, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v82 |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v80, 8, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v71, 8, v71 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v24 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v81 |
| ; GFX11-NEXT: v_and_b32_e32 v70, 0xff, v70 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v83, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v25, v80 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v71 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v70, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v22 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v66 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v65 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v55 |
| ; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v23, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v25, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v52 |
| ; GFX11-NEXT: v_lshlrev_b16 v20, 8, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v50 |
| ; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v48, 8, v48 |
| ; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v39 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v23, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v25, v49 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v48 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v39, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v38 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v37 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v35 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v34 |
| ; GFX11-NEXT: v_lshlrev_b16 v18, 8, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_lshlrev_b16 v69, 8, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v68, 0xff, v68 |
| ; GFX11-NEXT: v_lshlrev_b16 v67, 8, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v23, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v32 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v30 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v29 |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v26, 8, v26 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v69 |
| ; GFX11-NEXT: v_or_b32_e32 v67, v68, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v23, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v25, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v67 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; GFX11-NEXT: s_clause 0x3 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off |
| ; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <8 x i64> %a, splat (i64 3) |
| %a2 = bitcast <8 x i64> %a1 to <64 x i8> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x i64> %a to <64 x i8> |
| br label %end |
| |
| end: |
| %phi = phi <64 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <64 x i8> %phi |
| } |
| |
| define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v64i8_to_v8i64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_mov_b32_e32 v38, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v31, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104 |
| ; GCN-NEXT: s_waitcnt expcnt(3) |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:100 |
| ; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:96 |
| ; GCN-NEXT: s_waitcnt expcnt(2) |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:92 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88 |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:84 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76 |
| ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:72 |
| ; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:68 |
| ; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64 |
| ; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:60 |
| ; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56 |
| ; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:52 |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8 |
| ; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:132 |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:128 |
| ; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:124 |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:120 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v52 |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 |
| ; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:36 |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 |
| ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:28 |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:24 |
| ; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:20 |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v5 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v7 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v9 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v11 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v13 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v15 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v17 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v19 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v21 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v23 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v25 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v27 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v45 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 8, v44 |
| ; GCN-NEXT: s_waitcnt vmcnt(10) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v59, 24, v59 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v58 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v57, 24, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v58, 24, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v56, 8, v14 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v12 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v10 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v6 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v4 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v2 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:44 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:12 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116 |
| ; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0 |
| ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v46, 24, v46 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB35_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v42 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v41 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v40 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v55 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v35 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v36 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v37 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v38 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v50 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v36, 0xff, v54 |
| ; GCN-NEXT: v_and_b32_e32 v37, 0xff, v53 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xff, v52 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v49 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v48 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v39 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v63 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v62 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v61 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v60 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v21 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v32 |
| ; GCN-NEXT: v_or_b32_e32 v17, v33, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v34 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v20, v20, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v23, v24, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26 |
| ; GCN-NEXT: v_or_b32_e32 v26, v28, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v30 |
| ; GCN-NEXT: v_or_b32_e32 v27, v35, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_or_b32_e32 v29, v36, v44 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v37 |
| ; GCN-NEXT: v_or_b32_e32 v31, v38, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GCN-NEXT: v_or_b32_e32 v18, v18, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v7, v32, v7 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v6, v32, v6 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v5, v32, v5 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v32, v4 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v32, v21 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v22, v32, v22 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v24, v32, v24 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v28, v32, v28 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27 |
| ; GCN-NEXT: v_or_b32_e32 v25, v59, v25 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_or_b32_e32 v30, v57, v30 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31 |
| ; GCN-NEXT: v_or_b32_e32 v32, v58, v8 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v9 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v34, v8, v10 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xffff, v11 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v8, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v14, v8, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v16, v8, v16 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GCN-NEXT: v_or_b32_e32 v19, v46, v19 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v7 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v6 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v5 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; GCN-NEXT: v_or_b32_e32 v4, v17, v21 |
| ; GCN-NEXT: v_or_b32_e32 v5, v20, v22 |
| ; GCN-NEXT: v_or_b32_e32 v6, v23, v24 |
| ; GCN-NEXT: v_or_b32_e32 v7, v26, v28 |
| ; GCN-NEXT: v_or_b32_e32 v8, v27, v25 |
| ; GCN-NEXT: v_or_b32_e32 v9, v29, v30 |
| ; GCN-NEXT: v_or_b32_e32 v10, v31, v32 |
| ; GCN-NEXT: v_or_b32_e32 v11, v33, v34 |
| ; GCN-NEXT: v_or_b32_e32 v12, v35, v12 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; GCN-NEXT: v_or_b32_e32 v14, v15, v16 |
| ; GCN-NEXT: v_or_b32_e32 v15, v18, v19 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; kill: killed $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: .LBB35_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB35_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GCN-NEXT: v_or_b32_e32 v0, v42, v0 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GCN-NEXT: v_or_b32_e32 v1, v41, v1 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GCN-NEXT: v_or_b32_e32 v2, v40, v2 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GCN-NEXT: v_or_b32_e32 v3, v55, v3 |
| ; GCN-NEXT: s_movk_i32 s7, 0x300 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35 |
| ; GCN-NEXT: s_mov_b32 s6, 0x3000000 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v50 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v25 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v54 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v53 |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v52 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v49 |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v48 |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v39 |
| ; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v63 |
| ; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v62 |
| ; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v61 |
| ; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v60 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19 |
| ; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v51 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v7 |
| ; GCN-NEXT: v_or_b32_e32 v6, v17, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v9 |
| ; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v11 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v13 |
| ; GCN-NEXT: v_or_b32_e32 v13, v29, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15 |
| ; GCN-NEXT: v_or_b32_e32 v15, v27, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18 |
| ; GCN-NEXT: v_or_b32_e32 v17, v44, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v22 |
| ; GCN-NEXT: v_or_b32_e32 v20, v45, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v23 |
| ; GCN-NEXT: v_or_b32_e32 v23, v56, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v25, v25, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v30 |
| ; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v27, v27, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v32 |
| ; GCN-NEXT: v_or_b32_e32 v29, v43, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GCN-NEXT: v_or_b32_e32 v19, v47, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v34 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v31, v31, v36 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v8, v32, v8 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v5, v32, v5 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v32, v4 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v7, v32, v7 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v10, v32, v10 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v32, v12 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v14, v32, v14 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15 |
| ; GCN-NEXT: v_or_b32_e32 v16, v59, v16 |
| ; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v17 |
| ; GCN-NEXT: v_or_b32_e32 v18, v57, v18 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20 |
| ; GCN-NEXT: v_or_b32_e32 v22, v58, v22 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, s7, v23 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v24, v32, v24 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v26, v32, v26 |
| ; GCN-NEXT: v_add_i32_e32 v27, vcc, s7, v27 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v28, v32, v28 |
| ; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v32, v21 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 0x300, v19 |
| ; GCN-NEXT: v_or_b32_e32 v30, v46, v30 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GCN-NEXT: v_or_b32_e32 v0, v31, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v8, v1 |
| ; GCN-NEXT: v_or_b32_e32 v2, v5, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; GCN-NEXT: v_or_b32_e32 v4, v7, v6 |
| ; GCN-NEXT: v_or_b32_e32 v5, v10, v9 |
| ; GCN-NEXT: v_or_b32_e32 v6, v12, v11 |
| ; GCN-NEXT: v_or_b32_e32 v7, v14, v13 |
| ; GCN-NEXT: v_or_b32_e32 v8, v16, v15 |
| ; GCN-NEXT: v_or_b32_e32 v9, v18, v17 |
| ; GCN-NEXT: v_or_b32_e32 v10, v22, v20 |
| ; GCN-NEXT: v_or_b32_e32 v11, v24, v23 |
| ; GCN-NEXT: v_or_b32_e32 v12, v26, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v28, v27 |
| ; GCN-NEXT: v_or_b32_e32 v14, v21, v29 |
| ; GCN-NEXT: v_or_b32_e32 v15, v30, v19 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x3000000, v14 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x3000000, v15 |
| ; GCN-NEXT: .LBB35_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v64i8_to_v8i64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132 |
| ; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72 |
| ; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80 |
| ; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88 |
| ; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96 |
| ; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104 |
| ; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112 |
| ; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120 |
| ; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128 |
| ; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20 |
| ; VI-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12 |
| ; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100 |
| ; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68 |
| ; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36 |
| ; VI-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4 |
| ; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v26, 8, v23 |
| ; VI-NEXT: v_lshlrev_b16_e32 v28, 8, v25 |
| ; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v27 |
| ; VI-NEXT: v_lshlrev_b16_e32 v62, 8, v29 |
| ; VI-NEXT: v_lshlrev_b16_e32 v18, 8, v17 |
| ; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v19 |
| ; VI-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116 |
| ; VI-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124 |
| ; VI-NEXT: v_lshlrev_b16_e32 v31, 8, v1 |
| ; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v3 |
| ; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v5 |
| ; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v7 |
| ; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v9 |
| ; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v11 |
| ; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v13 |
| ; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v15 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54 |
| ; VI-NEXT: v_lshlrev_b16_e32 v61, 8, v0 |
| ; VI-NEXT: v_lshlrev_b16_e32 v59, 8, v2 |
| ; VI-NEXT: v_lshlrev_b16_e32 v58, 8, v4 |
| ; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v6 |
| ; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v8 |
| ; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v10 |
| ; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v12 |
| ; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v14 |
| ; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v24 |
| ; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v30 |
| ; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v53 |
| ; VI-NEXT: s_waitcnt vmcnt(13) |
| ; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v42 |
| ; VI-NEXT: s_waitcnt vmcnt(12) |
| ; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v44 |
| ; VI-NEXT: s_waitcnt vmcnt(11) |
| ; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v47 |
| ; VI-NEXT: s_waitcnt vmcnt(10) |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v57 |
| ; VI-NEXT: s_waitcnt vmcnt(9) |
| ; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v60 |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_lshlrev_b16_e32 v57, 8, v63 |
| ; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108 |
| ; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92 |
| ; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84 |
| ; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76 |
| ; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60 |
| ; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44 |
| ; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB35_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr30 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr23 |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(7) |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr31 |
| ; VI-NEXT: ; implicit-def: $vgpr32 |
| ; VI-NEXT: ; implicit-def: $vgpr33 |
| ; VI-NEXT: ; implicit-def: $vgpr34 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr38 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr35 |
| ; VI-NEXT: ; implicit-def: $vgpr36 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr37 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr49 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr48 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr39 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: .LBB35_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB35_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: v_mov_b32_e32 v15, 0x300 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v51 |
| ; VI-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v9, v9, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v24 |
| ; VI-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v10, v10, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v30 |
| ; VI-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v11, v11, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v60 |
| ; VI-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v12, v12, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v47 |
| ; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v13, v13, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v42 |
| ; VI-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v14, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v17, 3, v17 |
| ; VI-NEXT: v_or_b32_sdwa v17, v57, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: s_waitcnt vmcnt(7) |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v1 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v0, 0x300, v0 |
| ; VI-NEXT: v_add_u16_sdwa v1, v1, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; VI-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; VI-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v4, v4, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; VI-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v5, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; VI-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v6, v6, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; VI-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v7, v7, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v53 |
| ; VI-NEXT: v_or_b32_sdwa v16, v23, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v16, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v8 |
| ; VI-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v8, v8, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: v_add_u16_sdwa v15, v17, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v3 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; VI-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v4, 0x300, v4 |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; VI-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v5, 0x300, v5 |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; VI-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v6, 0x300, v6 |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; VI-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v7, 0x300, v7 |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v52 |
| ; VI-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v8, 0x300, v8 |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v50 |
| ; VI-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v9, 0x300, v9 |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v49 |
| ; VI-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v10, 0x300, v10 |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v63 |
| ; VI-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v11, 0x300, v11 |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v48 |
| ; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v12, 0x300, v12 |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v44 |
| ; VI-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v13, 0x300, v13 |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v39 |
| ; VI-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v14, 0x300, v14 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v16 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v19 |
| ; VI-NEXT: v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v16, 0x300, v16 |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: .LBB35_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v64i8_to_v8i64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132 |
| ; GFX9-NEXT: s_nop 0 |
| ; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64 |
| ; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72 |
| ; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80 |
| ; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88 |
| ; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96 |
| ; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104 |
| ; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112 |
| ; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120 |
| ; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128 |
| ; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20 |
| ; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12 |
| ; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100 |
| ; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68 |
| ; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36 |
| ; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v21 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v26, 8, v23 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v28, 8, v25 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v27 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v62, 8, v29 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v18, 8, v17 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v19 |
| ; GFX9-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116 |
| ; GFX9-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v31, 8, v1 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v3 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v5 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v7 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v9 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v11 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v13 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v15 |
| ; GFX9-NEXT: s_waitcnt vmcnt(25) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54 |
| ; GFX9-NEXT: s_waitcnt vmcnt(24) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v61, 8, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(23) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v59, 8, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(22) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v58, 8, v4 |
| ; GFX9-NEXT: s_waitcnt vmcnt(21) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(20) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v8 |
| ; GFX9-NEXT: s_waitcnt vmcnt(19) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v10 |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v12 |
| ; GFX9-NEXT: s_waitcnt vmcnt(17) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v14 |
| ; GFX9-NEXT: s_waitcnt vmcnt(16) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v24 |
| ; GFX9-NEXT: s_waitcnt vmcnt(15) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v30 |
| ; GFX9-NEXT: s_waitcnt vmcnt(14) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v53 |
| ; GFX9-NEXT: s_waitcnt vmcnt(13) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v42 |
| ; GFX9-NEXT: s_waitcnt vmcnt(12) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v44 |
| ; GFX9-NEXT: s_waitcnt vmcnt(11) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v47 |
| ; GFX9-NEXT: s_waitcnt vmcnt(10) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v57 |
| ; GFX9-NEXT: s_waitcnt vmcnt(9) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v60 |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v57, 8, v63 |
| ; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108 |
| ; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92 |
| ; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84 |
| ; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76 |
| ; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60 |
| ; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52 |
| ; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44 |
| ; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB35_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: .LBB35_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB35_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_movk_i32 s6, 0x300 |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_add_u16_e32 v9, 3, v51 |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_add_u16_e32 v10, 3, v24 |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v11, 3, v30 |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v12, 3, v60 |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v13, 3, v47 |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v14, 3, v42 |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v15, 3, v53 |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_add_u16_e32 v1, 3, v1 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GFX9-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v16, 3, v17 |
| ; GFX9-NEXT: v_or_b32_sdwa v16, v57, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_add_u16_e32 v8, 3, v8 |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3 |
| ; GFX9-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4 |
| ; GFX9-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5 |
| ; GFX9-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6 |
| ; GFX9-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7 |
| ; GFX9-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; GFX9-NEXT: v_add_u16_e32 v8, 3, v52 |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8 |
| ; GFX9-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; GFX9-NEXT: v_add_u16_e32 v9, 3, v50 |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9 |
| ; GFX9-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; GFX9-NEXT: v_add_u16_e32 v10, 3, v49 |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10 |
| ; GFX9-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; GFX9-NEXT: v_add_u16_e32 v11, 3, v63 |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11 |
| ; GFX9-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GFX9-NEXT: v_add_u16_e32 v12, 3, v48 |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12 |
| ; GFX9-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; GFX9-NEXT: v_add_u16_e32 v13, 3, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13 |
| ; GFX9-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; GFX9-NEXT: v_add_u16_e32 v14, 3, v39 |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14 |
| ; GFX9-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; GFX9-NEXT: v_add_u16_e32 v15, 3, v19 |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15 |
| ; GFX9-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; GFX9-NEXT: .LBB35_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v64i8_to_v8i64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12 |
| ; GFX11-NEXT: v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8 |
| ; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4 |
| ; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0 |
| ; GFX11-NEXT: s_clause 0x1f |
| ; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:128 |
| ; GFX11-NEXT: scratch_load_u16 v39, off, s32 offset:124 |
| ; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:120 |
| ; GFX11-NEXT: scratch_load_u16 v48, off, s32 offset:116 |
| ; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:112 |
| ; GFX11-NEXT: scratch_load_u16 v49, off, s32 offset:108 |
| ; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:104 |
| ; GFX11-NEXT: scratch_load_u16 v50, off, s32 offset:100 |
| ; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:96 |
| ; GFX11-NEXT: scratch_load_u16 v51, off, s32 offset:92 |
| ; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:88 |
| ; GFX11-NEXT: scratch_load_b32 v12, off, s32 offset:132 |
| ; GFX11-NEXT: scratch_load_u16 v14, off, s32 |
| ; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:8 |
| ; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:16 |
| ; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:24 |
| ; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:32 |
| ; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:40 |
| ; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:48 |
| ; GFX11-NEXT: scratch_load_u16 v128, off, s32 offset:56 |
| ; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:64 |
| ; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:72 |
| ; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:80 |
| ; GFX11-NEXT: scratch_load_u16 v52, off, s32 offset:84 |
| ; GFX11-NEXT: scratch_load_u16 v53, off, s32 offset:76 |
| ; GFX11-NEXT: scratch_load_u16 v54, off, s32 offset:68 |
| ; GFX11-NEXT: scratch_load_u16 v55, off, s32 offset:60 |
| ; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:52 |
| ; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:44 |
| ; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:36 |
| ; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:28 |
| ; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:20 |
| ; GFX11-NEXT: s_clause 0x1 |
| ; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:12 |
| ; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:4 |
| ; GFX11-NEXT: v_lshlrev_b16 v118, 8, v1 |
| ; GFX11-NEXT: v_lshlrev_b16 v119, 8, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v117, 8, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v114, 8, v7 |
| ; GFX11-NEXT: v_lshlrev_b16 v115, 8, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v116, 8, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v113, 8, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v100, 8, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v101, 8, v17 |
| ; GFX11-NEXT: v_lshlrev_b16 v102, 8, v19 |
| ; GFX11-NEXT: v_lshlrev_b16 v103, 8, v21 |
| ; GFX11-NEXT: v_lshlrev_b16 v112, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v87, 8, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v96, 8, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v97, 8, v29 |
| ; GFX11-NEXT: s_waitcnt vmcnt(33) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(31) |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v2 |
| ; GFX11-NEXT: s_waitcnt vmcnt(29) |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v4 |
| ; GFX11-NEXT: s_waitcnt vmcnt(27) |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v6 |
| ; GFX11-NEXT: s_waitcnt vmcnt(25) |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v8 |
| ; GFX11-NEXT: s_waitcnt vmcnt(22) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12 |
| ; GFX11-NEXT: s_waitcnt vmcnt(21) |
| ; GFX11-NEXT: v_lshlrev_b16 v98, 8, v14 |
| ; GFX11-NEXT: s_waitcnt vmcnt(20) |
| ; GFX11-NEXT: v_lshlrev_b16 v99, 8, v65 |
| ; GFX11-NEXT: s_waitcnt vmcnt(19) |
| ; GFX11-NEXT: v_lshlrev_b16 v81, 8, v66 |
| ; GFX11-NEXT: s_waitcnt vmcnt(18) |
| ; GFX11-NEXT: v_lshlrev_b16 v82, 8, v67 |
| ; GFX11-NEXT: s_waitcnt vmcnt(17) |
| ; GFX11-NEXT: v_lshlrev_b16 v83, 8, v83 |
| ; GFX11-NEXT: s_waitcnt vmcnt(16) |
| ; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84 |
| ; GFX11-NEXT: s_waitcnt vmcnt(15) |
| ; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85 |
| ; GFX11-NEXT: s_waitcnt vmcnt(14) |
| ; GFX11-NEXT: v_lshlrev_b16 v27, 8, v128 |
| ; GFX11-NEXT: s_waitcnt vmcnt(13) |
| ; GFX11-NEXT: v_lshlrev_b16 v29, 8, v129 |
| ; GFX11-NEXT: s_waitcnt vmcnt(12) |
| ; GFX11-NEXT: v_lshlrev_b16 v65, 8, v130 |
| ; GFX11-NEXT: s_waitcnt vmcnt(11) |
| ; GFX11-NEXT: v_lshlrev_b16 v66, 8, v131 |
| ; GFX11-NEXT: v_lshlrev_b16 v67, 8, v10 |
| ; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB35_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB35_4 |
| ; GFX11-NEXT: .LBB35_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB35_3: ; %cmp.false |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v37 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v118 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v119 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v35 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v117 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v6, v113 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v114 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v115 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v116 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v101 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v38 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v102 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v103 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v112 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v100 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v1 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v4, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v6, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v7, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v9, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v30 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v86 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v71 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v70 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v68 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v87 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v96 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v97 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v98 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v99 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v81 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v82 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v8, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v10, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v12, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v14, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v55 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v53 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v52 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v50 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v49 |
| ; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v48 |
| ; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v39 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v27 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v29 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v65 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v66 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v18, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v22, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v24, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v13, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v15, v16 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v17, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v19, v20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr16 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr118 |
| ; GFX11-NEXT: ; implicit-def: $vgpr119 |
| ; GFX11-NEXT: ; implicit-def: $vgpr117 |
| ; GFX11-NEXT: ; implicit-def: $vgpr114 |
| ; GFX11-NEXT: ; implicit-def: $vgpr115 |
| ; GFX11-NEXT: ; implicit-def: $vgpr116 |
| ; GFX11-NEXT: ; implicit-def: $vgpr113 |
| ; GFX11-NEXT: ; implicit-def: $vgpr100 |
| ; GFX11-NEXT: ; implicit-def: $vgpr101 |
| ; GFX11-NEXT: ; implicit-def: $vgpr102 |
| ; GFX11-NEXT: ; implicit-def: $vgpr103 |
| ; GFX11-NEXT: ; implicit-def: $vgpr112 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr97 |
| ; GFX11-NEXT: ; implicit-def: $vgpr98 |
| ; GFX11-NEXT: ; implicit-def: $vgpr99 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB35_2 |
| ; GFX11-NEXT: .LBB35_4: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u16 v0, v31, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v32, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v33, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v34, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v35, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v5, v36, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v6, v37, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v118, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v119, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v117, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v114, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v115, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v116, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v113, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v7, v16, 3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v5 |
| ; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v6, v38, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v8, v18, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v9, v20, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v10, v22, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v101, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v100, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v102, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v103, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v112, v10 |
| ; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7 |
| ; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8 |
| ; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9 |
| ; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v3, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v5, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v7, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v9, v10 |
| ; GFX11-NEXT: v_add_nc_u16 v6, v24, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v7, v26, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v8, v28, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v9, v30, 3 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_add_nc_u16 v10, v86, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v11, v80, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v12, v71, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v13, v70, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v14, v69, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v15, v68, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v87, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v96, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v97, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v98, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v99, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v81, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v82, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v83, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v84, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v85, v15 |
| ; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7 |
| ; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8 |
| ; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9 |
| ; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10 |
| ; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11 |
| ; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12 |
| ; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13 |
| ; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14 |
| ; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v8, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v10, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v12, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v14, v15 |
| ; GFX11-NEXT: v_add_nc_u16 v11, v64, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v12, v55, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v13, v54, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v14, v53, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v15, v52, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v16, v51, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v18, v50, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v20, v49, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v22, v48, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v24, v39, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v27, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v29, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v65, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v66, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v67, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v19, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v21, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v23, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v25, v24 |
| ; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11 |
| ; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12 |
| ; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13 |
| ; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14 |
| ; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15 |
| ; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16 |
| ; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17 |
| ; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18 |
| ; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19 |
| ; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v13, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v15, v16 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v17, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v19, v20 |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <64 x i8> %a, splat (i8 3) |
| %a2 = bitcast <64 x i8> %a1 to <8 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <64 x i8> %a to <8 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <8 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x i64> %phi |
| } |
| |
| define <32 x i16> @bitcast_v8f64_to_v32i16(<8 x double> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8f64_to_v32i16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v55, v15 |
| ; GCN-NEXT: v_mov_b32_e32 v54, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v53, v13 |
| ; GCN-NEXT: v_mov_b32_e32 v52, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v51, v11 |
| ; GCN-NEXT: v_mov_b32_e32 v50, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v49, v9 |
| ; GCN-NEXT: v_mov_b32_e32 v48, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v38, v7 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v5 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v3 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v2 |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB36_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_alignbit_b32 v29, v55, v54, 16 |
| ; GCN-NEXT: v_alignbit_b32 v25, v53, v52, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v51, v50, 16 |
| ; GCN-NEXT: v_alignbit_b32 v17, v49, v48, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v38, v37, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v36, v35, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v34, v33, 16 |
| ; GCN-NEXT: v_alignbit_b32 v32, v1, v0, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v55 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v51 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v49 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v38 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v36 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v34 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v1 |
| ; GCN-NEXT: .LBB36_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB36_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GCN-NEXT: v_add_f64 v[33:34], v[33:34], 1.0 |
| ; GCN-NEXT: v_add_f64 v[35:36], v[35:36], 1.0 |
| ; GCN-NEXT: v_add_f64 v[37:38], v[37:38], 1.0 |
| ; GCN-NEXT: v_add_f64 v[48:49], v[48:49], 1.0 |
| ; GCN-NEXT: v_add_f64 v[50:51], v[50:51], 1.0 |
| ; GCN-NEXT: v_add_f64 v[52:53], v[52:53], 1.0 |
| ; GCN-NEXT: v_add_f64 v[54:55], v[54:55], 1.0 |
| ; GCN-NEXT: v_alignbit_b32 v29, v55, v54, 16 |
| ; GCN-NEXT: v_alignbit_b32 v25, v53, v52, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v51, v50, 16 |
| ; GCN-NEXT: v_alignbit_b32 v17, v49, v48, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v38, v37, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v36, v35, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v34, v33, 16 |
| ; GCN-NEXT: v_alignbit_b32 v32, v1, v0, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v55 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v51 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v49 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v38 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v36 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v34 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v1 |
| ; GCN-NEXT: .LBB36_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_mov_b32_e32 v2, v1 |
| ; GCN-NEXT: v_mov_b32_e32 v4, v33 |
| ; GCN-NEXT: v_mov_b32_e32 v6, v34 |
| ; GCN-NEXT: v_mov_b32_e32 v8, v35 |
| ; GCN-NEXT: v_mov_b32_e32 v10, v36 |
| ; GCN-NEXT: v_mov_b32_e32 v12, v37 |
| ; GCN-NEXT: v_mov_b32_e32 v14, v38 |
| ; GCN-NEXT: v_mov_b32_e32 v16, v48 |
| ; GCN-NEXT: v_mov_b32_e32 v18, v49 |
| ; GCN-NEXT: v_mov_b32_e32 v20, v50 |
| ; GCN-NEXT: v_mov_b32_e32 v22, v51 |
| ; GCN-NEXT: v_mov_b32_e32 v24, v52 |
| ; GCN-NEXT: v_mov_b32_e32 v26, v53 |
| ; GCN-NEXT: v_mov_b32_e32 v28, v54 |
| ; GCN-NEXT: v_mov_b32_e32 v30, v55 |
| ; GCN-NEXT: v_mov_b32_e32 v1, v32 |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8f64_to_v32i16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB36_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; VI-NEXT: .LBB36_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8f64_to_v32i16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB36_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX9-NEXT: .LBB36_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8f64_to_v32i16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB36_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX11-NEXT: .LBB36_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <8 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <8 x double> %a1 to <32 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x double> %a to <32 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x i16> %phi |
| } |
| |
| define <8 x double> @bitcast_v32i16_to_v8f64(<32 x i16> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32i16_to_v8f64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v38, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v31, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v0 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB37_3 |
| ; GCN-NEXT: ; %bb.1: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB37_4 |
| ; GCN-NEXT: .LBB37_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; GCN-NEXT: .LBB37_3: ; %cmp.false |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v31 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v32 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v54 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v55 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v33 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v34 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v35 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v36 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v37 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v38 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v30 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v39 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v48 |
| ; GCN-NEXT: v_or_b32_e32 v4, v4, v49 |
| ; GCN-NEXT: v_or_b32_e32 v5, v5, v50 |
| ; GCN-NEXT: v_or_b32_e32 v6, v6, v51 |
| ; GCN-NEXT: v_or_b32_e32 v7, v7, v52 |
| ; GCN-NEXT: v_or_b32_e32 v8, v8, v17 |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v19 |
| ; GCN-NEXT: v_or_b32_e32 v10, v10, v21 |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v23 |
| ; GCN-NEXT: v_or_b32_e32 v12, v12, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v27 |
| ; GCN-NEXT: v_or_b32_e32 v14, v14, v29 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v53 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB37_2 |
| ; GCN-NEXT: .LBB37_4: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31 |
| ; GCN-NEXT: s_mov_b32 s6, 0x30000 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_or_b32_e32 v0, v54, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v55, v1 |
| ; GCN-NEXT: v_or_b32_e32 v2, v39, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v48, v3 |
| ; GCN-NEXT: v_or_b32_e32 v4, v49, v4 |
| ; GCN-NEXT: v_or_b32_e32 v5, v50, v5 |
| ; GCN-NEXT: v_or_b32_e32 v6, v51, v6 |
| ; GCN-NEXT: v_or_b32_e32 v7, v52, v7 |
| ; GCN-NEXT: v_or_b32_e32 v8, v17, v8 |
| ; GCN-NEXT: v_or_b32_e32 v9, v19, v9 |
| ; GCN-NEXT: v_or_b32_e32 v10, v21, v10 |
| ; GCN-NEXT: v_or_b32_e32 v11, v23, v11 |
| ; GCN-NEXT: v_or_b32_e32 v12, v25, v12 |
| ; GCN-NEXT: v_or_b32_e32 v13, v27, v13 |
| ; GCN-NEXT: v_or_b32_e32 v14, v29, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v53, v15 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v13 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v14 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x30000, v15 |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32i16_to_v8f64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB37_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v17, 3 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v15 |
| ; VI-NEXT: v_add_u16_sdwa v15, v15, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v14 |
| ; VI-NEXT: v_add_u16_sdwa v14, v14, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v14, v16, v14 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v13 |
| ; VI-NEXT: v_add_u16_sdwa v13, v13, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v13, v16, v13 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v12 |
| ; VI-NEXT: v_add_u16_sdwa v12, v12, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v12, v16, v12 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v11 |
| ; VI-NEXT: v_add_u16_sdwa v11, v11, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v11, v16, v11 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v10 |
| ; VI-NEXT: v_add_u16_sdwa v10, v10, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v10, v16, v10 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v9 |
| ; VI-NEXT: v_add_u16_sdwa v9, v9, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v9, v16, v9 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v8 |
| ; VI-NEXT: v_add_u16_sdwa v8, v8, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v8, v16, v8 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v7 |
| ; VI-NEXT: v_add_u16_sdwa v7, v7, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v7, v16, v7 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v6 |
| ; VI-NEXT: v_add_u16_sdwa v6, v6, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v6, v16, v6 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v5 |
| ; VI-NEXT: v_add_u16_sdwa v5, v5, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v5, v16, v5 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v4 |
| ; VI-NEXT: v_add_u16_sdwa v4, v4, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v4, v16, v4 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v3 |
| ; VI-NEXT: v_add_u16_sdwa v3, v3, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v3, v16, v3 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v2 |
| ; VI-NEXT: v_add_u16_sdwa v2, v2, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v2, v16, v2 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v1 |
| ; VI-NEXT: v_add_u16_sdwa v1, v1, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v1, v16, v1 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v0 |
| ; VI-NEXT: v_add_u16_sdwa v0, v0, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v0, v16, v0 |
| ; VI-NEXT: .LBB37_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32i16_to_v8f64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB37_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB37_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32i16_to_v8f64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB37_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: .LBB37_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <32 x i16> %a, splat (i16 3) |
| %a2 = bitcast <32 x i16> %a1 to <8 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x i16> %a to <8 x double> |
| br label %end |
| |
| end: |
| %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x double> %phi |
| } |
| |
| define <32 x half> @bitcast_v8f64_to_v32f16(<8 x double> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8f64_to_v32f16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB38_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v15 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v13 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v9 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v7 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v5 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v3 |
| ; GCN-NEXT: s_waitcnt expcnt(2) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v2 |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v41, 16, v1 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v42, 16, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v14 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v12 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v10 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v38, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v37, v6 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v36, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v35, v4 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v34, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v33, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v32, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v48, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v49, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v50, v53 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v51, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v52, v55 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v53, v40 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v54, v41 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v55, v42 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v39, v0 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: .LBB38_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB38_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v0 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v1 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v5 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v7 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v9 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v14 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v12 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v10 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v8 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v38, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v37, v6 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v36, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v35, v4 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v34, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v33, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v32, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v39, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v49, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v50, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v51, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v52, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v53, v53 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v54, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v55, v55 |
| ; GCN-NEXT: .LBB38_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_mov_b32_e32 v0, v39 |
| ; GCN-NEXT: v_mov_b32_e32 v1, v55 |
| ; GCN-NEXT: v_mov_b32_e32 v2, v32 |
| ; GCN-NEXT: v_mov_b32_e32 v3, v54 |
| ; GCN-NEXT: v_mov_b32_e32 v4, v33 |
| ; GCN-NEXT: v_mov_b32_e32 v5, v53 |
| ; GCN-NEXT: v_mov_b32_e32 v6, v34 |
| ; GCN-NEXT: v_mov_b32_e32 v7, v52 |
| ; GCN-NEXT: v_mov_b32_e32 v8, v35 |
| ; GCN-NEXT: v_mov_b32_e32 v9, v51 |
| ; GCN-NEXT: v_mov_b32_e32 v10, v36 |
| ; GCN-NEXT: v_mov_b32_e32 v11, v50 |
| ; GCN-NEXT: v_mov_b32_e32 v12, v37 |
| ; GCN-NEXT: v_mov_b32_e32 v13, v49 |
| ; GCN-NEXT: v_mov_b32_e32 v14, v38 |
| ; GCN-NEXT: v_mov_b32_e32 v15, v48 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8f64_to_v32f16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB38_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; VI-NEXT: .LBB38_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8f64_to_v32f16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB38_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX9-NEXT: .LBB38_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8f64_to_v32f16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB38_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX11-NEXT: .LBB38_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <8 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <8 x double> %a1 to <32 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x double> %a to <32 x half> |
| br label %end |
| |
| end: |
| %phi = phi <32 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x half> %phi |
| } |
| |
| define <8 x double> @bitcast_v32f16_to_v8f64(<32 x half> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32f16_to_v8f64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v45, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v44, v0 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v43, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v42, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v41, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v52, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v40, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v50, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v55, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v48, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v54, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v38, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v53, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v36, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v51, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v34, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v49, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v33, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v39, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v32, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v37, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v31, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v35, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v23, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v28 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v20, v46 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v30 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB39_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v43 |
| ; GCN-NEXT: v_or_b32_e32 v0, v44, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v42, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v40 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v20 |
| ; GCN-NEXT: v_or_b32_e32 v2, v52, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v50, v3 |
| ; GCN-NEXT: v_or_b32_e32 v4, v48, v4 |
| ; GCN-NEXT: v_or_b32_e32 v5, v38, v5 |
| ; GCN-NEXT: v_or_b32_e32 v6, v36, v6 |
| ; GCN-NEXT: v_or_b32_e32 v7, v34, v7 |
| ; GCN-NEXT: v_or_b32_e32 v8, v33, v8 |
| ; GCN-NEXT: v_or_b32_e32 v9, v32, v9 |
| ; GCN-NEXT: v_or_b32_e32 v10, v31, v10 |
| ; GCN-NEXT: v_or_b32_e32 v11, v21, v11 |
| ; GCN-NEXT: v_or_b32_e32 v12, v19, v12 |
| ; GCN-NEXT: v_or_b32_e32 v13, v18, v13 |
| ; GCN-NEXT: v_or_b32_e32 v14, v17, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: .LBB39_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB39_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v45 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v44 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v43 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v42 |
| ; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GCN-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v3, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v41 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v40 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v55 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v53 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v31 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GCN-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v5, v4 |
| ; GCN-NEXT: v_or_b32_e32 v4, v7, v6 |
| ; GCN-NEXT: v_or_b32_e32 v5, v9, v8 |
| ; GCN-NEXT: v_or_b32_e32 v6, v11, v10 |
| ; GCN-NEXT: v_or_b32_e32 v7, v13, v12 |
| ; GCN-NEXT: v_or_b32_e32 v8, v15, v14 |
| ; GCN-NEXT: v_or_b32_e32 v9, v26, v24 |
| ; GCN-NEXT: v_or_b32_e32 v10, v28, v27 |
| ; GCN-NEXT: v_or_b32_e32 v11, v21, v29 |
| ; GCN-NEXT: v_or_b32_e32 v12, v19, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v18, v23 |
| ; GCN-NEXT: v_or_b32_e32 v14, v17, v22 |
| ; GCN-NEXT: v_or_b32_e32 v15, v16, v20 |
| ; GCN-NEXT: .LBB39_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32f16_to_v8f64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB39_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v16, 0x200 |
| ; VI-NEXT: v_add_f16_sdwa v17, v15, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v15, 0x200, v15 |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v14, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v14, 0x200, v14 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v13, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v13, 0x200, v13 |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v12, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v12, 0x200, v12 |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v11, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v11, 0x200, v11 |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v10, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v10, 0x200, v10 |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v9, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v9, 0x200, v9 |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v8, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v8, 0x200, v8 |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v7, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v7, 0x200, v7 |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v6, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v6, 0x200, v6 |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v5, 0x200, v5 |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v4, 0x200, v4 |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, 0x200, v3 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v2, 0x200, v2 |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v17 |
| ; VI-NEXT: v_add_f16_sdwa v17, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v1, 0x200, v1 |
| ; VI-NEXT: v_add_f16_sdwa v16, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v0, 0x200, v0 |
| ; VI-NEXT: v_or_b32_e32 v1, v1, v17 |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v16 |
| ; VI-NEXT: .LBB39_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32f16_to_v8f64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB39_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: s_movk_i32 s6, 0x200 |
| ; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB39_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32f16_to_v8f64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB39_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-NEXT: .LBB39_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <32 x half> %a1 to <8 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x half> %a to <8 x double> |
| br label %end |
| |
| end: |
| %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x double> %phi |
| } |
| |
| define <32 x bfloat> @bitcast_v8f64_to_v32bf16(<8 x double> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8f64_to_v32bf16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB40_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v15 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v14 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v13 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v12 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v11 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v10 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v9 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v7 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v6 |
| ; GCN-NEXT: v_and_b32_e32 v36, 0xffff0000, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v5 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xffff0000, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v4 |
| ; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v3 |
| ; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v2 |
| ; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v1 |
| ; GCN-NEXT: v_and_b32_e32 v54, 0xffff0000, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v0 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: .LBB40_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB40_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GCN-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GCN-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GCN-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GCN-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GCN-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GCN-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GCN-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v15 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v14 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v13 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v12 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v11 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v10 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v9 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v7 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xffff0000, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v6 |
| ; GCN-NEXT: v_and_b32_e32 v36, 0xffff0000, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v5 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xffff0000, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v4 |
| ; GCN-NEXT: v_and_b32_e32 v48, 0xffff0000, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v3 |
| ; GCN-NEXT: v_and_b32_e32 v50, 0xffff0000, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v2 |
| ; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v1 |
| ; GCN-NEXT: v_and_b32_e32 v54, 0xffff0000, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v0 |
| ; GCN-NEXT: .LBB40_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_mov_b32_e32 v0, v55 |
| ; GCN-NEXT: v_mov_b32_e32 v1, v54 |
| ; GCN-NEXT: v_mov_b32_e32 v2, v53 |
| ; GCN-NEXT: v_mov_b32_e32 v3, v52 |
| ; GCN-NEXT: v_mov_b32_e32 v4, v51 |
| ; GCN-NEXT: v_mov_b32_e32 v5, v50 |
| ; GCN-NEXT: v_mov_b32_e32 v6, v49 |
| ; GCN-NEXT: v_mov_b32_e32 v7, v48 |
| ; GCN-NEXT: v_mov_b32_e32 v8, v39 |
| ; GCN-NEXT: v_mov_b32_e32 v9, v38 |
| ; GCN-NEXT: v_mov_b32_e32 v10, v37 |
| ; GCN-NEXT: v_mov_b32_e32 v11, v36 |
| ; GCN-NEXT: v_mov_b32_e32 v12, v35 |
| ; GCN-NEXT: v_mov_b32_e32 v13, v34 |
| ; GCN-NEXT: v_mov_b32_e32 v14, v33 |
| ; GCN-NEXT: v_mov_b32_e32 v15, v32 |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8f64_to_v32bf16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB40_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; VI-NEXT: .LBB40_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8f64_to_v32bf16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB40_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX9-NEXT: .LBB40_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8f64_to_v32bf16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB40_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX11-NEXT: .LBB40_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <8 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <8 x double> %a1 to <32 x bfloat> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x double> %a to <32 x bfloat> |
| br label %end |
| |
| end: |
| %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x bfloat> %phi |
| } |
| |
| define <8 x double> @bitcast_v32bf16_to_v8f64(<32 x bfloat> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32bf16_to_v8f64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v1 |
| ; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v0 |
| ; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v3 |
| ; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v2 |
| ; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v5 |
| ; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v4 |
| ; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v7 |
| ; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v6 |
| ; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v9 |
| ; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v8 |
| ; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v11 |
| ; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v10 |
| ; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v13 |
| ; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v12 |
| ; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v15 |
| ; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v14 |
| ; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v17 |
| ; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v16 |
| ; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v19 |
| ; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v18 |
| ; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v21 |
| ; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v20 |
| ; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v23 |
| ; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v22 |
| ; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25 |
| ; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v24 |
| ; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v27 |
| ; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v26 |
| ; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v29 |
| ; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v28 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v47 |
| ; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v46 |
| ; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v30 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB41_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v44 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v42 |
| ; GCN-NEXT: v_alignbit_b32 v0, v0, v45, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v1, v43, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v41 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v40 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v55 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v54 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v52 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v50 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v48 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v38 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v35 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v25 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v23 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v20 |
| ; GCN-NEXT: v_alignbit_b32 v2, v2, v51, 16 |
| ; GCN-NEXT: v_alignbit_b32 v3, v3, v49, 16 |
| ; GCN-NEXT: v_alignbit_b32 v4, v4, v39, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v5, v37, 16 |
| ; GCN-NEXT: v_alignbit_b32 v6, v6, v36, 16 |
| ; GCN-NEXT: v_alignbit_b32 v7, v7, v34, 16 |
| ; GCN-NEXT: v_alignbit_b32 v8, v8, v33, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v9, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v10, v10, v31, 16 |
| ; GCN-NEXT: v_alignbit_b32 v11, v11, v21, 16 |
| ; GCN-NEXT: v_alignbit_b32 v12, v12, v19, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v13, v18, 16 |
| ; GCN-NEXT: v_alignbit_b32 v14, v14, v17, 16 |
| ; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: .LBB41_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB41_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v45 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v44 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v43 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v42 |
| ; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v51 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v41 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v49 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v40 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v39 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v55 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v37 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v54 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v36 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v53 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v34 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v52 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v33 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v50 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v32 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v48 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v31 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v38 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v35 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20 |
| ; GCN-NEXT: v_alignbit_b32 v2, v3, v2, 16 |
| ; GCN-NEXT: v_alignbit_b32 v3, v5, v4, 16 |
| ; GCN-NEXT: v_alignbit_b32 v4, v7, v6, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v9, v8, 16 |
| ; GCN-NEXT: v_alignbit_b32 v6, v11, v10, 16 |
| ; GCN-NEXT: v_alignbit_b32 v7, v13, v12, 16 |
| ; GCN-NEXT: v_alignbit_b32 v8, v15, v14, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v26, v24, 16 |
| ; GCN-NEXT: v_alignbit_b32 v10, v28, v27, 16 |
| ; GCN-NEXT: v_alignbit_b32 v11, v29, v21, 16 |
| ; GCN-NEXT: v_alignbit_b32 v12, v25, v19, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v23, v18, 16 |
| ; GCN-NEXT: v_alignbit_b32 v14, v22, v17, 16 |
| ; GCN-NEXT: v_alignbit_b32 v15, v20, v16, 16 |
| ; GCN-NEXT: .LBB41_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32bf16_to_v8f64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB41_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v15 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17 |
| ; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v15, 16, 1 |
| ; VI-NEXT: s_movk_i32 s6, 0x7fff |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v15 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v15 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15 |
| ; VI-NEXT: v_alignbit_b32 v15, v15, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v14 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v14, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v14 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v14 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14 |
| ; VI-NEXT: v_alignbit_b32 v14, v14, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v13 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v13, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v13 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v13 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13 |
| ; VI-NEXT: v_alignbit_b32 v13, v13, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v12 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v12, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v12 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v12 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12 |
| ; VI-NEXT: v_alignbit_b32 v12, v12, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v11 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v11, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v11 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v11 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11 |
| ; VI-NEXT: v_alignbit_b32 v11, v11, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v10 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v10, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v10 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v10 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10 |
| ; VI-NEXT: v_alignbit_b32 v10, v10, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v9 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v9, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v9 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v9 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9 |
| ; VI-NEXT: v_alignbit_b32 v9, v9, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v8, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v8 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v8 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8 |
| ; VI-NEXT: v_alignbit_b32 v8, v8, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v7 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v7, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v7 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v7 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7 |
| ; VI-NEXT: v_alignbit_b32 v7, v7, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v6 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v6, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v6 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v6 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6 |
| ; VI-NEXT: v_alignbit_b32 v6, v6, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v5 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v5, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v5 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v5 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5 |
| ; VI-NEXT: v_alignbit_b32 v5, v5, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v4 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v4, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v4 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v4 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4 |
| ; VI-NEXT: v_alignbit_b32 v4, v4, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v3 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v3, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v3 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v3 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; VI-NEXT: v_alignbit_b32 v3, v3, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v2 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v2, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v2 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v2 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 |
| ; VI-NEXT: v_alignbit_b32 v2, v2, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v1 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v1, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v1 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1 |
| ; VI-NEXT: v_alignbit_b32 v1, v1, v16, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v0, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v0 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v0 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0 |
| ; VI-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 |
| ; VI-NEXT: v_alignbit_b32 v0, v0, v16, 16 |
| ; VI-NEXT: .LBB41_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32bf16_to_v8f64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB41_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: s_movk_i32 s6, 0x7fff |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v15, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v15, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc |
| ; GFX9-NEXT: s_mov_b32 s7, 0x7060302 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v14, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v14, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v14 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v13, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v13, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v13 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v12, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v12, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v12 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v11, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v11, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v11 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v11, v11, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v10, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v10, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v10 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v10, v10, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v9, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v9, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v9 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v9, v9, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v8, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v8, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v8 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v8, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v8, v8, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v7, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v7, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v7 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v7, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v7, v7, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v6, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v6, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v6 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v6, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v6, v6, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v5, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v5, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v5 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v5, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v5, v5, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v4, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v4, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v4 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v4, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v4, v4, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v3, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v3, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v3 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v3, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v3, v3, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v2, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v2, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v2 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v2, v2, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v1, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v1, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v1 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v1, v1, v16, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v0, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v0, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v0 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc |
| ; GFX9-NEXT: v_perm_b32 v0, v0, v16, s7 |
| ; GFX9-NEXT: .LBB41_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32bf16_to_v8f64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB41_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16 |
| ; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v18, v16, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v16 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_bfe_u32 v23, v14, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v17, 0x7fff |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v16, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v18, v19, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v19, v23, v14, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v17 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX11-NEXT: v_bfe_u32 v20, v15, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v15 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v20, v20, v15, 0x7fff |
| ; GFX11-NEXT: v_dual_cndmask_b32 v15, v20, v22 :: v_dual_lshlrev_b32 v20, 16, v13 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_perm_b32 v15, v15, v16, 0x7060302 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v17, v21, v18 :: v_dual_add_f32 v18, 0x40c00000, v20 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v14 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v13 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v14, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v12 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_perm_b32 v14, v14, v17, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v17, v13, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_cndmask_b32 v16, v16, v20 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v13, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v11 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_perm_b32 v13, v13, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v21 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX11-NEXT: v_bfe_u32 v18, v12, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v12 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v12, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v12, v18, v22, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v12, v12, v17, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v17, v11, 16, 1 |
| ; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v11, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v19, v10, 16, 1 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v19, v19, v10, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v11, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v10 |
| ; GFX11-NEXT: v_perm_b32 v11, v11, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v10, v19, v22 :: v_dual_lshlrev_b32 v21, 16, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_lshlrev_b32 v19, 16, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX11-NEXT: v_perm_b32 v10, v10, v17, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v16, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_add_f32 v8, 0x40c00000, v8 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v18, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v18, v8, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v9, 0x40c00000, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v8, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v17, v9, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v9 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v9, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v9, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v19, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v8 |
| ; GFX11-NEXT: v_perm_b32 v9, v9, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v8, v18, v22, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v7 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v8, v8, v17, 0x7060302 |
| ; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v19, 0x40c00000, v21 |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v16, v19, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v19 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v19, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GFX11-NEXT: v_bfe_u32 v19, v6, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_f32 v7, 0x40c00000, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_add3_u32 v19, v19, v6, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v17, v7, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v7 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 |
| ; GFX11-NEXT: v_add3_u32 v17, v17, v7, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v7, v17, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v17, v22, v18, 0x7fff |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v5 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v6 |
| ; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_cndmask_b32 v17, v17, v20 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v20, v18, 16, 1 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v6, v19, v22 :: v_dual_lshlrev_b32 v19, 16, v4 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX11-NEXT: v_perm_b32 v7, v7, v16, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v6, v6, v17, 0x7060302 |
| ; GFX11-NEXT: v_add_f32_e32 v17, 0x40c00000, v19 |
| ; GFX11-NEXT: v_add3_u32 v19, v20, v18, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GFX11-NEXT: v_bfe_u32 v22, v17, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v16, v5, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v5 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 |
| ; GFX11-NEXT: v_add3_u32 v16, v16, v5, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v5, v16, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v16, v22, v17, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: v_perm_b32 v5, v5, v18, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v18, v4, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v19, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v4 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 |
| ; GFX11-NEXT: v_add3_u32 v18, v18, v4, 0x7fff |
| ; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v20 :: v_dual_lshlrev_b32 v20, 16, v2 |
| ; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v4, v18, v19, vcc_lo |
| ; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v18, 0x40c00000, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: v_perm_b32 v4, v4, v16, 0x7060302 |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v17, 0x7fff |
| ; GFX11-NEXT: v_bfe_u32 v21, v3, 16, 1 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v18 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v3, 0x7fff |
| ; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v3 |
| ; GFX11-NEXT: v_add3_u32 v21, v22, v18, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v1 |
| ; GFX11-NEXT: v_bfe_u32 v24, v2, 16, 1 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v22 |
| ; GFX11-NEXT: v_add3_u32 v20, v24, v2, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_perm_b32 v3, v3, v17, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v18, v21, v23, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v2 |
| ; GFX11-NEXT: v_bfe_u32 v22, v19, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v2, v20, v21 :: v_dual_lshlrev_b32 v23, 16, v0 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) |
| ; GFX11-NEXT: v_add3_u32 v21, v22, v19, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v19 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; GFX11-NEXT: v_add_f32_e32 v20, 0x40c00000, v23 |
| ; GFX11-NEXT: v_perm_b32 v2, v2, v18, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v19, v21, v22 :: v_dual_add_f32 v0, 0x40c00000, v0 |
| ; GFX11-NEXT: v_bfe_u32 v24, v20, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v20 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v21, v0, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v0 |
| ; GFX11-NEXT: v_add3_u32 v24, v24, v20, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v0, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX11-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v23, v1, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 |
| ; GFX11-NEXT: v_add3_u32 v22, v23, v1, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v1, v22, v23, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 |
| ; GFX11-NEXT: v_perm_b32 v1, v1, v19, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v20, v24, v25, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v0, v21, v26, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) |
| ; GFX11-NEXT: v_perm_b32 v0, v0, v20, 0x7060302 |
| ; GFX11-NEXT: .LBB41_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x bfloat> %a, splat (bfloat 0xR40C0) |
| %a2 = bitcast <32 x bfloat> %a1 to <8 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x bfloat> %a to <8 x double> |
| br label %end |
| |
| end: |
| %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x double> %phi |
| } |
| |
| define <64 x i8> @bitcast_v8f64_to_v64i8(<8 x double> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v8f64_to_v64i8: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; kill: killed $vgpr18 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB42_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 8 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 16 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v24, v14, v13, 8 |
| ; GCN-NEXT: v_alignbit_b32 v22, v12, v11, 24 |
| ; GCN-NEXT: v_alignbit_b32 v21, v12, v11, 16 |
| ; GCN-NEXT: v_alignbit_b32 v30, v12, v11, 8 |
| ; GCN-NEXT: v_alignbit_b32 v27, v10, v9, 24 |
| ; GCN-NEXT: v_alignbit_b32 v28, v10, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v34, v10, v9, 8 |
| ; GCN-NEXT: v_alignbit_b32 v32, v8, v7, 24 |
| ; GCN-NEXT: v_alignbit_b32 v33, v8, v7, 16 |
| ; GCN-NEXT: v_alignbit_b32 v48, v8, v7, 8 |
| ; GCN-NEXT: v_alignbit_b32 v37, v6, v5, 24 |
| ; GCN-NEXT: v_alignbit_b32 v39, v6, v5, 16 |
| ; GCN-NEXT: v_alignbit_b32 v40, v6, v5, 8 |
| ; GCN-NEXT: v_alignbit_b32 v49, v4, v3, 24 |
| ; GCN-NEXT: v_alignbit_b32 v52, v4, v3, 16 |
| ; GCN-NEXT: v_alignbit_b32 v42, v4, v3, 8 |
| ; GCN-NEXT: v_alignbit_b32 v54, v2, v1, 24 |
| ; GCN-NEXT: v_alignbit_b32 v55, v2, v1, 16 |
| ; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 24, v16 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 24, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v50, 8, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v51, 24, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v56, 8, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v41, 24, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v46, 24, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v61, 8, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v60, 24, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2 |
| ; GCN-NEXT: .LBB42_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB42_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_f64 v[1:2], v[1:2], 1.0 |
| ; GCN-NEXT: v_add_f64 v[3:4], v[3:4], 1.0 |
| ; GCN-NEXT: v_add_f64 v[5:6], v[5:6], 1.0 |
| ; GCN-NEXT: v_add_f64 v[7:8], v[7:8], 1.0 |
| ; GCN-NEXT: v_add_f64 v[9:10], v[9:10], 1.0 |
| ; GCN-NEXT: v_add_f64 v[11:12], v[11:12], 1.0 |
| ; GCN-NEXT: v_add_f64 v[13:14], v[13:14], 1.0 |
| ; GCN-NEXT: v_add_f64 v[15:16], v[15:16], 1.0 |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 16 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v16, v15, 8 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 24 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v17, v14, v13, 16 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v24, v14, v13, 8 |
| ; GCN-NEXT: v_alignbit_b32 v22, v12, v11, 24 |
| ; GCN-NEXT: v_alignbit_b32 v21, v12, v11, 16 |
| ; GCN-NEXT: v_alignbit_b32 v30, v12, v11, 8 |
| ; GCN-NEXT: v_alignbit_b32 v27, v10, v9, 24 |
| ; GCN-NEXT: v_alignbit_b32 v28, v10, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v34, v10, v9, 8 |
| ; GCN-NEXT: v_alignbit_b32 v32, v8, v7, 24 |
| ; GCN-NEXT: v_alignbit_b32 v33, v8, v7, 16 |
| ; GCN-NEXT: v_alignbit_b32 v48, v8, v7, 8 |
| ; GCN-NEXT: v_alignbit_b32 v37, v6, v5, 24 |
| ; GCN-NEXT: v_alignbit_b32 v39, v6, v5, 16 |
| ; GCN-NEXT: v_alignbit_b32 v40, v6, v5, 8 |
| ; GCN-NEXT: v_alignbit_b32 v49, v4, v3, 24 |
| ; GCN-NEXT: v_alignbit_b32 v52, v4, v3, 16 |
| ; GCN-NEXT: v_alignbit_b32 v42, v4, v3, 8 |
| ; GCN-NEXT: v_alignbit_b32 v54, v2, v1, 24 |
| ; GCN-NEXT: v_alignbit_b32 v55, v2, v1, 16 |
| ; GCN-NEXT: v_alignbit_b32 v45, v2, v1, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 24, v16 |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v35, 8, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 24, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v50, 8, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v51, 24, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v56, 8, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v41, 24, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v43, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v59, 8, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v46, 24, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v47, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v61, 8, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v60, 24, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v62, 16, v2 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 8, v2 |
| ; GCN-NEXT: .LBB42_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v45 |
| ; GCN-NEXT: v_or_b32_e32 v45, v1, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v17 |
| ; GCN-NEXT: v_or_b32_e32 v18, v2, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v42 |
| ; GCN-NEXT: v_or_b32_e32 v42, v3, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v63 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v63, v4, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v40 |
| ; GCN-NEXT: v_or_b32_e32 v40, v5, v1 |
| ; GCN-NEXT: v_and_b32_e32 v55, 0xff, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v54, 24, v54 |
| ; GCN-NEXT: v_and_b32_e32 v62, 0xff, v62 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v60, 24, v60 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v61 |
| ; GCN-NEXT: v_or_b32_e32 v14, v6, v1 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0 |
| ; GCN-NEXT: v_and_b32_e32 v61, 0xff, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 24, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v48 |
| ; GCN-NEXT: v_or_b32_e32 v15, v7, v2 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0 |
| ; GCN-NEXT: v_and_b32_e32 v58, 0xff, v58 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v57, 24, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v59 |
| ; GCN-NEXT: v_or_b32_e32 v16, v8, v3 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 12, v0 |
| ; GCN-NEXT: v_and_b32_e32 v52, 0xff, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v34 |
| ; GCN-NEXT: v_or_b32_e32 v17, v9, v4 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 16, v0 |
| ; GCN-NEXT: v_and_b32_e32 v47, 0xff, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 24, v46 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v56 |
| ; GCN-NEXT: v_or_b32_e32 v34, v10, v5 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 20, v0 |
| ; GCN-NEXT: v_and_b32_e32 v46, 0xff, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v30 |
| ; GCN-NEXT: v_or_b32_e32 v30, v11, v6 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0 |
| ; GCN-NEXT: v_and_b32_e32 v43, 0xff, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 8, v44 |
| ; GCN-NEXT: v_or_b32_e32 v11, v12, v7 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 28, v0 |
| ; GCN-NEXT: v_and_b32_e32 v41, 0xff, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 24, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 8, v24 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v8 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 32, v0 |
| ; GCN-NEXT: v_and_b32_e32 v53, 0xff, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v48, 24, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v50 |
| ; GCN-NEXT: v_or_b32_e32 v24, v19, v9 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 36, v0 |
| ; GCN-NEXT: v_and_b32_e32 v50, 0xff, v21 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v22 |
| ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 8, v10 |
| ; GCN-NEXT: v_or_b32_e32 v21, v20, v10 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 40, v0 |
| ; GCN-NEXT: v_and_b32_e32 v51, 0xff, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 24, v36 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 8, v35 |
| ; GCN-NEXT: v_or_b32_e32 v23, v23, v12 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 44, v0 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v44, 0xff, v19 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v35, 24, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v55 |
| ; GCN-NEXT: v_or_b32_e32 v36, v54, v19 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 48, v0 |
| ; GCN-NEXT: v_and_b32_e32 v54, 0xff, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 24, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v62 |
| ; GCN-NEXT: v_or_b32_e32 v31, v60, v20 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 52, v0 |
| ; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v55, 0xff, v28 |
| ; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v38, 24, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v61 |
| ; GCN-NEXT: v_or_b32_e32 v56, v25, v28 |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 56, v0 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v59, 24, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v58 |
| ; GCN-NEXT: v_or_b32_e32 v57, v57, v25 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, 60, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v46 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 16, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v44 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; GCN-NEXT: v_and_b32_e32 v45, 0xffff, v45 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v42, 0xffff, v42 |
| ; GCN-NEXT: v_and_b32_e32 v58, 0xffff, v63 |
| ; GCN-NEXT: v_and_b32_e32 v40, 0xffff, v40 |
| ; GCN-NEXT: v_or_b32_e32 v37, v37, v52 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GCN-NEXT: v_or_b32_e32 v39, v39, v47 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_or_b32_e32 v32, v32, v46 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GCN-NEXT: v_or_b32_e32 v33, v33, v43 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: v_or_b32_e32 v27, v27, v41 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xffff, v34 |
| ; GCN-NEXT: v_or_b32_e32 v48, v48, v53 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30 |
| ; GCN-NEXT: v_or_b32_e32 v49, v49, v50 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_or_b32_e32 v22, v22, v51 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_or_b32_e32 v35, v35, v44 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24 |
| ; GCN-NEXT: v_or_b32_e32 v29, v29, v54 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21 |
| ; GCN-NEXT: v_or_b32_e32 v38, v38, v55 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23 |
| ; GCN-NEXT: v_or_b32_e32 v26, v59, v26 |
| ; GCN-NEXT: v_or_b32_e32 v36, v45, v36 |
| ; GCN-NEXT: v_or_b32_e32 v18, v18, v31 |
| ; GCN-NEXT: v_or_b32_e32 v31, v42, v56 |
| ; GCN-NEXT: v_or_b32_e32 v50, v58, v57 |
| ; GCN-NEXT: v_or_b32_e32 v37, v40, v37 |
| ; GCN-NEXT: v_or_b32_e32 v14, v14, v39 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v32 |
| ; GCN-NEXT: v_or_b32_e32 v16, v16, v33 |
| ; GCN-NEXT: v_or_b32_e32 v17, v17, v27 |
| ; GCN-NEXT: v_or_b32_e32 v27, v34, v48 |
| ; GCN-NEXT: v_or_b32_e32 v30, v30, v49 |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v22 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v35 |
| ; GCN-NEXT: v_or_b32_e32 v22, v24, v29 |
| ; GCN-NEXT: v_or_b32_e32 v21, v21, v38 |
| ; GCN-NEXT: v_or_b32_e32 v23, v23, v26 |
| ; GCN-NEXT: buffer_store_dword v36, v0, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v18, v1, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v31, v2, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v50, v3, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v37, v4, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v14, v5, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v15, v6, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v16, v7, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v17, v8, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v27, v9, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v30, v10, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v11, v12, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v13, v19, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v22, v20, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v21, v28, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v23, v25, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v8f64_to_v64i8: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr49 |
| ; VI-NEXT: ; implicit-def: $vgpr48 |
| ; VI-NEXT: ; implicit-def: $vgpr39 |
| ; VI-NEXT: ; implicit-def: $vgpr38 |
| ; VI-NEXT: ; implicit-def: $vgpr37 |
| ; VI-NEXT: ; implicit-def: $vgpr36 |
| ; VI-NEXT: ; implicit-def: $vgpr35 |
| ; VI-NEXT: ; implicit-def: $vgpr34 |
| ; VI-NEXT: ; implicit-def: $vgpr33 |
| ; VI-NEXT: ; implicit-def: $vgpr32 |
| ; VI-NEXT: ; implicit-def: $vgpr31 |
| ; VI-NEXT: ; implicit-def: $vgpr30 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB42_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v16 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8] |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v16 |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; VI-NEXT: v_mov_b32_e32 v26, v22 |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4] |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2] |
| ; VI-NEXT: .LBB42_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB42_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_add_f64 v[15:16], v[15:16], 1.0 |
| ; VI-NEXT: v_add_f64 v[13:14], v[13:14], 1.0 |
| ; VI-NEXT: v_add_f64 v[11:12], v[11:12], 1.0 |
| ; VI-NEXT: v_add_f64 v[9:10], v[9:10], 1.0 |
| ; VI-NEXT: v_add_f64 v[3:4], v[3:4], 1.0 |
| ; VI-NEXT: v_add_f64 v[1:2], v[1:2], 1.0 |
| ; VI-NEXT: v_add_f64 v[7:8], v[7:8], 1.0 |
| ; VI-NEXT: v_add_f64 v[5:6], v[5:6], 1.0 |
| ; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4] |
| ; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2] |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8] |
| ; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 8, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: .LBB42_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v29 |
| ; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v22 |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v22, v27, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v28 |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v62 |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61 |
| ; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v59 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v57 |
| ; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v56 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v46 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44 |
| ; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v43 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v24 |
| ; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v55 |
| ; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v54 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v52 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v50 |
| ; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v49 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; VI-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37 |
| ; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; VI-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32 |
| ; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; VI-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v26 |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 60, v0 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v8f64_to_v64i8: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB42_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v16 |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_mov_b32_e32 v26, v23 |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: .LBB42_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB42_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: v_add_f64 v[15:16], v[15:16], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[13:14], v[13:14], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[11:12], v[11:12], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[9:10], v[9:10], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[7:8], v[7:8], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[3:4], v[3:4], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[1:2], v[1:2], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[5:6], v[5:6], 1.0 |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: .LBB42_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v29 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v23 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v23, v27, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28 |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v22 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v60 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v59 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v57 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v56 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v46 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v21 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v55 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v50 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v49 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56 |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v26 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60 |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v8f64_to_v64i8: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB42_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB42_2: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB42_4 |
| ; GFX11-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-NEXT: v_add_f64 v[15:16], v[15:16], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[13:14], v[13:14], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[11:12], v[11:12], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[9:10], v[9:10], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[7:8], v[7:8], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[5:6], v[5:6], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[3:4], v[3:4], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[1:2], v[1:2], 1.0 |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB42_4: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v96, 0xff, v96 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v86, 0xff, v86 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v87 |
| ; GFX11-NEXT: v_or_b32_e32 v24, v96, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v25 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v86, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84 |
| ; GFX11-NEXT: v_and_b32_e32 v83, 0xff, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v82 |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v80, 8, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v71, 8, v71 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v24 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v81 |
| ; GFX11-NEXT: v_and_b32_e32 v70, 0xff, v70 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v83, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v25, v80 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v71 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v70, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v22 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v66 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v65 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v55 |
| ; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v23, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v25, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v52 |
| ; GFX11-NEXT: v_lshlrev_b16 v20, 8, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v50 |
| ; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v48, 8, v48 |
| ; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v39 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v23, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v25, v49 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v48 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v39, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v38 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v37 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v35 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v34 |
| ; GFX11-NEXT: v_lshlrev_b16 v18, 8, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_lshlrev_b16 v69, 8, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v68, 0xff, v68 |
| ; GFX11-NEXT: v_lshlrev_b16 v67, 8, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v23, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v32 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v30 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v29 |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v26, 8, v26 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v69 |
| ; GFX11-NEXT: v_or_b32_e32 v67, v68, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v23, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v25, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v67 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; GFX11-NEXT: s_clause 0x3 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off |
| ; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <8 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <8 x double> %a1 to <64 x i8> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <8 x double> %a to <64 x i8> |
| br label %end |
| |
| end: |
| %phi = phi <64 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <64 x i8> %phi |
| } |
| |
| define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v64i8_to_v8f64: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_mov_b32_e32 v38, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v31, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:112 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104 |
| ; GCN-NEXT: s_waitcnt expcnt(3) |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:100 |
| ; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:96 |
| ; GCN-NEXT: s_waitcnt expcnt(2) |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:92 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88 |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:84 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76 |
| ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:72 |
| ; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:68 |
| ; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64 |
| ; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:60 |
| ; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56 |
| ; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:52 |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:8 |
| ; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:132 |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:128 |
| ; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:124 |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:120 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v52 |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 |
| ; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:36 |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 |
| ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:28 |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:24 |
| ; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:20 |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 8, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v3 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v41, 8, v5 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v7 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v40, 8, v9 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v11 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 8, v13 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v15 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 8, v17 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v19 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v21 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v23 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v25 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v27 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v45 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 8, v44 |
| ; GCN-NEXT: s_waitcnt vmcnt(10) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v59, 24, v59 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v58 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v57, 24, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v58, 24, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v56, 8, v14 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v12 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v10 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v6 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v4 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v2 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:44 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:12 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116 |
| ; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v0 |
| ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v46, 24, v46 |
| ; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB43_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xff, v31 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v42 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v41 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v40 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v34 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v55 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v35 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v36 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v37 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v38 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v50 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v36, 0xff, v54 |
| ; GCN-NEXT: v_and_b32_e32 v37, 0xff, v53 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xff, v52 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v49 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v48 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v39 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v63 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v62 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v61 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v60 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v21 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v32 |
| ; GCN-NEXT: v_or_b32_e32 v17, v33, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v34 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v20, v20, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v23, v24, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26 |
| ; GCN-NEXT: v_or_b32_e32 v26, v28, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v30 |
| ; GCN-NEXT: v_or_b32_e32 v27, v35, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_or_b32_e32 v29, v36, v44 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v37 |
| ; GCN-NEXT: v_or_b32_e32 v31, v38, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GCN-NEXT: v_or_b32_e32 v18, v18, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v7, v32, v7 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v6, v32, v6 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v5, v32, v5 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v32, v4 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v32, v21 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v22, v32, v22 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v24, v32, v24 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v28, v32, v28 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27 |
| ; GCN-NEXT: v_or_b32_e32 v25, v59, v25 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_or_b32_e32 v30, v57, v30 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31 |
| ; GCN-NEXT: v_or_b32_e32 v32, v58, v8 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v9 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v34, v8, v10 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xffff, v11 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v8, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v14, v8, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v16, v8, v16 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GCN-NEXT: v_or_b32_e32 v19, v46, v19 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v7 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v6 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v5 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; GCN-NEXT: v_or_b32_e32 v4, v17, v21 |
| ; GCN-NEXT: v_or_b32_e32 v5, v20, v22 |
| ; GCN-NEXT: v_or_b32_e32 v6, v23, v24 |
| ; GCN-NEXT: v_or_b32_e32 v7, v26, v28 |
| ; GCN-NEXT: v_or_b32_e32 v8, v27, v25 |
| ; GCN-NEXT: v_or_b32_e32 v9, v29, v30 |
| ; GCN-NEXT: v_or_b32_e32 v10, v31, v32 |
| ; GCN-NEXT: v_or_b32_e32 v11, v33, v34 |
| ; GCN-NEXT: v_or_b32_e32 v12, v35, v12 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; GCN-NEXT: v_or_b32_e32 v14, v15, v16 |
| ; GCN-NEXT: v_or_b32_e32 v15, v18, v19 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; kill: killed $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; kill: killed $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; kill: killed $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: .LBB43_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB43_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v31 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GCN-NEXT: v_or_b32_e32 v0, v42, v0 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v32 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GCN-NEXT: v_or_b32_e32 v1, v41, v1 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v33 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GCN-NEXT: v_or_b32_e32 v2, v40, v2 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v34 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GCN-NEXT: v_or_b32_e32 v3, v55, v3 |
| ; GCN-NEXT: s_movk_i32 s7, 0x300 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v35 |
| ; GCN-NEXT: s_mov_b32 s6, 0x3000000 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v36 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v37 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v16 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v18 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v20 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v22 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v24 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v26 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v28 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v30 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v50 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v25 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v54 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v53 |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v52 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v23 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v49 |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v48 |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v39 |
| ; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v63 |
| ; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v62 |
| ; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v61 |
| ; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v60 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v21 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v19 |
| ; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v51 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v32 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xff, v33 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v34 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v7 |
| ; GCN-NEXT: v_or_b32_e32 v6, v17, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v9 |
| ; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v11 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v13 |
| ; GCN-NEXT: v_or_b32_e32 v13, v29, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15 |
| ; GCN-NEXT: v_or_b32_e32 v15, v27, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v18 |
| ; GCN-NEXT: v_or_b32_e32 v17, v44, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v22 |
| ; GCN-NEXT: v_or_b32_e32 v20, v45, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v23 |
| ; GCN-NEXT: v_or_b32_e32 v23, v56, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v26 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v25, v25, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v30 |
| ; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v27, v27, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v32 |
| ; GCN-NEXT: v_or_b32_e32 v29, v43, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GCN-NEXT: v_or_b32_e32 v19, v47, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v34 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x300, v0 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v31, v31, v36 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s7, v1 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v8, v32, v8 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v5, v32, v5 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v32, v4 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s7, v6 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v7, v32, v7 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v10, v32, v10 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v32, v12 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v14, v32, v14 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, s7, v15 |
| ; GCN-NEXT: v_or_b32_e32 v16, v59, v16 |
| ; GCN-NEXT: v_add_i32_e32 v17, vcc, s7, v17 |
| ; GCN-NEXT: v_or_b32_e32 v18, v57, v18 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20 |
| ; GCN-NEXT: v_or_b32_e32 v22, v58, v22 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, s7, v23 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v24, v32, v24 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v26, v32, v26 |
| ; GCN-NEXT: v_add_i32_e32 v27, vcc, s7, v27 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v28, v32, v28 |
| ; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v32, v21 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 0x300, v19 |
| ; GCN-NEXT: v_or_b32_e32 v30, v46, v30 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GCN-NEXT: v_or_b32_e32 v0, v31, v0 |
| ; GCN-NEXT: v_or_b32_e32 v1, v8, v1 |
| ; GCN-NEXT: v_or_b32_e32 v2, v5, v2 |
| ; GCN-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; GCN-NEXT: v_or_b32_e32 v4, v7, v6 |
| ; GCN-NEXT: v_or_b32_e32 v5, v10, v9 |
| ; GCN-NEXT: v_or_b32_e32 v6, v12, v11 |
| ; GCN-NEXT: v_or_b32_e32 v7, v14, v13 |
| ; GCN-NEXT: v_or_b32_e32 v8, v16, v15 |
| ; GCN-NEXT: v_or_b32_e32 v9, v18, v17 |
| ; GCN-NEXT: v_or_b32_e32 v10, v22, v20 |
| ; GCN-NEXT: v_or_b32_e32 v11, v24, v23 |
| ; GCN-NEXT: v_or_b32_e32 v12, v26, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v28, v27 |
| ; GCN-NEXT: v_or_b32_e32 v14, v21, v29 |
| ; GCN-NEXT: v_or_b32_e32 v15, v30, v19 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v3 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v13 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x3000000, v14 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x3000000, v15 |
| ; GCN-NEXT: .LBB43_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v64i8_to_v8f64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132 |
| ; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72 |
| ; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80 |
| ; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88 |
| ; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96 |
| ; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104 |
| ; VI-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112 |
| ; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120 |
| ; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128 |
| ; VI-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20 |
| ; VI-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12 |
| ; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100 |
| ; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68 |
| ; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36 |
| ; VI-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4 |
| ; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v26, 8, v23 |
| ; VI-NEXT: v_lshlrev_b16_e32 v28, 8, v25 |
| ; VI-NEXT: v_lshlrev_b16_e32 v38, 8, v27 |
| ; VI-NEXT: v_lshlrev_b16_e32 v62, 8, v29 |
| ; VI-NEXT: v_lshlrev_b16_e32 v18, 8, v17 |
| ; VI-NEXT: v_lshlrev_b16_e32 v20, 8, v19 |
| ; VI-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116 |
| ; VI-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124 |
| ; VI-NEXT: v_lshlrev_b16_e32 v31, 8, v1 |
| ; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v3 |
| ; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v5 |
| ; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v7 |
| ; VI-NEXT: v_lshlrev_b16_e32 v35, 8, v9 |
| ; VI-NEXT: v_lshlrev_b16_e32 v36, 8, v11 |
| ; VI-NEXT: v_lshlrev_b16_e32 v37, 8, v13 |
| ; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v15 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54 |
| ; VI-NEXT: v_lshlrev_b16_e32 v61, 8, v0 |
| ; VI-NEXT: v_lshlrev_b16_e32 v59, 8, v2 |
| ; VI-NEXT: v_lshlrev_b16_e32 v58, 8, v4 |
| ; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v6 |
| ; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v8 |
| ; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v10 |
| ; VI-NEXT: v_lshlrev_b16_e32 v43, 8, v12 |
| ; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v14 |
| ; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v24 |
| ; VI-NEXT: v_lshlrev_b16_e32 v55, 8, v30 |
| ; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v53 |
| ; VI-NEXT: s_waitcnt vmcnt(13) |
| ; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v42 |
| ; VI-NEXT: s_waitcnt vmcnt(12) |
| ; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v44 |
| ; VI-NEXT: s_waitcnt vmcnt(11) |
| ; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v47 |
| ; VI-NEXT: s_waitcnt vmcnt(10) |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v57 |
| ; VI-NEXT: s_waitcnt vmcnt(9) |
| ; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v60 |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_lshlrev_b16_e32 v57, 8, v63 |
| ; VI-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108 |
| ; VI-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92 |
| ; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84 |
| ; VI-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76 |
| ; VI-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60 |
| ; VI-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44 |
| ; VI-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB43_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr30 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr23 |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(7) |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr31 |
| ; VI-NEXT: ; implicit-def: $vgpr32 |
| ; VI-NEXT: ; implicit-def: $vgpr33 |
| ; VI-NEXT: ; implicit-def: $vgpr34 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr38 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr35 |
| ; VI-NEXT: ; implicit-def: $vgpr36 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr37 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr49 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr48 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr39 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: .LBB43_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB43_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: v_mov_b32_e32 v15, 0x300 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v51 |
| ; VI-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v9, v9, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v24 |
| ; VI-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v10, v10, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v30 |
| ; VI-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v11, v11, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v60 |
| ; VI-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v12, v12, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v47 |
| ; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v13, v13, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v42 |
| ; VI-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v14, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v17, 3, v17 |
| ; VI-NEXT: v_or_b32_sdwa v17, v57, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: s_waitcnt vmcnt(7) |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v1 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v0, 0x300, v0 |
| ; VI-NEXT: v_add_u16_sdwa v1, v1, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; VI-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; VI-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v4, v4, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; VI-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v5, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; VI-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v6, v6, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; VI-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v7, v7, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v53 |
| ; VI-NEXT: v_or_b32_sdwa v16, v23, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v16, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v8 |
| ; VI-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v8, v8, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; VI-NEXT: v_add_u16_sdwa v3, v3, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: v_add_u16_sdwa v15, v17, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v3 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; VI-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v4, 0x300, v4 |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; VI-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v5, 0x300, v5 |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; VI-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v6, 0x300, v6 |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; VI-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v7, 0x300, v7 |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v52 |
| ; VI-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v8, 0x300, v8 |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v50 |
| ; VI-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v9, 0x300, v9 |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v49 |
| ; VI-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v10, 0x300, v10 |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v63 |
| ; VI-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v11, 0x300, v11 |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v48 |
| ; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v12, 0x300, v12 |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v44 |
| ; VI-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v13, 0x300, v13 |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v39 |
| ; VI-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v14, 0x300, v14 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v16 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v19 |
| ; VI-NEXT: v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v16, 0x300, v16 |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: .LBB43_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v64i8_to_v8f64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132 |
| ; GFX9-NEXT: s_nop 0 |
| ; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:64 |
| ; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:72 |
| ; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:80 |
| ; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:88 |
| ; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:96 |
| ; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:104 |
| ; GFX9-NEXT: buffer_load_ushort v57, off, s[0:3], s32 offset:112 |
| ; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:120 |
| ; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:128 |
| ; GFX9-NEXT: buffer_load_ushort v50, off, s[0:3], s32 offset:20 |
| ; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:12 |
| ; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:100 |
| ; GFX9-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:68 |
| ; GFX9-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:36 |
| ; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v21 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v26, 8, v23 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v28, 8, v25 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v38, 8, v27 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v62, 8, v29 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v18, 8, v17 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v19 |
| ; GFX9-NEXT: buffer_load_ushort v19, off, s[0:3], s32 offset:116 |
| ; GFX9-NEXT: buffer_load_ushort v17, off, s[0:3], s32 offset:124 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v31, 8, v1 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v32, 8, v3 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v5 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v34, 8, v7 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v35, 8, v9 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v11 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v37, 8, v13 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v16, 8, v15 |
| ; GFX9-NEXT: s_waitcnt vmcnt(25) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v54 |
| ; GFX9-NEXT: s_waitcnt vmcnt(24) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v61, 8, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(23) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v59, 8, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(22) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v58, 8, v4 |
| ; GFX9-NEXT: s_waitcnt vmcnt(21) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(20) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v8 |
| ; GFX9-NEXT: s_waitcnt vmcnt(19) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v10 |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v12 |
| ; GFX9-NEXT: s_waitcnt vmcnt(17) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v14 |
| ; GFX9-NEXT: s_waitcnt vmcnt(16) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v24 |
| ; GFX9-NEXT: s_waitcnt vmcnt(15) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v30 |
| ; GFX9-NEXT: s_waitcnt vmcnt(14) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v53 |
| ; GFX9-NEXT: s_waitcnt vmcnt(13) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v42 |
| ; GFX9-NEXT: s_waitcnt vmcnt(12) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v44 |
| ; GFX9-NEXT: s_waitcnt vmcnt(11) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v47 |
| ; GFX9-NEXT: s_waitcnt vmcnt(10) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v57 |
| ; GFX9-NEXT: s_waitcnt vmcnt(9) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v60 |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v57, 8, v63 |
| ; GFX9-NEXT: buffer_load_ushort v53, off, s[0:3], s32 offset:108 |
| ; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:92 |
| ; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:84 |
| ; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:76 |
| ; GFX9-NEXT: buffer_load_ushort v60, off, s[0:3], s32 offset:60 |
| ; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:52 |
| ; GFX9-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:44 |
| ; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB43_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v0, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: .LBB43_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB43_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_movk_i32 s6, 0x300 |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_add_u16_e32 v9, 3, v51 |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_add_u16_e32 v10, 3, v24 |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v11, 3, v30 |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v12, 3, v60 |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v13, 3, v47 |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v14, 3, v42 |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v15, 3, v53 |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_add_u16_e32 v1, 3, v1 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v34, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GFX9-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v16, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v20, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_add_u16_e32 v16, 3, v17 |
| ; GFX9-NEXT: v_or_b32_sdwa v16, v57, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_add_u16_e32 v8, 3, v8 |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v36, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v2 |
| ; GFX9-NEXT: v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v37, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v3 |
| ; GFX9-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v18, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v4 |
| ; GFX9-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v22, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v5 |
| ; GFX9-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v28, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v6 |
| ; GFX9-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v7 |
| ; GFX9-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; GFX9-NEXT: v_add_u16_e32 v8, 3, v52 |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v8 |
| ; GFX9-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; GFX9-NEXT: v_add_u16_e32 v9, 3, v50 |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v9 |
| ; GFX9-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; GFX9-NEXT: v_add_u16_e32 v10, 3, v49 |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v10 |
| ; GFX9-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; GFX9-NEXT: v_add_u16_e32 v11, 3, v63 |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v11 |
| ; GFX9-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GFX9-NEXT: v_add_u16_e32 v12, 3, v48 |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v12 |
| ; GFX9-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; GFX9-NEXT: v_add_u16_e32 v13, 3, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v13 |
| ; GFX9-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; GFX9-NEXT: v_add_u16_e32 v14, 3, v39 |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v14 |
| ; GFX9-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; GFX9-NEXT: v_add_u16_e32 v15, 3, v19 |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v15 |
| ; GFX9-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; GFX9-NEXT: .LBB43_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v64i8_to_v8f64: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12 |
| ; GFX11-NEXT: v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8 |
| ; GFX11-NEXT: v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4 |
| ; GFX11-NEXT: v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0 |
| ; GFX11-NEXT: s_clause 0x1f |
| ; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:128 |
| ; GFX11-NEXT: scratch_load_u16 v39, off, s32 offset:124 |
| ; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:120 |
| ; GFX11-NEXT: scratch_load_u16 v48, off, s32 offset:116 |
| ; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:112 |
| ; GFX11-NEXT: scratch_load_u16 v49, off, s32 offset:108 |
| ; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:104 |
| ; GFX11-NEXT: scratch_load_u16 v50, off, s32 offset:100 |
| ; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:96 |
| ; GFX11-NEXT: scratch_load_u16 v51, off, s32 offset:92 |
| ; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:88 |
| ; GFX11-NEXT: scratch_load_b32 v12, off, s32 offset:132 |
| ; GFX11-NEXT: scratch_load_u16 v14, off, s32 |
| ; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:8 |
| ; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:16 |
| ; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:24 |
| ; GFX11-NEXT: scratch_load_u16 v83, off, s32 offset:32 |
| ; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:40 |
| ; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:48 |
| ; GFX11-NEXT: scratch_load_u16 v128, off, s32 offset:56 |
| ; GFX11-NEXT: scratch_load_u16 v129, off, s32 offset:64 |
| ; GFX11-NEXT: scratch_load_u16 v130, off, s32 offset:72 |
| ; GFX11-NEXT: scratch_load_u16 v131, off, s32 offset:80 |
| ; GFX11-NEXT: scratch_load_u16 v52, off, s32 offset:84 |
| ; GFX11-NEXT: scratch_load_u16 v53, off, s32 offset:76 |
| ; GFX11-NEXT: scratch_load_u16 v54, off, s32 offset:68 |
| ; GFX11-NEXT: scratch_load_u16 v55, off, s32 offset:60 |
| ; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:52 |
| ; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:44 |
| ; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:36 |
| ; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:28 |
| ; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:20 |
| ; GFX11-NEXT: s_clause 0x1 |
| ; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:12 |
| ; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:4 |
| ; GFX11-NEXT: v_lshlrev_b16 v118, 8, v1 |
| ; GFX11-NEXT: v_lshlrev_b16 v119, 8, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v117, 8, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v114, 8, v7 |
| ; GFX11-NEXT: v_lshlrev_b16 v115, 8, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v116, 8, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v113, 8, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v100, 8, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v101, 8, v17 |
| ; GFX11-NEXT: v_lshlrev_b16 v102, 8, v19 |
| ; GFX11-NEXT: v_lshlrev_b16 v103, 8, v21 |
| ; GFX11-NEXT: v_lshlrev_b16 v112, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v87, 8, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v96, 8, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v97, 8, v29 |
| ; GFX11-NEXT: s_waitcnt vmcnt(33) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(31) |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v2 |
| ; GFX11-NEXT: s_waitcnt vmcnt(29) |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v4 |
| ; GFX11-NEXT: s_waitcnt vmcnt(27) |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v6 |
| ; GFX11-NEXT: s_waitcnt vmcnt(25) |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v8 |
| ; GFX11-NEXT: s_waitcnt vmcnt(22) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12 |
| ; GFX11-NEXT: s_waitcnt vmcnt(21) |
| ; GFX11-NEXT: v_lshlrev_b16 v98, 8, v14 |
| ; GFX11-NEXT: s_waitcnt vmcnt(20) |
| ; GFX11-NEXT: v_lshlrev_b16 v99, 8, v65 |
| ; GFX11-NEXT: s_waitcnt vmcnt(19) |
| ; GFX11-NEXT: v_lshlrev_b16 v81, 8, v66 |
| ; GFX11-NEXT: s_waitcnt vmcnt(18) |
| ; GFX11-NEXT: v_lshlrev_b16 v82, 8, v67 |
| ; GFX11-NEXT: s_waitcnt vmcnt(17) |
| ; GFX11-NEXT: v_lshlrev_b16 v83, 8, v83 |
| ; GFX11-NEXT: s_waitcnt vmcnt(16) |
| ; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84 |
| ; GFX11-NEXT: s_waitcnt vmcnt(15) |
| ; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85 |
| ; GFX11-NEXT: s_waitcnt vmcnt(14) |
| ; GFX11-NEXT: v_lshlrev_b16 v27, 8, v128 |
| ; GFX11-NEXT: s_waitcnt vmcnt(13) |
| ; GFX11-NEXT: v_lshlrev_b16 v29, 8, v129 |
| ; GFX11-NEXT: s_waitcnt vmcnt(12) |
| ; GFX11-NEXT: v_lshlrev_b16 v65, 8, v130 |
| ; GFX11-NEXT: s_waitcnt vmcnt(11) |
| ; GFX11-NEXT: v_lshlrev_b16 v66, 8, v131 |
| ; GFX11-NEXT: v_lshlrev_b16 v67, 8, v10 |
| ; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB43_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB43_4 |
| ; GFX11-NEXT: .LBB43_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB43_3: ; %cmp.false |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v32 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v37 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v34 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v118 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v119 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v35 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v117 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v6, v113 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v114 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v115 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v116 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v101 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v38 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v102 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v103 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v112 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v100 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v1 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v4, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v6, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v7, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v9, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v30 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v86 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v71 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v70 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v68 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v87 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v96 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v97 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v98 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v99 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v81 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v82 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v8, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v10, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v12, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v14, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v55 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v53 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v52 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v50 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v49 |
| ; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v48 |
| ; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v39 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v27 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v29 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v65 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v66 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v18, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v22, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v24, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v13, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v15, v16 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v17, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v19, v20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr16 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr118 |
| ; GFX11-NEXT: ; implicit-def: $vgpr119 |
| ; GFX11-NEXT: ; implicit-def: $vgpr117 |
| ; GFX11-NEXT: ; implicit-def: $vgpr114 |
| ; GFX11-NEXT: ; implicit-def: $vgpr115 |
| ; GFX11-NEXT: ; implicit-def: $vgpr116 |
| ; GFX11-NEXT: ; implicit-def: $vgpr113 |
| ; GFX11-NEXT: ; implicit-def: $vgpr100 |
| ; GFX11-NEXT: ; implicit-def: $vgpr101 |
| ; GFX11-NEXT: ; implicit-def: $vgpr102 |
| ; GFX11-NEXT: ; implicit-def: $vgpr103 |
| ; GFX11-NEXT: ; implicit-def: $vgpr112 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr97 |
| ; GFX11-NEXT: ; implicit-def: $vgpr98 |
| ; GFX11-NEXT: ; implicit-def: $vgpr99 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB43_2 |
| ; GFX11-NEXT: .LBB43_4: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u16 v0, v31, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v32, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v33, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v34, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v35, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v5, v36, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v6, v37, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v118, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v119, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GFX11-NEXT: v_add_nc_u16 v0, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v117, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v114, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v115, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v116, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v113, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v7, v16, 3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, 0x300, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v5 |
| ; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v6, v38, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v8, v18, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v9, v20, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v10, v22, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v101, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v100, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v102, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v103, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v112, v10 |
| ; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7 |
| ; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8 |
| ; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9 |
| ; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v3, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v5, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v7, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v9, v10 |
| ; GFX11-NEXT: v_add_nc_u16 v6, v24, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v7, v26, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v8, v28, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v9, v30, 3 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_add_nc_u16 v10, v86, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v11, v80, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v12, v71, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v13, v70, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v14, v69, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v15, v68, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v87, v6 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v96, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v97, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v98, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v99, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v81, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v82, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v83, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v84, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v85, v15 |
| ; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v6 |
| ; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v7 |
| ; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v8 |
| ; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v9 |
| ; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v10 |
| ; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11 |
| ; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12 |
| ; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13 |
| ; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14 |
| ; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v8, v9 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v10, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v12, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v14, v15 |
| ; GFX11-NEXT: v_add_nc_u16 v11, v64, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v12, v55, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v13, v54, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v14, v53, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v15, v52, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v16, v51, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v18, v50, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v20, v49, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v22, v48, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v24, v39, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v27, v11 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v29, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v65, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v66, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v67, v15 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v19, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v21, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v23, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v25, v24 |
| ; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v11 |
| ; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v12 |
| ; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v13 |
| ; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v14 |
| ; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v15 |
| ; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v16 |
| ; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17 |
| ; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v18 |
| ; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v19 |
| ; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v13, v14 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v15, v16 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v17, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v19, v20 |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <64 x i8> %a, splat (i8 3) |
| %a2 = bitcast <64 x i8> %a1 to <8 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <64 x i8> %a to <8 x double> |
| br label %end |
| |
| end: |
| %phi = phi <8 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <8 x double> %phi |
| } |
| |
| define <32 x half> @bitcast_v32i16_to_v32f16(<32 x i16> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32i16_to_v32f16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: v_mov_b32_e32 v62, v30 |
| ; GCN-NEXT: v_mov_b32_e32 v61, v29 |
| ; GCN-NEXT: v_mov_b32_e32 v60, v28 |
| ; GCN-NEXT: v_mov_b32_e32 v59, v27 |
| ; GCN-NEXT: v_mov_b32_e32 v58, v26 |
| ; GCN-NEXT: v_mov_b32_e32 v57, v25 |
| ; GCN-NEXT: v_mov_b32_e32 v56, v24 |
| ; GCN-NEXT: v_mov_b32_e32 v47, v23 |
| ; GCN-NEXT: v_mov_b32_e32 v46, v22 |
| ; GCN-NEXT: v_mov_b32_e32 v45, v21 |
| ; GCN-NEXT: v_mov_b32_e32 v44, v20 |
| ; GCN-NEXT: v_mov_b32_e32 v43, v19 |
| ; GCN-NEXT: v_mov_b32_e32 v42, v18 |
| ; GCN-NEXT: v_mov_b32_e32 v41, v17 |
| ; GCN-NEXT: v_mov_b32_e32 v40, v16 |
| ; GCN-NEXT: v_mov_b32_e32 v55, v15 |
| ; GCN-NEXT: v_mov_b32_e32 v54, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v53, v13 |
| ; GCN-NEXT: v_mov_b32_e32 v52, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v51, v11 |
| ; GCN-NEXT: v_mov_b32_e32 v50, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v49, v9 |
| ; GCN-NEXT: v_mov_b32_e32 v48, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v39, v7 |
| ; GCN-NEXT: v_mov_b32_e32 v38, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v5 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v3 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v1 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB44_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v53 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v55 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v40 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v41 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v42 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v43 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v44 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v45 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v46 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v47 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v56 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v57 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v58 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v59 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v60 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v61 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v62 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v63 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: .LBB44_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB44_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v63 |
| ; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v62 |
| ; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v61 |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v60 |
| ; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v59 |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v58 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v57 |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v56 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v47 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v46 |
| ; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v45 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v44 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v43 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v42 |
| ; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v41 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v40 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v55 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v54 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v53 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v52 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v51 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v50 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v49 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v48 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v39 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v37 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v36 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v35 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v34 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v33 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31 |
| ; GCN-NEXT: .LBB44_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32i16_to_v32f16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB44_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v16, 3 |
| ; VI-NEXT: v_add_u16_sdwa v19, v15, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v15, 3, v15 |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v14, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v14 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v13, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v13 |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v12, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v12 |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v11, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v11 |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v10, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v10 |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v9, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v9 |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v8, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v8 |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v7, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v6, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; VI-NEXT: v_add_u16_sdwa v17, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v18, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v16, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v1 |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v16 |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v19 |
| ; VI-NEXT: v_or_b32_e32 v1, v1, v18 |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v17 |
| ; VI-NEXT: .LBB44_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32i16_to_v32f16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB44_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB44_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32i16_to_v32f16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB44_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: .LBB44_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <32 x i16> %a, splat (i16 3) |
| %a2 = bitcast <32 x i16> %a1 to <32 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x i16> %a to <32 x half> |
| br label %end |
| |
| end: |
| %phi = phi <32 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x half> %phi |
| } |
| |
| define <32 x i16> @bitcast_v32f16_to_v32i16(<32 x half> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32f16_to_v32i16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v30, v30 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v31, v32 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB45_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.true |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30 |
| ; GCN-NEXT: v_add_f32_e32 v31, 0x38000000, v31 |
| ; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v31, v31 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v30, v30 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v31 |
| ; GCN-NEXT: v_or_b32_e32 v30, v30, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26 |
| ; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v27 |
| ; GCN-NEXT: v_or_b32_e32 v26, v26, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v23 |
| ; GCN-NEXT: v_or_b32_e32 v22, v22, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v19 |
| ; GCN-NEXT: v_or_b32_e32 v18, v18, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v15 |
| ; GCN-NEXT: v_or_b32_e32 v14, v14, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v11 |
| ; GCN-NEXT: v_or_b32_e32 v10, v10, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v7 |
| ; GCN-NEXT: v_or_b32_e32 v6, v6, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v3 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GCN-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; GCN-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; GCN-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; GCN-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; GCN-NEXT: v_or_b32_e32 v20, v20, v21 |
| ; GCN-NEXT: v_or_b32_e32 v24, v24, v25 |
| ; GCN-NEXT: v_or_b32_e32 v28, v28, v29 |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v6, v5, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v10, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v14, v13, 16 |
| ; GCN-NEXT: v_alignbit_b32 v17, v18, v17, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v22, v21, 16 |
| ; GCN-NEXT: v_alignbit_b32 v25, v26, v25, 16 |
| ; GCN-NEXT: v_alignbit_b32 v29, v30, v29, 16 |
| ; GCN-NEXT: .LBB45_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32f16_to_v32i16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB45_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v17, 0x200 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v15 |
| ; VI-NEXT: v_add_f16_sdwa v15, v15, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v15, v19, v15 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v14 |
| ; VI-NEXT: v_add_f16_sdwa v14, v14, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v14, v19, v14 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v13 |
| ; VI-NEXT: v_add_f16_sdwa v13, v13, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v13, v19, v13 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v12 |
| ; VI-NEXT: v_add_f16_sdwa v12, v12, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v12, v19, v12 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v11 |
| ; VI-NEXT: v_add_f16_sdwa v11, v11, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v11, v19, v11 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v10 |
| ; VI-NEXT: v_add_f16_sdwa v10, v10, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v10, v19, v10 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v9 |
| ; VI-NEXT: v_add_f16_sdwa v9, v9, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v9, v19, v9 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v8 |
| ; VI-NEXT: v_add_f16_sdwa v8, v8, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v8, v19, v8 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v7 |
| ; VI-NEXT: v_add_f16_sdwa v7, v7, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v7, v19, v7 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v6 |
| ; VI-NEXT: v_add_f16_sdwa v6, v6, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v6, v19, v6 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v5 |
| ; VI-NEXT: v_add_f16_sdwa v5, v5, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v5, v19, v5 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v4 |
| ; VI-NEXT: v_add_f16_sdwa v4, v4, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v16, 0x200, v0 |
| ; VI-NEXT: v_add_f16_sdwa v0, v0, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v18, 0x200, v1 |
| ; VI-NEXT: v_add_f16_sdwa v1, v1, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v4, v19, v4 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v2 |
| ; VI-NEXT: v_add_f16_sdwa v2, v2, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_sdwa v17, v3, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, 0x200, v3 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v17 |
| ; VI-NEXT: v_or_b32_e32 v2, v19, v2 |
| ; VI-NEXT: v_or_b32_e32 v1, v18, v1 |
| ; VI-NEXT: v_or_b32_e32 v0, v16, v0 |
| ; VI-NEXT: .LBB45_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32f16_to_v32i16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB45_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: s_movk_i32 s6, 0x200 |
| ; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB45_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32f16_to_v32i16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB45_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-NEXT: .LBB45_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <32 x half> %a1 to <32 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x half> %a to <32 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x i16> %phi |
| } |
| |
| define <32 x bfloat> @bitcast_v32i16_to_v32bf16(<32 x i16> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32i16_to_v32bf16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v55, v30 |
| ; GCN-NEXT: v_mov_b32_e32 v54, v28 |
| ; GCN-NEXT: v_mov_b32_e32 v53, v26 |
| ; GCN-NEXT: v_mov_b32_e32 v52, v24 |
| ; GCN-NEXT: v_mov_b32_e32 v51, v22 |
| ; GCN-NEXT: v_mov_b32_e32 v50, v20 |
| ; GCN-NEXT: v_mov_b32_e32 v49, v18 |
| ; GCN-NEXT: v_mov_b32_e32 v48, v16 |
| ; GCN-NEXT: v_mov_b32_e32 v39, v14 |
| ; GCN-NEXT: v_mov_b32_e32 v38, v12 |
| ; GCN-NEXT: v_mov_b32_e32 v37, v10 |
| ; GCN-NEXT: v_mov_b32_e32 v36, v8 |
| ; GCN-NEXT: v_mov_b32_e32 v35, v6 |
| ; GCN-NEXT: v_mov_b32_e32 v34, v4 |
| ; GCN-NEXT: v_mov_b32_e32 v33, v2 |
| ; GCN-NEXT: v_mov_b32_e32 v32, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v0 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB46_3 |
| ; GCN-NEXT: ; %bb.1: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execnz .LBB46_4 |
| ; GCN-NEXT: .LBB46_2: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; GCN-NEXT: .LBB46_3: ; %cmp.false |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v34 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v36 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v48 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v55 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB46_2 |
| ; GCN-NEXT: .LBB46_4: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v55 |
| ; GCN-NEXT: s_mov_b32 s6, 0x30000 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v54 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v53 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v52 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v51 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v50 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v49 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v48 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v39 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v37 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v36 |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v35 |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v34 |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v33 |
| ; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v32 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xffff, v30 |
| ; GCN-NEXT: v_or_b32_e32 v0, v31, v0 |
| ; GCN-NEXT: v_or_b32_e32 v2, v29, v2 |
| ; GCN-NEXT: v_or_b32_e32 v4, v27, v4 |
| ; GCN-NEXT: v_or_b32_e32 v6, v25, v6 |
| ; GCN-NEXT: v_or_b32_e32 v8, v23, v8 |
| ; GCN-NEXT: v_or_b32_e32 v10, v21, v10 |
| ; GCN-NEXT: v_or_b32_e32 v12, v19, v12 |
| ; GCN-NEXT: v_or_b32_e32 v14, v17, v14 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v18 |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v20 |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v22 |
| ; GCN-NEXT: v_or_b32_e32 v7, v7, v24 |
| ; GCN-NEXT: v_or_b32_e32 v5, v5, v26 |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v28 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v30 |
| ; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x30000, v0 |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, s6, v2 |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v14 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, s6, v15 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v13 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v3 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v1 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v14 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v18 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v26 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v30 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32i16_to_v32bf16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB46_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v16, 3 |
| ; VI-NEXT: v_add_u16_sdwa v19, v15, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v15, 3, v15 |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v14, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v14 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v13, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v13 |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v12, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v12 |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v11, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v11 |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v10, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v10 |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v9, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v9 |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v8, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v8 |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v7, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v6, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v4, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; VI-NEXT: v_add_u16_sdwa v17, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v18, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v19 |
| ; VI-NEXT: v_add_u16_sdwa v19, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v16, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v1 |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v16 |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v19 |
| ; VI-NEXT: v_or_b32_e32 v1, v1, v18 |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v17 |
| ; VI-NEXT: .LBB46_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32i16_to_v32bf16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB46_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB46_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32i16_to_v32bf16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB46_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: .LBB46_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <32 x i16> %a, splat (i16 3) |
| %a2 = bitcast <32 x i16> %a1 to <32 x bfloat> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x i16> %a to <32 x bfloat> |
| br label %end |
| |
| end: |
| %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x bfloat> %phi |
| } |
| |
| define <32 x i16> @bitcast_v32bf16_to_v32i16(<32 x bfloat> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32bf16_to_v32i16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_mul_f32_e32 v63, 1.0, v0 |
| ; GCN-NEXT: v_mul_f32_e32 v62, 1.0, v1 |
| ; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v2 |
| ; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v3 |
| ; GCN-NEXT: v_mul_f32_e32 v61, 1.0, v4 |
| ; GCN-NEXT: v_mul_f32_e32 v60, 1.0, v5 |
| ; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v6 |
| ; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v7 |
| ; GCN-NEXT: v_mul_f32_e32 v59, 1.0, v8 |
| ; GCN-NEXT: v_mul_f32_e32 v58, 1.0, v9 |
| ; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v10 |
| ; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v11 |
| ; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v12 |
| ; GCN-NEXT: v_mul_f32_e32 v56, 1.0, v13 |
| ; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v14 |
| ; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v15 |
| ; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v16 |
| ; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v17 |
| ; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v18 |
| ; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v19 |
| ; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v20 |
| ; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v21 |
| ; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v22 |
| ; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v23 |
| ; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v24 |
| ; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v25 |
| ; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v26 |
| ; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v27 |
| ; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v28 |
| ; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v29 |
| ; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v30 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v55 |
| ; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v31 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB47_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v63 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v62 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v33 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v32 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v61 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v60 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v35 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v34 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v59 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v58 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v37 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v36 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v57 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v56 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v39 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v38 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v47 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v46 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v49 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v48 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v45 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v44 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v51 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v50 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v43 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v42 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v52 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v41 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v40 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v54 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v55 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: .LBB47_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB47_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v63 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v62 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v61 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v60 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v59 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v58 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v57 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v56 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v47 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v46 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v45 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v44 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v43 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v42 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v41 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v40 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v54 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v55 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v53 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v52 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v51 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v50 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v49 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v48 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v39 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v38 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v37 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v36 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v35 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v34 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v33 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v32 |
| ; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v32, 0x40c00000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v33, 0x40c00000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v16 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v34, 0x40c00000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v35, 0x40c00000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v30 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v31 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v1 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v7 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v15 |
| ; GCN-NEXT: v_and_b32_e32 v51, 0xffff0000, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GCN-NEXT: v_and_b32_e32 v52, 0xffff0000, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v19 |
| ; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v20 |
| ; GCN-NEXT: v_and_b32_e32 v54, 0xffff0000, v7 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v22 |
| ; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v24 |
| ; GCN-NEXT: v_and_b32_e32 v40, 0xffff0000, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v26 |
| ; GCN-NEXT: v_and_b32_e32 v41, 0xffff0000, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v28 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GCN-NEXT: v_alignbit_b32 v0, v30, v0, 16 |
| ; GCN-NEXT: v_alignbit_b32 v4, v36, v2, 16 |
| ; GCN-NEXT: v_alignbit_b32 v8, v37, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v12, v38, v5, 16 |
| ; GCN-NEXT: v_alignbit_b32 v16, v39, v33, 16 |
| ; GCN-NEXT: v_alignbit_b32 v20, v48, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v24, v49, v10, 16 |
| ; GCN-NEXT: v_alignbit_b32 v28, v50, v13, 16 |
| ; GCN-NEXT: v_alignbit_b32 v30, v31, v14, 16 |
| ; GCN-NEXT: v_alignbit_b32 v26, v27, v17, 16 |
| ; GCN-NEXT: v_alignbit_b32 v22, v23, v18, 16 |
| ; GCN-NEXT: v_alignbit_b32 v18, v19, v21, 16 |
| ; GCN-NEXT: v_alignbit_b32 v14, v15, v34, 16 |
| ; GCN-NEXT: v_alignbit_b32 v10, v11, v25, 16 |
| ; GCN-NEXT: v_alignbit_b32 v6, v7, v35, 16 |
| ; GCN-NEXT: v_alignbit_b32 v2, v3, v29, 16 |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v6, v41, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v10, v40, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v14, v55, 16 |
| ; GCN-NEXT: v_alignbit_b32 v17, v18, v54, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v22, v53, 16 |
| ; GCN-NEXT: v_alignbit_b32 v25, v26, v52, 16 |
| ; GCN-NEXT: v_alignbit_b32 v29, v30, v51, 16 |
| ; GCN-NEXT: .LBB47_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32bf16_to_v32i16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB47_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17 |
| ; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v0, 16, 1 |
| ; VI-NEXT: s_movk_i32 s6, 0x7fff |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v0 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v0 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0 |
| ; VI-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v1 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v1, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v1 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v1, v18, v19, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v2 |
| ; VI-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; VI-NEXT: v_bfe_u32 v19, v18, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v19, vcc, v19, v18 |
| ; VI-NEXT: v_add_u32_e32 v19, vcc, s6, v19 |
| ; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; VI-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v18, v18 |
| ; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc |
| ; VI-NEXT: v_bfe_u32 v19, v2, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v19, vcc, v19, v2 |
| ; VI-NEXT: v_add_u32_e32 v19, vcc, s6, v19 |
| ; VI-NEXT: v_or_b32_e32 v20, 0x400000, v2 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v2, v19, v20, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v19, 16, v3 |
| ; VI-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 |
| ; VI-NEXT: v_bfe_u32 v20, v19, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v20, vcc, v20, v19 |
| ; VI-NEXT: v_add_u32_e32 v20, vcc, s6, v20 |
| ; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; VI-NEXT: v_or_b32_e32 v21, 0x400000, v19 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v19, v19 |
| ; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v19, v20, v21, vcc |
| ; VI-NEXT: v_bfe_u32 v20, v3, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v20, vcc, v20, v3 |
| ; VI-NEXT: v_add_u32_e32 v20, vcc, s6, v20 |
| ; VI-NEXT: v_or_b32_e32 v21, 0x400000, v3 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v4 |
| ; VI-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 |
| ; VI-NEXT: v_bfe_u32 v21, v20, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v21, vcc, v21, v20 |
| ; VI-NEXT: v_add_u32_e32 v21, vcc, s6, v21 |
| ; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; VI-NEXT: v_or_b32_e32 v22, 0x400000, v20 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v20, v20 |
| ; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v20, v21, v22, vcc |
| ; VI-NEXT: v_bfe_u32 v21, v4, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v21, vcc, v21, v4 |
| ; VI-NEXT: v_add_u32_e32 v21, vcc, s6, v21 |
| ; VI-NEXT: v_or_b32_e32 v22, 0x400000, v4 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v4, v21, v22, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v5 |
| ; VI-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 |
| ; VI-NEXT: v_bfe_u32 v22, v21, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v22, vcc, v22, v21 |
| ; VI-NEXT: v_add_u32_e32 v22, vcc, s6, v22 |
| ; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; VI-NEXT: v_or_b32_e32 v23, 0x400000, v21 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v21, v21 |
| ; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v21, v22, v23, vcc |
| ; VI-NEXT: v_bfe_u32 v22, v5, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v22, vcc, v22, v5 |
| ; VI-NEXT: v_add_u32_e32 v22, vcc, s6, v22 |
| ; VI-NEXT: v_or_b32_e32 v23, 0x400000, v5 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v5, v22, v23, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v6 |
| ; VI-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; VI-NEXT: v_bfe_u32 v23, v22, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v23, vcc, v23, v22 |
| ; VI-NEXT: v_add_u32_e32 v23, vcc, s6, v23 |
| ; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; VI-NEXT: v_or_b32_e32 v24, 0x400000, v22 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v22, v22 |
| ; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v22, v23, v24, vcc |
| ; VI-NEXT: v_bfe_u32 v23, v6, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v23, vcc, v23, v6 |
| ; VI-NEXT: v_add_u32_e32 v23, vcc, s6, v23 |
| ; VI-NEXT: v_or_b32_e32 v24, 0x400000, v6 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v6, v23, v24, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v7 |
| ; VI-NEXT: v_add_f32_e32 v23, 0x40c00000, v23 |
| ; VI-NEXT: v_bfe_u32 v24, v23, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v24, vcc, v24, v23 |
| ; VI-NEXT: v_add_u32_e32 v24, vcc, s6, v24 |
| ; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; VI-NEXT: v_or_b32_e32 v25, 0x400000, v23 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v23, v23 |
| ; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v23, v24, v25, vcc |
| ; VI-NEXT: v_bfe_u32 v24, v7, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v24, vcc, v24, v7 |
| ; VI-NEXT: v_add_u32_e32 v24, vcc, s6, v24 |
| ; VI-NEXT: v_or_b32_e32 v25, 0x400000, v7 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v7, v24, v25, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v8 |
| ; VI-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 |
| ; VI-NEXT: v_bfe_u32 v25, v24, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v25, vcc, v25, v24 |
| ; VI-NEXT: v_add_u32_e32 v25, vcc, s6, v25 |
| ; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; VI-NEXT: v_or_b32_e32 v26, 0x400000, v24 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v24, v24 |
| ; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v24, v25, v26, vcc |
| ; VI-NEXT: v_bfe_u32 v25, v8, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v25, vcc, v25, v8 |
| ; VI-NEXT: v_add_u32_e32 v25, vcc, s6, v25 |
| ; VI-NEXT: v_or_b32_e32 v26, 0x400000, v8 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v8, v25, v26, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v9 |
| ; VI-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 |
| ; VI-NEXT: v_bfe_u32 v26, v25, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v26, vcc, v26, v25 |
| ; VI-NEXT: v_add_u32_e32 v26, vcc, s6, v26 |
| ; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; VI-NEXT: v_or_b32_e32 v27, 0x400000, v25 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v25, v25 |
| ; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v25, v26, v27, vcc |
| ; VI-NEXT: v_bfe_u32 v26, v9, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v26, vcc, v26, v9 |
| ; VI-NEXT: v_add_u32_e32 v26, vcc, s6, v26 |
| ; VI-NEXT: v_or_b32_e32 v27, 0x400000, v9 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v9, v26, v27, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v10 |
| ; VI-NEXT: v_add_f32_e32 v26, 0x40c00000, v26 |
| ; VI-NEXT: v_bfe_u32 v27, v26, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v27, vcc, v27, v26 |
| ; VI-NEXT: v_add_u32_e32 v27, vcc, s6, v27 |
| ; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; VI-NEXT: v_or_b32_e32 v28, 0x400000, v26 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v26, v26 |
| ; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v26, v27, v28, vcc |
| ; VI-NEXT: v_bfe_u32 v27, v10, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v27, vcc, v27, v10 |
| ; VI-NEXT: v_add_u32_e32 v27, vcc, s6, v27 |
| ; VI-NEXT: v_or_b32_e32 v28, 0x400000, v10 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v10, v27, v28, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v11 |
| ; VI-NEXT: v_add_f32_e32 v27, 0x40c00000, v27 |
| ; VI-NEXT: v_bfe_u32 v28, v27, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v28, vcc, v28, v27 |
| ; VI-NEXT: v_add_u32_e32 v28, vcc, s6, v28 |
| ; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; VI-NEXT: v_or_b32_e32 v29, 0x400000, v27 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v27, v27 |
| ; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v27, v28, v29, vcc |
| ; VI-NEXT: v_bfe_u32 v28, v11, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v28, vcc, v28, v11 |
| ; VI-NEXT: v_add_u32_e32 v28, vcc, s6, v28 |
| ; VI-NEXT: v_or_b32_e32 v29, 0x400000, v11 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v11, v28, v29, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v12 |
| ; VI-NEXT: v_add_f32_e32 v28, 0x40c00000, v28 |
| ; VI-NEXT: v_bfe_u32 v29, v28, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v29, vcc, v29, v28 |
| ; VI-NEXT: v_add_u32_e32 v29, vcc, s6, v29 |
| ; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; VI-NEXT: v_or_b32_e32 v30, 0x400000, v28 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v28, v28 |
| ; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v28, v29, v30, vcc |
| ; VI-NEXT: v_bfe_u32 v29, v12, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v29, vcc, v29, v12 |
| ; VI-NEXT: v_add_u32_e32 v29, vcc, s6, v29 |
| ; VI-NEXT: v_or_b32_e32 v30, 0x400000, v12 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v12, v29, v30, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v13 |
| ; VI-NEXT: v_add_f32_e32 v29, 0x40c00000, v29 |
| ; VI-NEXT: v_bfe_u32 v30, v29, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v30, vcc, v30, v29 |
| ; VI-NEXT: v_add_u32_e32 v30, vcc, s6, v30 |
| ; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; VI-NEXT: v_or_b32_e32 v31, 0x400000, v29 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v29, v29 |
| ; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v29, v30, v31, vcc |
| ; VI-NEXT: v_bfe_u32 v30, v13, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v30, vcc, v30, v13 |
| ; VI-NEXT: v_add_u32_e32 v30, vcc, s6, v30 |
| ; VI-NEXT: v_or_b32_e32 v31, 0x400000, v13 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v13, v30, v31, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v14 |
| ; VI-NEXT: v_add_f32_e32 v30, 0x40c00000, v30 |
| ; VI-NEXT: v_bfe_u32 v31, v30, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v31, vcc, v31, v30 |
| ; VI-NEXT: v_add_u32_e32 v31, vcc, s6, v31 |
| ; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; VI-NEXT: v_or_b32_e32 v32, 0x400000, v30 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v30, v30 |
| ; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v30, v31, v32, vcc |
| ; VI-NEXT: v_bfe_u32 v31, v14, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v31, vcc, v31, v14 |
| ; VI-NEXT: v_add_u32_e32 v31, vcc, s6, v31 |
| ; VI-NEXT: v_or_b32_e32 v32, 0x400000, v14 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v14, v31, v32, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v15 |
| ; VI-NEXT: v_add_f32_e32 v31, 0x40c00000, v31 |
| ; VI-NEXT: v_bfe_u32 v32, v31, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v32, vcc, v32, v31 |
| ; VI-NEXT: v_add_u32_e32 v32, vcc, s6, v32 |
| ; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; VI-NEXT: v_or_b32_e32 v33, 0x400000, v31 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v31, v31 |
| ; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc |
| ; VI-NEXT: v_bfe_u32 v32, v15, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v32, vcc, v32, v15 |
| ; VI-NEXT: v_add_u32_e32 v32, vcc, s6, v32 |
| ; VI-NEXT: v_or_b32_e32 v33, 0x400000, v15 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v15, v32, v33, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 |
| ; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; VI-NEXT: v_alignbit_b32 v15, v15, v31, 16 |
| ; VI-NEXT: v_alignbit_b32 v14, v14, v30, 16 |
| ; VI-NEXT: v_alignbit_b32 v13, v13, v29, 16 |
| ; VI-NEXT: v_alignbit_b32 v12, v12, v28, 16 |
| ; VI-NEXT: v_alignbit_b32 v11, v11, v27, 16 |
| ; VI-NEXT: v_alignbit_b32 v10, v10, v26, 16 |
| ; VI-NEXT: v_alignbit_b32 v9, v9, v25, 16 |
| ; VI-NEXT: v_alignbit_b32 v8, v8, v24, 16 |
| ; VI-NEXT: v_alignbit_b32 v7, v7, v23, 16 |
| ; VI-NEXT: v_alignbit_b32 v6, v6, v22, 16 |
| ; VI-NEXT: v_alignbit_b32 v5, v5, v21, 16 |
| ; VI-NEXT: v_alignbit_b32 v4, v4, v20, 16 |
| ; VI-NEXT: v_alignbit_b32 v3, v3, v19, 16 |
| ; VI-NEXT: v_alignbit_b32 v2, v2, v18, 16 |
| ; VI-NEXT: v_alignbit_b32 v1, v1, v17, 16 |
| ; VI-NEXT: v_alignbit_b32 v0, v0, v16, 16 |
| ; VI-NEXT: .LBB47_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32bf16_to_v32i16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB47_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: s_movk_i32 s6, 0x7fff |
| ; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v0, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v0, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v0 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; GFX9-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX9-NEXT: v_add3_u32 v18, v18, v17, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; GFX9-NEXT: v_bfe_u32 v18, v1, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v18, v18, v1, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v19, 0x400000, v1 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v1, v18, v19, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v18, 16, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GFX9-NEXT: v_bfe_u32 v19, v18, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX9-NEXT: v_add3_u32 v19, v19, v18, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc |
| ; GFX9-NEXT: v_bfe_u32 v19, v2, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v19, v19, v2, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v2 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v2, v19, v20, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v19, 16, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 |
| ; GFX9-NEXT: v_bfe_u32 v20, v19, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX9-NEXT: v_add3_u32 v20, v20, v19, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v21, 0x400000, v19 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v19, v20, v21, vcc |
| ; GFX9-NEXT: v_bfe_u32 v20, v3, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v20, v20, v3, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v21, 0x400000, v3 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v20, 16, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 |
| ; GFX9-NEXT: v_bfe_u32 v21, v20, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX9-NEXT: v_add3_u32 v21, v21, v20, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v22, 0x400000, v20 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v20, v21, v22, vcc |
| ; GFX9-NEXT: v_bfe_u32 v21, v4, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v21, v21, v4, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v22, 0x400000, v4 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v4, v21, v22, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v21, 16, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 |
| ; GFX9-NEXT: v_bfe_u32 v22, v21, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX9-NEXT: v_add3_u32 v22, v22, v21, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v23, 0x400000, v21 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v21, v22, v23, vcc |
| ; GFX9-NEXT: v_bfe_u32 v22, v5, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v22, v22, v5, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v23, 0x400000, v5 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v5, v22, v23, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v22, 16, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; GFX9-NEXT: v_bfe_u32 v23, v22, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX9-NEXT: v_add3_u32 v23, v23, v22, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v24, 0x400000, v22 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v22, v23, v24, vcc |
| ; GFX9-NEXT: v_bfe_u32 v23, v6, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v23, v23, v6, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v24, 0x400000, v6 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v6, v23, v24, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v23, 16, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v23, 0x40c00000, v23 |
| ; GFX9-NEXT: v_bfe_u32 v24, v23, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GFX9-NEXT: v_add3_u32 v24, v24, v23, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v25, 0x400000, v23 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v23, v24, v25, vcc |
| ; GFX9-NEXT: v_bfe_u32 v24, v7, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v24, v24, v7, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v25, 0x400000, v7 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v7, v24, v25, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v24, 16, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 |
| ; GFX9-NEXT: v_bfe_u32 v25, v24, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX9-NEXT: v_add3_u32 v25, v25, v24, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v26, 0x400000, v24 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v24, v25, v26, vcc |
| ; GFX9-NEXT: v_bfe_u32 v25, v8, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v25, v25, v8, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v26, 0x400000, v8 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v8, v25, v26, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v25, 16, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 |
| ; GFX9-NEXT: v_bfe_u32 v26, v25, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX9-NEXT: v_add3_u32 v26, v26, v25, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v27, 0x400000, v25 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v25, v26, v27, vcc |
| ; GFX9-NEXT: v_bfe_u32 v26, v9, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v26, v26, v9, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v27, 0x400000, v9 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v9, v26, v27, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v26, 16, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v26, 0x40c00000, v26 |
| ; GFX9-NEXT: v_bfe_u32 v27, v26, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX9-NEXT: v_add3_u32 v27, v27, v26, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v28, 0x400000, v26 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v26, v27, v28, vcc |
| ; GFX9-NEXT: v_bfe_u32 v27, v10, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v27, v27, v10, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v28, 0x400000, v10 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v10, v27, v28, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v27, 16, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v27, 0x40c00000, v27 |
| ; GFX9-NEXT: v_bfe_u32 v28, v27, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; GFX9-NEXT: v_add3_u32 v28, v28, v27, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v29, 0x400000, v27 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v27, v28, v29, vcc |
| ; GFX9-NEXT: v_bfe_u32 v28, v11, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v28, v28, v11, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v29, 0x400000, v11 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v11, v28, v29, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v28, 16, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v28, 0x40c00000, v28 |
| ; GFX9-NEXT: v_bfe_u32 v29, v28, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX9-NEXT: v_add3_u32 v29, v29, v28, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v30, 0x400000, v28 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v28, v29, v30, vcc |
| ; GFX9-NEXT: v_bfe_u32 v29, v12, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v29, v29, v12, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v30, 0x400000, v12 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v12, v29, v30, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v29, 16, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v29, 0x40c00000, v29 |
| ; GFX9-NEXT: v_bfe_u32 v30, v29, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX9-NEXT: v_add3_u32 v30, v30, v29, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v31, 0x400000, v29 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v29, v30, v31, vcc |
| ; GFX9-NEXT: v_bfe_u32 v30, v13, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v30, v30, v13, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v31, 0x400000, v13 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v13, v30, v31, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v30, 16, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v30, 0x40c00000, v30 |
| ; GFX9-NEXT: v_bfe_u32 v31, v30, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX9-NEXT: v_add3_u32 v31, v31, v30, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v30 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v30, v31, v32, vcc |
| ; GFX9-NEXT: v_bfe_u32 v31, v14, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v31, v31, v14, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v14 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v14, v31, v32, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v31, 16, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v31, 0x40c00000, v31 |
| ; GFX9-NEXT: v_bfe_u32 v32, v31, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX9-NEXT: v_add3_u32 v32, v32, v31, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v31 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31 |
| ; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc |
| ; GFX9-NEXT: v_bfe_u32 v32, v15, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v32, v32, v15, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v15 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v15, v32, v33, vcc |
| ; GFX9-NEXT: s_mov_b32 s6, 0x7060302 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v31, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v30, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v29, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v28, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v11, v27, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v10, v26, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v9, v25, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v8, v24, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v7, v23, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v6, v22, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v5, v21, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v4, v20, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v3, v19, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v2, v18, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v1, v17, s6 |
| ; GFX9-NEXT: v_perm_b32 v0, v0, v16, s6 |
| ; GFX9-NEXT: .LBB47_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32bf16_to_v32i16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB47_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v1 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v11 |
| ; GFX11-NEXT: v_dual_add_f32 v24, 0x40c00000, v24 :: v_dual_lshlrev_b32 v25, 16, v6 |
| ; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v19, v16, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v16 |
| ; GFX11-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v17, 0x7fff |
| ; GFX11-NEXT: v_add3_u32 v19, v19, v16, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX11-NEXT: v_dual_add_f32 v26, 0x40c00000, v26 :: v_dual_lshlrev_b32 v27, 16, v8 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v19, v22, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v22, 16, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX11-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX11-NEXT: v_add_f32_e32 v27, 0x40c00000, v27 |
| ; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v10 |
| ; GFX11-NEXT: v_bfe_u32 v20, v0, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v0 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 |
| ; GFX11-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GFX11-NEXT: v_dual_add_f32 v28, 0x40c00000, v28 :: v_dual_add_f32 v29, 0x40c00000, v29 |
| ; GFX11-NEXT: v_add3_u32 v20, v20, v0, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX11-NEXT: v_dual_add_f32 v30, 0x40c00000, v30 :: v_dual_lshlrev_b32 v31, 16, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v0, v20, v23 :: v_dual_lshlrev_b32 v23, 16, v4 |
| ; GFX11-NEXT: v_bfe_u32 v20, v1, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX11-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_add_f32 v23, 0x40c00000, v23 |
| ; GFX11-NEXT: v_perm_b32 v0, v0, v16, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v21, v19, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v19, v20, v1, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX11-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v31, 0x40c00000, v31 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v1, v19, v20 :: v_dual_add_f32 v18, 0x40c00000, v18 |
| ; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX11-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; GFX11-NEXT: v_bfe_u32 v21, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v33, 0x400000, v31 |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v18, 0x7fff |
| ; GFX11-NEXT: v_bfe_u32 v21, v2, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v34, v12, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v1, v1, v17, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v18, v19, v20 :: v_dual_and_b32 v7, 0xffff0000, v7 |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v2, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v2 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 |
| ; GFX11-NEXT: v_bfe_u32 v21, v22, 16, 1 |
| ; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v2, v19, v20 :: v_dual_and_b32 v11, 0xffff0000, v11 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v22, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v22 |
| ; GFX11-NEXT: v_bfe_u32 v21, v3, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 |
| ; GFX11-NEXT: v_bfe_u32 v22, v23, 16, 1 |
| ; GFX11-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GFX11-NEXT: v_perm_b32 v2, v2, v18, 0x7060302 |
| ; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v19, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v20, v21, v3, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v3 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v32, 0x400000, v11 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v20, v22, v23, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v23 |
| ; GFX11-NEXT: v_bfe_u32 v22, v4, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 |
| ; GFX11-NEXT: v_bfe_u32 v23, v24, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v3, v3, v19, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v20, v20, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v21, v22, v4, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v4 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v4, v21, v22, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v21, v23, v24, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v24 |
| ; GFX11-NEXT: v_bfe_u32 v23, v5, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24 |
| ; GFX11-NEXT: v_bfe_u32 v24, v25, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v4, v4, v20, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v21, v21, v22, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v22, v23, v5, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v5 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v5, v22, v23, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v22, v24, v25, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v25 |
| ; GFX11-NEXT: v_bfe_u32 v24, v6, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25 |
| ; GFX11-NEXT: v_bfe_u32 v25, v26, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v5, v5, v21, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v22, v22, v23, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v23, v24, v6, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v6 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v6, v23, v24, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v23, v25, v26, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v26 |
| ; GFX11-NEXT: v_bfe_u32 v25, v7, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26 |
| ; GFX11-NEXT: v_bfe_u32 v26, v27, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v6, v6, v22, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v23, v23, v24, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v24, v25, v7, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v7 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v7, v24, v25, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v24, v26, v27, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v27 |
| ; GFX11-NEXT: v_bfe_u32 v26, v8, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27 |
| ; GFX11-NEXT: v_bfe_u32 v27, v28, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v7, v7, v23, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v24, v24, v25, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v25, v26, v8, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v8 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v8, v25, v26, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v25, v27, v28, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v28 |
| ; GFX11-NEXT: v_bfe_u32 v27, v9, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28 |
| ; GFX11-NEXT: v_bfe_u32 v28, v29, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v8, v8, v24, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v25, v25, v26, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v26, v27, v9, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v27, 0x400000, v9 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v9, v26, v27, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v26, v28, v29, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v27, 0x400000, v29 |
| ; GFX11-NEXT: v_bfe_u32 v28, v10, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29 |
| ; GFX11-NEXT: v_bfe_u32 v29, v30, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v9, v9, v25, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v26, v26, v27, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v27, v28, v10, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v28, 0x400000, v10 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v10, v27, v28, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v27, v29, v30, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v28, 0x400000, v30 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30 |
| ; GFX11-NEXT: v_bfe_u32 v30, v31, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v29, v11, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v27, v27, v28 :: v_dual_lshlrev_b32 v28, 16, v13 |
| ; GFX11-NEXT: v_add3_u32 v30, v30, v31, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX11-NEXT: v_add3_u32 v31, v34, v12, 0x7fff |
| ; GFX11-NEXT: v_add_f32_e32 v28, 0x40c00000, v28 |
| ; GFX11-NEXT: v_add3_u32 v29, v29, v11, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v30, v30, v33, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v33, 0x400000, v12 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 |
| ; GFX11-NEXT: v_bfe_u32 v35, v28, 16, 1 |
| ; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v28 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v12, v31, v33, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v34, v35, v28, 0x7fff |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v35, 16, v14 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX11-NEXT: v_bfe_u32 v37, v13, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v10, v10, v26, 0x7060302 |
| ; GFX11-NEXT: v_dual_add_f32 v31, 0x40c00000, v35 :: v_dual_cndmask_b32 v28, v34, v36 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v15 |
| ; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX11-NEXT: v_add3_u32 v33, v37, v13, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v35, v31, 16, 1 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34 |
| ; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v31 |
| ; GFX11-NEXT: v_bfe_u32 v38, v14, 16, 1 |
| ; GFX11-NEXT: v_add3_u32 v35, v35, v31, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 |
| ; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX11-NEXT: v_bfe_u32 v39, v34, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v34 |
| ; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v13 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v31, v35, v37, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v37, v38, v14, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v14 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14 |
| ; GFX11-NEXT: v_bfe_u32 v35, v15, 16, 1 |
| ; GFX11-NEXT: v_add3_u32 v39, v39, v34, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v15 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v14, v37, v38, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34 |
| ; GFX11-NEXT: v_add3_u32 v35, v35, v15, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v14, v14, v31, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v34, v39, v48, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v15, v35, v49, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_perm_b32 v15, v15, v34, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v13, v33, v36, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 |
| ; GFX11-NEXT: v_perm_b32 v12, v12, v30, 0x7060302 |
| ; GFX11-NEXT: v_perm_b32 v13, v13, v28, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v11, v29, v32, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) |
| ; GFX11-NEXT: v_perm_b32 v11, v11, v27, 0x7060302 |
| ; GFX11-NEXT: .LBB47_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x bfloat> %a, splat (bfloat 0xR40C0) |
| %a2 = bitcast <32 x bfloat> %a1 to <32 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x bfloat> %a to <32 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x i16> %phi |
| } |
| |
| define <64 x i8> @bitcast_v32i16_to_v64i8(<32 x i16> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32i16_to_v64i8: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v4 |
| ; GCN-NEXT: s_waitcnt expcnt(6) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v6 |
| ; GCN-NEXT: s_waitcnt expcnt(4) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v59, 16, v8 |
| ; GCN-NEXT: s_waitcnt expcnt(3) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v60, 16, v10 |
| ; GCN-NEXT: s_waitcnt expcnt(2) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v61, 16, v12 |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v62, 16, v14 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v63, 16, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v18 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v49 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; kill: killed $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; kill: killed $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; kill: killed $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB48_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v3 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v21 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v27 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v29 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v38 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v4 |
| ; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v8 |
| ; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v12 |
| ; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v16 |
| ; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_and_b32_e32 v40, 0xffff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v53, 0xffff, v24 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v28 |
| ; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v49 |
| ; GCN-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_bfe_u32 v4, v4, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v4, v8, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_bfe_u32 v44, v12, 8, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v4, v16, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v4, v20, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v4, v24, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v4, v28, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v4, v49, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_or_b32_e32 v52, v1, v46 |
| ; GCN-NEXT: v_or_b32_e32 v51, v2, v47 |
| ; GCN-NEXT: v_or_b32_e32 v48, v3, v57 |
| ; GCN-NEXT: v_or_b32_e32 v37, v5, v59 |
| ; GCN-NEXT: v_or_b32_e32 v32, v6, v60 |
| ; GCN-NEXT: v_or_b32_e32 v6, v7, v61 |
| ; GCN-NEXT: v_or_b32_e32 v58, v9, v62 |
| ; GCN-NEXT: v_or_b32_e32 v56, v11, v63 |
| ; GCN-NEXT: v_or_b32_e32 v42, v13, v10 |
| ; GCN-NEXT: v_or_b32_e32 v54, v15, v14 |
| ; GCN-NEXT: v_or_b32_e32 v50, v17, v22 |
| ; GCN-NEXT: v_or_b32_e32 v39, v18, v33 |
| ; GCN-NEXT: v_or_b32_e32 v36, v19, v26 |
| ; GCN-NEXT: v_or_b32_e32 v31, v21, v34 |
| ; GCN-NEXT: v_or_b32_e32 v18, v23, v30 |
| ; GCN-NEXT: v_or_b32_e32 v2, v25, v35 |
| ; GCN-NEXT: v_alignbit_b32 v1, v51, v52, 24 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v51, v52, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v51, v52, 8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v37, v48, 24 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v37, v48, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v37, v48, 8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v55, v6, v32, 24 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v6, v32, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v6, v32, 8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v41, v56, v58, 24 |
| ; GCN-NEXT: v_alignbit_b32 v43, v56, v58, 16 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v56, v58, 8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v54, v42, 24 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v54, v42, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v54, v42, 8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v39, v50, 24 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v39, v50, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v45, v39, v50, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v31, v36, 24 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v31, v36, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v31, v36, 8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v18, 24 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v18, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v18, 8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v51 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v37 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v6 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v56 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v54 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v39 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v31 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v2 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: .LBB48_2: ; %Flow |
| ; GCN-NEXT: s_or_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: v_mov_b32_e32 v49, v40 |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_xor_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB48_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v29 |
| ; GCN-NEXT: s_mov_b32 s6, 0x30000 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v38 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v25 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v27 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v21 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v23 |
| ; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v17 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v19 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v13 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v15 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v9 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v11 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v5 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v7 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v1 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v3 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: v_or_b32_e32 v2, v30, v2 |
| ; GCN-NEXT: v_or_b32_e32 v4, v35, v4 |
| ; GCN-NEXT: v_or_b32_e32 v6, v26, v6 |
| ; GCN-NEXT: v_or_b32_e32 v8, v34, v8 |
| ; GCN-NEXT: v_or_b32_e32 v12, v22, v12 |
| ; GCN-NEXT: v_or_b32_e32 v16, v33, v16 |
| ; GCN-NEXT: v_or_b32_e32 v10, v10, v17 |
| ; GCN-NEXT: v_or_b32_e32 v14, v14, v18 |
| ; GCN-NEXT: v_or_b32_e32 v13, v62, v13 |
| ; GCN-NEXT: v_or_b32_e32 v15, v63, v15 |
| ; GCN-NEXT: v_or_b32_e32 v9, v60, v9 |
| ; GCN-NEXT: v_or_b32_e32 v11, v61, v11 |
| ; GCN-NEXT: v_or_b32_e32 v5, v57, v5 |
| ; GCN-NEXT: v_or_b32_e32 v7, v59, v7 |
| ; GCN-NEXT: v_or_b32_e32 v1, v46, v1 |
| ; GCN-NEXT: v_or_b32_e32 v3, v47, v3 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 0x30000, v2 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v36, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v31, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v50, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v39, vcc, s6, v16 |
| ; GCN-NEXT: v_add_i32_e32 v42, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v54, vcc, s6, v14 |
| ; GCN-NEXT: v_add_i32_e32 v58, vcc, s6, v13 |
| ; GCN-NEXT: v_add_i32_e32 v56, vcc, s6, v15 |
| ; GCN-NEXT: v_add_i32_e32 v32, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v48, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v37, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v52, vcc, s6, v1 |
| ; GCN-NEXT: v_add_i32_e32 v51, vcc, s6, v3 |
| ; GCN-NEXT: v_alignbit_b32 v1, v51, v52, 24 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v51, v52, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v51, v52, 8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v4, v37, v48, 24 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v37, v48, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v8, v37, v48, 8 |
| ; GCN-NEXT: v_alignbit_b32 v55, v6, v32, 24 |
| ; GCN-NEXT: v_alignbit_b32 v20, v6, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v12, v6, v32, 8 |
| ; GCN-NEXT: v_alignbit_b32 v41, v56, v58, 24 |
| ; GCN-NEXT: v_alignbit_b32 v43, v56, v58, 16 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v56, v58, 8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v54, v42, 24 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v54, v42, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v54, v42, 8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v39, v50, 24 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v39, v50, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v45, v39, v50, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v31, v36, 24 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v31, v36, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v31, v36, 8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v18, 24 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v18, 16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v1, v2, v18, 8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v51 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v51 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v51 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v24, 24, v37 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v37 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v37 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v40, 16, v6 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v6 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v56 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v56 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v56 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v54 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v54 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v54 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v39 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v39 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v39 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v31 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v31 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v31 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 24, v2 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v2 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 8, v2 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: .LBB48_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v52 |
| ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3 |
| ; GCN-NEXT: v_or_b32_e32 v5, v1, v3 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v51 |
| ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3 |
| ; GCN-NEXT: v_or_b32_e32 v7, v1, v3 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v48 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v8 |
| ; GCN-NEXT: v_or_b32_e32 v8, v1, v3 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v37 |
| ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3 |
| ; GCN-NEXT: v_or_b32_e32 v10, v1, v3 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v12 |
| ; GCN-NEXT: v_or_b32_e32 v12, v1, v3 |
| ; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v1 |
| ; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v1 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v28 |
| ; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 24, v1 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v6 |
| ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v3 |
| ; GCN-NEXT: v_or_b32_e32 v13, v1, v3 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 4, v0 |
| ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 24, v4 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v58 |
| ; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 8, v4 |
| ; GCN-NEXT: v_or_b32_e32 v14, v3, v4 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 8, v0 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v35, 24, v24 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v56 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 8, v6 |
| ; GCN-NEXT: v_or_b32_e32 v16, v4, v6 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 12, v0 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xff, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 24, v55 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v42 |
| ; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v9 |
| ; GCN-NEXT: v_or_b32_e32 v17, v6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 16, v0 |
| ; GCN-NEXT: v_and_b32_e32 v48, 0xff, v40 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 24, v44 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v54 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v11 |
| ; GCN-NEXT: v_or_b32_e32 v21, v9, v11 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 20, v0 |
| ; GCN-NEXT: v_and_b32_e32 v51, 0xff, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 24, v41 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 8, v45 |
| ; GCN-NEXT: v_or_b32_e32 v24, v11, v15 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 24, v0 |
| ; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v55, 0xff, v15 |
| ; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v15 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v39 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 8, v19 |
| ; GCN-NEXT: v_or_b32_e32 v22, v15, v19 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 28, v0 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v42, 0xff, v19 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 24, v19 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v36 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v25 |
| ; GCN-NEXT: v_or_b32_e32 v36, v19, v25 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 32, v0 |
| ; GCN-NEXT: v_and_b32_e32 v45, 0xff, v49 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v25 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v31 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v31, 8, v31 |
| ; GCN-NEXT: v_or_b32_e32 v31, v25, v31 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, 36, v0 |
| ; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v46, 0xff, v50 |
| ; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 24, v50 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18 |
| ; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 8, v50 |
| ; GCN-NEXT: v_or_b32_e32 v50, v18, v50 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 40, v0 |
| ; GCN-NEXT: v_and_b32_e32 v47, 0xff, v53 |
| ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 24, v53 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v54, 8, v54 |
| ; GCN-NEXT: v_or_b32_e32 v54, v2, v54 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 44, v0 |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v56, 0xff, v40 |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v40, 24, v40 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; GCN-NEXT: v_or_b32_e32 v33, v33, v32 |
| ; GCN-NEXT: v_add_i32_e32 v32, vcc, 48, v0 |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v57, 0xff, v41 |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v41, 24, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; GCN-NEXT: v_or_b32_e32 v43, v27, v26 |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, 52, v0 |
| ; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v58, 0xff, v27 |
| ; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v44, 24, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v28 |
| ; GCN-NEXT: v_or_b32_e32 v29, v29, v27 |
| ; GCN-NEXT: v_add_i32_e32 v27, vcc, 56, v0 |
| ; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v59, 0xff, v28 |
| ; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v60, 24, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v34 |
| ; GCN-NEXT: v_or_b32_e32 v34, v35, v28 |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 60, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v48 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v42 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v46 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v58 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v59 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GCN-NEXT: v_or_b32_e32 v20, v20, v35 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_or_b32_e32 v23, v23, v38 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GCN-NEXT: v_or_b32_e32 v30, v30, v48 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GCN-NEXT: v_or_b32_e32 v35, v37, v51 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: v_or_b32_e32 v37, v39, v55 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21 |
| ; GCN-NEXT: v_or_b32_e32 v38, v49, v42 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24 |
| ; GCN-NEXT: v_or_b32_e32 v39, v52, v45 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22 |
| ; GCN-NEXT: v_or_b32_e32 v48, v53, v46 |
| ; GCN-NEXT: v_and_b32_e32 v36, 0xffff, v36 |
| ; GCN-NEXT: v_or_b32_e32 v49, v40, v47 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff, v31 |
| ; GCN-NEXT: v_or_b32_e32 v51, v41, v56 |
| ; GCN-NEXT: v_and_b32_e32 v50, 0xffff, v50 |
| ; GCN-NEXT: v_or_b32_e32 v52, v44, v57 |
| ; GCN-NEXT: v_and_b32_e32 v53, 0xffff, v54 |
| ; GCN-NEXT: v_or_b32_e32 v54, v60, v58 |
| ; GCN-NEXT: v_or_b32_e32 v5, v5, v33 |
| ; GCN-NEXT: v_or_b32_e32 v7, v7, v43 |
| ; GCN-NEXT: v_or_b32_e32 v8, v8, v29 |
| ; GCN-NEXT: v_or_b32_e32 v10, v10, v34 |
| ; GCN-NEXT: v_or_b32_e32 v12, v12, v20 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; GCN-NEXT: v_or_b32_e32 v14, v14, v30 |
| ; GCN-NEXT: v_or_b32_e32 v16, v16, v35 |
| ; GCN-NEXT: v_or_b32_e32 v17, v17, v37 |
| ; GCN-NEXT: v_or_b32_e32 v20, v21, v38 |
| ; GCN-NEXT: v_or_b32_e32 v21, v24, v39 |
| ; GCN-NEXT: v_or_b32_e32 v22, v22, v48 |
| ; GCN-NEXT: v_or_b32_e32 v23, v36, v49 |
| ; GCN-NEXT: v_or_b32_e32 v24, v31, v51 |
| ; GCN-NEXT: v_or_b32_e32 v29, v50, v52 |
| ; GCN-NEXT: v_or_b32_e32 v30, v53, v54 |
| ; GCN-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v7, v1, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v8, v3, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v10, v4, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v12, v6, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v13, v9, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v14, v11, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v16, v15, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v17, v19, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v20, v25, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v21, v18, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v22, v2, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v23, v32, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v24, v26, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v29, v27, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v30, v28, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32i16_to_v64i8: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v1 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; kill: killed $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr23 |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB48_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v15 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v14 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v14 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v13 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v12 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v12 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v11 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v10 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v10 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 24, v8 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v8 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v7 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v6 |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[15:16] |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v5 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[13:14] |
| ; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; VI-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v16 |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 24, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v46, 24, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v59, 24, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 8, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v18, 8, v1 |
| ; VI-NEXT: v_mov_b32_e32 v25, v50 |
| ; VI-NEXT: v_mov_b32_e32 v41, v1 |
| ; VI-NEXT: v_mov_b32_e32 v54, v2 |
| ; VI-NEXT: v_mov_b32_e32 v57, v3 |
| ; VI-NEXT: v_mov_b32_e32 v47, v4 |
| ; VI-NEXT: v_mov_b32_e32 v61, v5 |
| ; VI-NEXT: v_mov_b32_e32 v60, v6 |
| ; VI-NEXT: v_mov_b32_e32 v52, v7 |
| ; VI-NEXT: v_mov_b32_e32 v63, v8 |
| ; VI-NEXT: v_mov_b32_e32 v40, v9 |
| ; VI-NEXT: v_mov_b32_e32 v53, v10 |
| ; VI-NEXT: v_mov_b32_e32 v17, v11 |
| ; VI-NEXT: v_mov_b32_e32 v44, v12 |
| ; VI-NEXT: v_mov_b32_e32 v58, v13 |
| ; VI-NEXT: v_mov_b32_e32 v56, v14 |
| ; VI-NEXT: v_mov_b32_e32 v50, v15 |
| ; VI-NEXT: v_mov_b32_e32 v62, v16 |
| ; VI-NEXT: ; implicit-def: $vgpr1 |
| ; VI-NEXT: ; implicit-def: $vgpr3 |
| ; VI-NEXT: ; implicit-def: $vgpr5 |
| ; VI-NEXT: ; implicit-def: $vgpr7 |
| ; VI-NEXT: ; implicit-def: $vgpr9 |
| ; VI-NEXT: ; implicit-def: $vgpr11 |
| ; VI-NEXT: ; implicit-def: $vgpr13 |
| ; VI-NEXT: ; implicit-def: $vgpr15 |
| ; VI-NEXT: .LBB48_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB48_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v18, 3 |
| ; VI-NEXT: v_add_u16_sdwa v26, v16, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v29, v15, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v62, 3, v16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v26 |
| ; VI-NEXT: v_add_u16_e32 v50, 3, v15 |
| ; VI-NEXT: v_lshlrev_b32_e32 v15, 16, v29 |
| ; VI-NEXT: v_or_b32_e32 v16, v62, v16 |
| ; VI-NEXT: v_or_b32_e32 v15, v50, v15 |
| ; VI-NEXT: v_add_u16_sdwa v38, v2, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v49, v1, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v36, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v48, v3, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v34, v6, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v39, v5, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v32, v8, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v37, v7, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v30, v10, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v35, v9, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v28, v12, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v33, v11, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v27, v14, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_sdwa v31, v13, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[15:16] |
| ; VI-NEXT: v_add_u16_e32 v56, 3, v14 |
| ; VI-NEXT: v_lshlrev_b32_e32 v14, 16, v27 |
| ; VI-NEXT: v_add_u16_e32 v58, 3, v13 |
| ; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v31 |
| ; VI-NEXT: v_or_b32_e32 v14, v56, v14 |
| ; VI-NEXT: v_or_b32_e32 v13, v58, v13 |
| ; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; VI-NEXT: v_add_u16_e32 v44, 3, v12 |
| ; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v28 |
| ; VI-NEXT: v_add_u16_e32 v17, 3, v11 |
| ; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v33 |
| ; VI-NEXT: v_add_u16_e32 v53, 3, v10 |
| ; VI-NEXT: v_lshlrev_b32_e32 v10, 16, v30 |
| ; VI-NEXT: v_add_u16_e32 v40, 3, v9 |
| ; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v35 |
| ; VI-NEXT: v_or_b32_e32 v12, v44, v12 |
| ; VI-NEXT: v_or_b32_e32 v11, v17, v11 |
| ; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; VI-NEXT: v_add_u16_e32 v63, 3, v8 |
| ; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v32 |
| ; VI-NEXT: v_add_u16_e32 v52, 3, v7 |
| ; VI-NEXT: v_lshlrev_b32_e32 v7, 16, v37 |
| ; VI-NEXT: v_or_b32_e32 v10, v53, v10 |
| ; VI-NEXT: v_or_b32_e32 v9, v40, v9 |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_add_u16_e32 v60, 3, v6 |
| ; VI-NEXT: v_lshlrev_b32_e32 v6, 16, v34 |
| ; VI-NEXT: v_add_u16_e32 v61, 3, v5 |
| ; VI-NEXT: v_lshlrev_b32_e32 v5, 16, v39 |
| ; VI-NEXT: v_or_b32_e32 v8, v63, v8 |
| ; VI-NEXT: v_or_b32_e32 v7, v52, v7 |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_add_u16_e32 v47, 3, v4 |
| ; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v36 |
| ; VI-NEXT: v_add_u16_e32 v57, 3, v3 |
| ; VI-NEXT: v_lshlrev_b32_e32 v3, 16, v48 |
| ; VI-NEXT: v_or_b32_e32 v6, v60, v6 |
| ; VI-NEXT: v_or_b32_e32 v5, v61, v5 |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; VI-NEXT: v_add_u16_e32 v54, 3, v2 |
| ; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v38 |
| ; VI-NEXT: v_add_u16_e32 v41, 3, v1 |
| ; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v49 |
| ; VI-NEXT: v_or_b32_e32 v4, v47, v4 |
| ; VI-NEXT: v_or_b32_e32 v3, v57, v3 |
| ; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; VI-NEXT: v_or_b32_e32 v2, v54, v2 |
| ; VI-NEXT: v_or_b32_e32 v1, v41, v1 |
| ; VI-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; VI-NEXT: v_lshrrev_b32_e32 v18, 8, v1 |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; VI-NEXT: v_bfe_u32 v1, v27, 8, 8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v8 |
| ; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill |
| ; VI-NEXT: v_bfe_u32 v1, v28, 8, 8 |
| ; VI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v12 |
| ; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v9, 8, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v7, 8, v6 |
| ; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill |
| ; VI-NEXT: v_bfe_u32 v1, v30, 8, 8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v16, 8, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v15, 8, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v14, 8, v14 |
| ; VI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v13, 8, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v11, 8, v10 |
| ; VI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v7, 8, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 8, v2 |
| ; VI-NEXT: v_bfe_u32 v25, v26, 8, 8 |
| ; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill |
| ; VI-NEXT: v_bfe_u32 v1, v32, 8, 8 |
| ; VI-NEXT: v_bfe_u32 v43, v34, 8, 8 |
| ; VI-NEXT: v_bfe_u32 v46, v36, 8, 8 |
| ; VI-NEXT: v_bfe_u32 v59, v38, 8, 8 |
| ; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill |
| ; VI-NEXT: .LBB48_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v18 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v24 |
| ; VI-NEXT: v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v49, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v45 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v59 |
| ; VI-NEXT: v_or_b32_sdwa v1, v54, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v51 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v23 |
| ; VI-NEXT: v_or_b32_sdwa v1, v57, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v55 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v46 |
| ; VI-NEXT: v_or_b32_sdwa v1, v47, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v36, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v22 |
| ; VI-NEXT: v_or_b32_sdwa v2, v39, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v43 |
| ; VI-NEXT: v_or_b32_sdwa v2, v34, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v21 |
| ; VI-NEXT: v_or_b32_sdwa v2, v37, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v52, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2 |
| ; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v32, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v42 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; VI-NEXT: v_or_b32_sdwa v1, v40, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2 |
| ; VI-NEXT: v_or_b32_sdwa v1, v53, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v17, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2 |
| ; VI-NEXT: v_or_b32_sdwa v1, v44, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v28, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2 |
| ; VI-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v31, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2 |
| ; VI-NEXT: v_or_b32_sdwa v1, v56, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v27, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v2 |
| ; VI-NEXT: v_or_b32_sdwa v1, v50, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v29, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; VI-NEXT: v_or_b32_sdwa v2, v26, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 60, v0 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v62, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32i16_to_v64i8: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB48_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v16 |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_mov_b32_e32 v26, v23 |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: .LBB48_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB48_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: .LBB48_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v29 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v23 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v23, v27, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28 |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v22 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v60 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v59 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v57 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v56 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v46 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v21 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v55 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v50 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v49 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56 |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v26 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60 |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32i16_to_v64i8: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB48_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB48_2: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB48_4 |
| ; GFX11-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB48_4: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v96, 0xff, v96 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v86, 0xff, v86 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v87 |
| ; GFX11-NEXT: v_or_b32_e32 v24, v96, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v25 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v86, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84 |
| ; GFX11-NEXT: v_and_b32_e32 v83, 0xff, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v82 |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v80, 8, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v71, 8, v71 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v24 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v81 |
| ; GFX11-NEXT: v_and_b32_e32 v70, 0xff, v70 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v83, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v25, v80 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v71 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v70, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v22 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v66 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v65 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v55 |
| ; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v23, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v25, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v52 |
| ; GFX11-NEXT: v_lshlrev_b16 v20, 8, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v50 |
| ; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v48, 8, v48 |
| ; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v39 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v23, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v25, v49 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v48 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v39, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v38 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v37 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v35 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v34 |
| ; GFX11-NEXT: v_lshlrev_b16 v18, 8, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_lshlrev_b16 v69, 8, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v68, 0xff, v68 |
| ; GFX11-NEXT: v_lshlrev_b16 v67, 8, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v23, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v32 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v30 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v29 |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v26, 8, v26 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v69 |
| ; GFX11-NEXT: v_or_b32_e32 v67, v68, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v23, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v25, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v67 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; GFX11-NEXT: s_clause 0x3 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off |
| ; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <32 x i16> %a, splat (i16 3) |
| %a2 = bitcast <32 x i16> %a1 to <64 x i8> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x i16> %a to <64 x i8> |
| br label %end |
| |
| end: |
| %phi = phi <64 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <64 x i8> %phi |
| } |
| |
| define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v64i8_to_v32i16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:128 |
| ; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:96 |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:92 |
| ; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:88 |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:84 |
| ; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:80 |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:72 |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68 |
| ; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:64 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:60 |
| ; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:56 |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:52 |
| ; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:48 |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:44 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:40 |
| ; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:132 |
| ; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:120 |
| ; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 |
| ; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:112 |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:108 |
| ; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:104 |
| ; GCN-NEXT: s_waitcnt vmcnt(5) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v39 |
| ; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:32 |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:28 |
| ; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:24 |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:20 |
| ; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:16 |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:12 |
| ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v5 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v7 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v13 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v15 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v11 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v21 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v23 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v19 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 8, v29 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v27 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v9 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v17 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v25 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v38 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt vmcnt(14) expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v51 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v39 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v52 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v36 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v35 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v37 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v43, 8, v33 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v16 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v34 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v57, 8, v48 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v8 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v49 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v53 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:36 |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:100 |
| ; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:124 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v32 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v50 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB49_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v4 |
| ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v3 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v12 |
| ; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v20 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v28 |
| ; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v5, v5, v7 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xff, v56 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v46 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v47 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v62 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v0 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xff, v61 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v41 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xff, v63 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xff, v40 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v58 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xff, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v0 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v59 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v60 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v42 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v18 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v20 |
| ; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v17, v21, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v23 |
| ; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v20, v25, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v26 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v27 |
| ; GCN-NEXT: v_or_b32_e32 v23, v28, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v30 |
| ; GCN-NEXT: v_or_b32_e32 v24, v24, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v32 |
| ; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v29, v33, v29 |
| ; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v30, v34, v30 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v31, v35, v31 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v12, v32 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v32 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v14, v14, v32 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v44 |
| ; GCN-NEXT: v_or_b32_e32 v16, v16, v45 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v10, v32, v10 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v40, v32, v11 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v8, v11, v8 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v9, v11, v9 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v6, v11, v6 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v41, v11, v7 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v0, v7, v0 |
| ; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v2, v7, v2 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v17 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v17, v11, v18 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v18, v11, v19 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v20 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v20, v19, v21 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v19, v22 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v23 |
| ; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v22, v22, v25 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v25, v23, v26 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v24 |
| ; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v24, v24, v27 |
| ; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v26, v26, v28 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v29 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v31 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GCN-NEXT: v_or_b32_e32 v33, v1, v10 |
| ; GCN-NEXT: v_or_b32_e32 v35, v3, v8 |
| ; GCN-NEXT: v_or_b32_e32 v37, v4, v6 |
| ; GCN-NEXT: v_or_b32_e32 v39, v5, v0 |
| ; GCN-NEXT: v_or_b32_e32 v49, v7, v17 |
| ; GCN-NEXT: v_or_b32_e32 v51, v11, v20 |
| ; GCN-NEXT: v_or_b32_e32 v53, v19, v22 |
| ; GCN-NEXT: v_or_b32_e32 v55, v23, v24 |
| ; GCN-NEXT: v_or_b32_e32 v32, v27, v40 |
| ; GCN-NEXT: v_or_b32_e32 v34, v28, v9 |
| ; GCN-NEXT: v_or_b32_e32 v36, v29, v41 |
| ; GCN-NEXT: v_or_b32_e32 v38, v12, v2 |
| ; GCN-NEXT: v_or_b32_e32 v48, v13, v18 |
| ; GCN-NEXT: v_or_b32_e32 v50, v14, v21 |
| ; GCN-NEXT: v_or_b32_e32 v52, v15, v25 |
| ; GCN-NEXT: v_or_b32_e32 v54, v16, v26 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v0 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v17 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v20 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v22 |
| ; GCN-NEXT: v_alignbit_b32 v1, v33, v40, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v35, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v37, v41, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v39, v2, 16 |
| ; GCN-NEXT: v_alignbit_b32 v17, v49, v18, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v51, v21, 16 |
| ; GCN-NEXT: v_alignbit_b32 v25, v53, v25, 16 |
| ; GCN-NEXT: v_alignbit_b32 v29, v55, v26, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v24 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; kill: killed $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; kill: killed $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; kill: killed $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; kill: killed $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; kill: killed $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: .LBB49_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB49_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: s_waitcnt vmcnt(2) expcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v8 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GCN-NEXT: v_or_b32_e32 v1, v45, v1 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v24 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GCN-NEXT: v_or_b32_e32 v3, v57, v3 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v42 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GCN-NEXT: v_or_b32_e32 v5, v44, v5 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v41 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GCN-NEXT: v_or_b32_e32 v7, v43, v7 |
| ; GCN-NEXT: s_movk_i32 s7, 0x300 |
| ; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v58 |
| ; GCN-NEXT: s_mov_b32 s6, 0x3000000 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v16 |
| ; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v40 |
| ; GCN-NEXT: v_add_i32_e32 v35, vcc, 3, v63 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v60 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v61 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v62 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v0 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v59 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v47 |
| ; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v56 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v46 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v0 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v26 |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v28 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v30 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v0 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v0 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v0 |
| ; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v2 |
| ; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v4 |
| ; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v6 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xff, v32 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v33 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v34 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v35 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xff, v29 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v6 |
| ; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v4, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v9 |
| ; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v9, v9, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v13 |
| ; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v16 |
| ; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v19 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v19, v19, v21 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v23 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v23, v23, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v25 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v25, v25, v26 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v20, v26, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v26, v26, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v27, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v27, v27, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v29 |
| ; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v29, v29, v30 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v31 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 0x300, v1 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v31, v31, v33 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s7, v3 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v8, v32, v8 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v2, v32, v2 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v7 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v0, v32, v0 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, s7, v4 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v6, v32, v6 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s7, v9 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v11, v32, v11 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v15, v32, v15 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, s7, v16 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v17, v32, v17 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, s7, v19 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v32, v21 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, s7, v23 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v24, v32, v24 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v18, v32, v18 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, s7, v20 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v22, v32, v22 |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, s7, v26 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v10, v32, v10 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, s7, v12 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v14, v32, v14 |
| ; GCN-NEXT: v_add_i32_e32 v27, vcc, s7, v27 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v28, v32, v28 |
| ; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v29 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v30, v32, v30 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff, v27 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_or_b32_e32 v1, v31, v1 |
| ; GCN-NEXT: v_or_b32_e32 v3, v8, v3 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v5 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v7 |
| ; GCN-NEXT: v_or_b32_e32 v4, v6, v4 |
| ; GCN-NEXT: v_or_b32_e32 v5, v11, v9 |
| ; GCN-NEXT: v_or_b32_e32 v6, v15, v13 |
| ; GCN-NEXT: v_or_b32_e32 v7, v17, v16 |
| ; GCN-NEXT: v_or_b32_e32 v8, v21, v19 |
| ; GCN-NEXT: v_or_b32_e32 v9, v24, v23 |
| ; GCN-NEXT: v_or_b32_e32 v11, v18, v25 |
| ; GCN-NEXT: v_or_b32_e32 v13, v22, v20 |
| ; GCN-NEXT: v_or_b32_e32 v10, v10, v26 |
| ; GCN-NEXT: v_or_b32_e32 v12, v14, v12 |
| ; GCN-NEXT: v_or_b32_e32 v14, v28, v27 |
| ; GCN-NEXT: v_or_b32_e32 v15, v30, v29 |
| ; GCN-NEXT: v_add_i32_e32 v54, vcc, s6, v1 |
| ; GCN-NEXT: v_add_i32_e32 v55, vcc, s6, v3 |
| ; GCN-NEXT: v_add_i32_e32 v52, vcc, s6, v2 |
| ; GCN-NEXT: v_add_i32_e32 v53, vcc, s6, v0 |
| ; GCN-NEXT: v_add_i32_e32 v50, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v51, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v48, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v49, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v38, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v39, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v36, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v37, vcc, s6, v13 |
| ; GCN-NEXT: v_add_i32_e32 v34, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v35, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v32, vcc, s6, v14 |
| ; GCN-NEXT: v_add_i32_e32 v33, vcc, s6, v15 |
| ; GCN-NEXT: v_alignbit_b32 v1, v33, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v35, v34, 16 |
| ; GCN-NEXT: v_alignbit_b32 v9, v37, v36, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v39, v38, 16 |
| ; GCN-NEXT: v_alignbit_b32 v17, v49, v48, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v51, v50, 16 |
| ; GCN-NEXT: v_alignbit_b32 v25, v53, v52, 16 |
| ; GCN-NEXT: v_alignbit_b32 v29, v55, v54, 16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v33 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v35 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v37 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v39 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v49 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v51 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v55 |
| ; GCN-NEXT: .LBB49_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_mov_b32_e32 v0, v32 |
| ; GCN-NEXT: v_mov_b32_e32 v2, v33 |
| ; GCN-NEXT: v_mov_b32_e32 v4, v34 |
| ; GCN-NEXT: v_mov_b32_e32 v6, v35 |
| ; GCN-NEXT: s_waitcnt vmcnt(2) |
| ; GCN-NEXT: v_mov_b32_e32 v8, v36 |
| ; GCN-NEXT: v_mov_b32_e32 v10, v37 |
| ; GCN-NEXT: v_mov_b32_e32 v12, v38 |
| ; GCN-NEXT: v_mov_b32_e32 v14, v39 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_mov_b32_e32 v16, v48 |
| ; GCN-NEXT: v_mov_b32_e32 v18, v49 |
| ; GCN-NEXT: v_mov_b32_e32 v20, v50 |
| ; GCN-NEXT: v_mov_b32_e32 v22, v51 |
| ; GCN-NEXT: v_mov_b32_e32 v24, v52 |
| ; GCN-NEXT: v_mov_b32_e32 v26, v53 |
| ; GCN-NEXT: v_mov_b32_e32 v28, v54 |
| ; GCN-NEXT: v_mov_b32_e32 v30, v55 |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v64i8_to_v32i16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:132 |
| ; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64 |
| ; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72 |
| ; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80 |
| ; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88 |
| ; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:96 |
| ; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:104 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:112 |
| ; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120 |
| ; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:128 |
| ; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:20 |
| ; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:12 |
| ; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:100 |
| ; VI-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:68 |
| ; VI-NEXT: buffer_load_ushort v55, off, s[0:3], s32 offset:36 |
| ; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:4 |
| ; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:116 |
| ; VI-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:124 |
| ; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v1 |
| ; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v3 |
| ; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v5 |
| ; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v7 |
| ; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v11 |
| ; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v13 |
| ; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v15 |
| ; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v17 |
| ; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v19 |
| ; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v23 |
| ; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v25 |
| ; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v27 |
| ; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v29 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24 |
| ; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v0 |
| ; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v2 |
| ; VI-NEXT: v_lshlrev_b16_e32 v47, 8, v4 |
| ; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v6 |
| ; VI-NEXT: v_lshlrev_b16_e32 v57, 8, v8 |
| ; VI-NEXT: v_lshlrev_b16_e32 v58, 8, v10 |
| ; VI-NEXT: v_lshlrev_b16_e32 v59, 8, v12 |
| ; VI-NEXT: v_lshlrev_b16_e32 v60, 8, v14 |
| ; VI-NEXT: v_lshlrev_b16_e32 v62, 8, v16 |
| ; VI-NEXT: v_lshlrev_b16_e32 v63, 8, v18 |
| ; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v20 |
| ; VI-NEXT: s_waitcnt vmcnt(13) |
| ; VI-NEXT: v_lshlrev_b16_e32 v24, 8, v22 |
| ; VI-NEXT: s_waitcnt vmcnt(12) |
| ; VI-NEXT: v_lshlrev_b16_e32 v18, 8, v26 |
| ; VI-NEXT: s_waitcnt vmcnt(11) |
| ; VI-NEXT: v_lshlrev_b16_e32 v26, 8, v28 |
| ; VI-NEXT: s_waitcnt vmcnt(10) |
| ; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v30 |
| ; VI-NEXT: s_waitcnt vmcnt(9) |
| ; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v31 |
| ; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:108 |
| ; VI-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:92 |
| ; VI-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:84 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:76 |
| ; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:60 |
| ; VI-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:52 |
| ; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:44 |
| ; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:28 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v32 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB49_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: v_or_b32_sdwa v9, v39, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v10, v20, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v28, v59 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v22, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v30, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v38, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr39 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: ; implicit-def: $vgpr30 |
| ; VI-NEXT: ; implicit-def: $vgpr35 |
| ; VI-NEXT: ; implicit-def: $vgpr31 |
| ; VI-NEXT: ; implicit-def: $vgpr38 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: ; implicit-def: $vgpr33 |
| ; VI-NEXT: ; implicit-def: $vgpr32 |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(7) |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr23 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v49, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v48, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v55, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v36, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v43, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v37, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v44, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v61, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr49 |
| ; VI-NEXT: ; implicit-def: $vgpr48 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr36 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr37 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr34 |
| ; VI-NEXT: .LBB49_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB49_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v38 |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v44 |
| ; VI-NEXT: v_or_b32_sdwa v0, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v14, v26, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v35 |
| ; VI-NEXT: v_mov_b32_e32 v3, 0x300 |
| ; VI-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v18, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v37 |
| ; VI-NEXT: v_or_b32_sdwa v24, v24, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v30 |
| ; VI-NEXT: v_or_b32_sdwa v0, v16, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v12, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v43 |
| ; VI-NEXT: v_or_b32_sdwa v16, v63, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v22 |
| ; VI-NEXT: v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v11, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v36 |
| ; VI-NEXT: v_or_b32_sdwa v22, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v28 |
| ; VI-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v10, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v55 |
| ; VI-NEXT: v_or_b32_sdwa v28, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v9, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v48 |
| ; VI-NEXT: v_or_b32_sdwa v20, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v39 |
| ; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v8, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v49 |
| ; VI-NEXT: v_or_b32_sdwa v30, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: v_add_u16_sdwa v13, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v61 |
| ; VI-NEXT: v_or_b32_sdwa v15, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v31 |
| ; VI-NEXT: v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v26, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v16, 0x300, v16 |
| ; VI-NEXT: v_or_b32_e32 v12, v16, v12 |
| ; VI-NEXT: v_add_u16_e32 v16, 0x300, v24 |
| ; VI-NEXT: v_add_u16_e32 v14, 0x300, v14 |
| ; VI-NEXT: v_add_u16_e32 v15, 0x300, v15 |
| ; VI-NEXT: v_or_b32_e32 v13, v16, v13 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v26 |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v18 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v7, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v29, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v6, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v25, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v5, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v21, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v17 |
| ; VI-NEXT: v_or_b32_e32 v4, v17, v4 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v21 |
| ; VI-NEXT: v_or_b32_e32 v5, v17, v5 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v25 |
| ; VI-NEXT: v_or_b32_e32 v6, v17, v6 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v29 |
| ; VI-NEXT: v_or_b32_e32 v7, v17, v7 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v30 |
| ; VI-NEXT: v_or_b32_e32 v8, v17, v8 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v20 |
| ; VI-NEXT: v_or_b32_e32 v9, v17, v9 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v28 |
| ; VI-NEXT: v_or_b32_e32 v10, v17, v10 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v22 |
| ; VI-NEXT: v_or_b32_e32 v11, v17, v11 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v19, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v23, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v2, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v27, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v1, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v31, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v3 |
| ; VI-NEXT: v_or_b32_e32 v0, v3, v0 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v31 |
| ; VI-NEXT: v_or_b32_e32 v1, v3, v1 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v27 |
| ; VI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v23 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v19 |
| ; VI-NEXT: .LBB49_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v64i8_to_v32i16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:132 |
| ; GFX9-NEXT: s_nop 0 |
| ; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64 |
| ; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72 |
| ; GFX9-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80 |
| ; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:88 |
| ; GFX9-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:96 |
| ; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:104 |
| ; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:112 |
| ; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:120 |
| ; GFX9-NEXT: buffer_load_ushort v33, off, s[0:3], s32 offset:128 |
| ; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:20 |
| ; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:12 |
| ; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:100 |
| ; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:68 |
| ; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:36 |
| ; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v1 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v3 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v5 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v7 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v9 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v11 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v13 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v15 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v17 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v19 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v21 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v23 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v25 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v27 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v29 |
| ; GFX9-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:124 |
| ; GFX9-NEXT: s_waitcnt vmcnt(24) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v22 |
| ; GFX9-NEXT: s_waitcnt vmcnt(23) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(22) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(21) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v4 |
| ; GFX9-NEXT: s_waitcnt vmcnt(20) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v57, 8, v6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(19) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v8 |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v59, 8, v10 |
| ; GFX9-NEXT: s_waitcnt vmcnt(17) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v58, 8, v12 |
| ; GFX9-NEXT: s_waitcnt vmcnt(16) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v61, 8, v14 |
| ; GFX9-NEXT: s_waitcnt vmcnt(15) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v60, 8, v16 |
| ; GFX9-NEXT: s_waitcnt vmcnt(14) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v18 |
| ; GFX9-NEXT: s_waitcnt vmcnt(13) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v20 |
| ; GFX9-NEXT: s_waitcnt vmcnt(12) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v30, 8, v24 |
| ; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:116 |
| ; GFX9-NEXT: s_waitcnt vmcnt(11) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28 |
| ; GFX9-NEXT: s_waitcnt vmcnt(10) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v31 |
| ; GFX9-NEXT: s_waitcnt vmcnt(9) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v31, 8, v32 |
| ; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:108 |
| ; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:92 |
| ; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:84 |
| ; GFX9-NEXT: buffer_load_ushort v34, off, s[0:3], s32 offset:76 |
| ; GFX9-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:60 |
| ; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:52 |
| ; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:44 |
| ; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v26, 8, v26 |
| ; GFX9-NEXT: s_waitcnt vmcnt(16) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB49_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(23) |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v39, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(9) |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v62, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v63, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v35, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v34, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v16, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v32, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v16, v38, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v0, v1, v0, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v3, v2, s6 |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v8, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6 |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6 |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6 |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v51, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v8, v9, v8, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v52, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v9, v10, v9, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v42, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v10, v11, v10, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v37, v61 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v11, v12, v11, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v44, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v12, v13, v12, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v28, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v13, v14, v13, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v47, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v14, v15, v14, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v18, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v15, v16, v15, s6 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: .LBB49_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB49_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v18 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v28 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v30, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v16 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v26, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v34 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v22, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v37 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v61, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v35 |
| ; GFX9-NEXT: v_add_u16_e32 v2, 3, v47 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v24, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v24, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v42 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v63 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v26, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v52 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v62 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v28, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v51 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v39 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v30, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v32 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v20, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v3 |
| ; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v2 |
| ; GFX9-NEXT: v_add_u16_e32 v1, 3, v38 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v1 |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v8, v30, v8, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v28, v9, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v26, v10, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v24, v11, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v22, v12, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v16, v13, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v20, v14, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v18, v15, s6 |
| ; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v31, 3, v31 |
| ; GFX9-NEXT: v_or_b32_sdwa v31, v48, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v31, 0x300, v31 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v7, v27, v7, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v6, v23, v6, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v5, v19, v5, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v17, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v4, v17, v4, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v21, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v3, v21, v3, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v55, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v25, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v2, v25, v2, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v29, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v1, v29, v1, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0 |
| ; GFX9-NEXT: v_perm_b32 v0, v31, v0, s6 |
| ; GFX9-NEXT: .LBB49_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v64i8_to_v32i16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_dual_mov_b32 v32, v14 :: v_dual_mov_b32 v31, v8 |
| ; GFX11-NEXT: v_dual_mov_b32 v34, v12 :: v_dual_mov_b32 v35, v10 |
| ; GFX11-NEXT: v_dual_mov_b32 v38, v6 :: v_dual_mov_b32 v33, v4 |
| ; GFX11-NEXT: v_dual_mov_b32 v37, v2 :: v_dual_mov_b32 v36, v0 |
| ; GFX11-NEXT: s_clause 0x1f |
| ; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:128 |
| ; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:124 |
| ; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:120 |
| ; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:116 |
| ; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:112 |
| ; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:108 |
| ; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:104 |
| ; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:100 |
| ; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:96 |
| ; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:92 |
| ; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:88 |
| ; GFX11-NEXT: scratch_load_b32 v12, off, s32 offset:132 |
| ; GFX11-NEXT: scratch_load_u16 v14, off, s32 |
| ; GFX11-NEXT: scratch_load_u16 v96, off, s32 offset:8 |
| ; GFX11-NEXT: scratch_load_u16 v100, off, s32 offset:16 |
| ; GFX11-NEXT: scratch_load_u16 v101, off, s32 offset:24 |
| ; GFX11-NEXT: scratch_load_u16 v102, off, s32 offset:32 |
| ; GFX11-NEXT: scratch_load_u16 v103, off, s32 offset:40 |
| ; GFX11-NEXT: scratch_load_u16 v113, off, s32 offset:48 |
| ; GFX11-NEXT: scratch_load_u16 v114, off, s32 offset:56 |
| ; GFX11-NEXT: scratch_load_u16 v115, off, s32 offset:64 |
| ; GFX11-NEXT: scratch_load_u16 v116, off, s32 offset:72 |
| ; GFX11-NEXT: scratch_load_u16 v117, off, s32 offset:80 |
| ; GFX11-NEXT: scratch_load_u16 v99, off, s32 offset:84 |
| ; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:76 |
| ; GFX11-NEXT: scratch_load_u16 v97, off, s32 offset:68 |
| ; GFX11-NEXT: scratch_load_u16 v82, off, s32 offset:60 |
| ; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:52 |
| ; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:44 |
| ; GFX11-NEXT: scratch_load_u16 v87, off, s32 offset:36 |
| ; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:28 |
| ; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:20 |
| ; GFX11-NEXT: s_clause 0x1 |
| ; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:12 |
| ; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:4 |
| ; GFX11-NEXT: v_lshlrev_b16 v50, 8, v1 |
| ; GFX11-NEXT: v_lshlrev_b16 v51, 8, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v49, 8, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v39, 8, v7 |
| ; GFX11-NEXT: v_lshlrev_b16 v54, 8, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v48, 8, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v55, 8, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v52, 8, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v53, 8, v17 |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v19 |
| ; GFX11-NEXT: v_lshlrev_b16 v83, 8, v21 |
| ; GFX11-NEXT: v_lshlrev_b16 v81, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v27, 8, v29 |
| ; GFX11-NEXT: s_waitcnt vmcnt(33) |
| ; GFX11-NEXT: v_lshlrev_b16 v119, 8, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(31) |
| ; GFX11-NEXT: v_lshlrev_b16 v118, 8, v2 |
| ; GFX11-NEXT: s_waitcnt vmcnt(22) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12 |
| ; GFX11-NEXT: s_waitcnt vmcnt(21) |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v14 |
| ; GFX11-NEXT: s_waitcnt vmcnt(20) |
| ; GFX11-NEXT: v_lshlrev_b16 v98, 8, v96 |
| ; GFX11-NEXT: s_waitcnt vmcnt(19) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v100 |
| ; GFX11-NEXT: s_waitcnt vmcnt(18) |
| ; GFX11-NEXT: v_lshlrev_b16 v96, 8, v101 |
| ; GFX11-NEXT: s_waitcnt vmcnt(17) |
| ; GFX11-NEXT: v_lshlrev_b16 v29, 8, v102 |
| ; GFX11-NEXT: s_waitcnt vmcnt(16) |
| ; GFX11-NEXT: v_lshlrev_b16 v112, 8, v103 |
| ; GFX11-NEXT: s_waitcnt vmcnt(15) |
| ; GFX11-NEXT: v_lshlrev_b16 v103, 8, v113 |
| ; GFX11-NEXT: s_waitcnt vmcnt(14) |
| ; GFX11-NEXT: v_lshlrev_b16 v101, 8, v114 |
| ; GFX11-NEXT: s_waitcnt vmcnt(13) |
| ; GFX11-NEXT: v_lshlrev_b16 v100, 8, v115 |
| ; GFX11-NEXT: s_waitcnt vmcnt(12) |
| ; GFX11-NEXT: v_lshlrev_b16 v114, 8, v116 |
| ; GFX11-NEXT: s_waitcnt vmcnt(11) |
| ; GFX11-NEXT: v_lshlrev_b16 v102, 8, v117 |
| ; GFX11-NEXT: v_lshlrev_b16 v117, 8, v10 |
| ; GFX11-NEXT: v_lshlrev_b16 v113, 8, v8 |
| ; GFX11-NEXT: v_lshlrev_b16 v116, 8, v6 |
| ; GFX11-NEXT: v_lshlrev_b16 v115, 8, v4 |
| ; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB49_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB49_4 |
| ; GFX11-NEXT: .LBB49_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB49_3: ; %cmp.false |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v37 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v38 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v35 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v50 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v34 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18 |
| ; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x5040100 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v3, v39 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v4, v54 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v5, v48 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v6, v55 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v32 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v49 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v52 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v81 |
| ; GFX11-NEXT: v_perm_b32 v1, v1, v2, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v2, v4, v3, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v3, v6, v5, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v4, v8, v7, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v5, v10, v9, 0x5040100 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v30 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v71 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v65 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v68 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v87 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v27 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v98 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v25 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v96 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v29 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v112 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v103 |
| ; GFX11-NEXT: v_perm_b32 v6, v7, v6, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v7, v9, v8, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v8, v11, v10, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v9, v13, v12, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v10, v15, v14, 0x5040100 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v86 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v82 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v97 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v99 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v66 |
| ; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v70 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v101 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v100 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v114 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v102 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v117 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v113 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v17, v116 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v18, v115 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v19, v118 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v20, v119 |
| ; GFX11-NEXT: v_perm_b32 v11, v12, v11, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v12, v14, v13, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v13, v16, v15, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v14, v18, v17, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v15, v20, v19, 0x5040100 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr16 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr97 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr99 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr98 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr112 |
| ; GFX11-NEXT: ; implicit-def: $vgpr103 |
| ; GFX11-NEXT: ; implicit-def: $vgpr101 |
| ; GFX11-NEXT: ; implicit-def: $vgpr100 |
| ; GFX11-NEXT: ; implicit-def: $vgpr114 |
| ; GFX11-NEXT: ; implicit-def: $vgpr102 |
| ; GFX11-NEXT: ; implicit-def: $vgpr117 |
| ; GFX11-NEXT: ; implicit-def: $vgpr113 |
| ; GFX11-NEXT: ; implicit-def: $vgpr116 |
| ; GFX11-NEXT: ; implicit-def: $vgpr115 |
| ; GFX11-NEXT: ; implicit-def: $vgpr118 |
| ; GFX11-NEXT: ; implicit-def: $vgpr119 |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB49_2 |
| ; GFX11-NEXT: .LBB49_4: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u16 v0, v70, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v67, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v69, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v66, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v5, v64, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: s_waitcnt vmcnt(8) |
| ; GFX11-NEXT: v_add_nc_u16 v6, v97, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v118, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v119, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v99, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v64, 0x300, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v116, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v115, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v113, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v66, 0x300, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v114, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v117, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v85, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v67, 0x300, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(6) |
| ; GFX11-NEXT: v_add_nc_u16 v0, v86, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v82, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: s_waitcnt vmcnt(4) |
| ; GFX11-NEXT: v_add_nc_u16 v3, v87, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v84, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v102, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v101, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v100, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v69, 0x300, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v112, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v103, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v70, 0x300, v1 |
| ; GFX11-NEXT: s_waitcnt vmcnt(2) |
| ; GFX11-NEXT: v_add_nc_u16 v1, v80, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v68, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v68, 0x300, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_add_nc_u16 v0, v71, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v65, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v28, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v96, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v29, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v98, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v25, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v27, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v0, v30, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v24, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v26, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v20, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v22, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v23, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v21, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v19, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v83, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v81, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v16, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v0, v34, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v31, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v18, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v32, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v53, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v55, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v54, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v17, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v52, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v33, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v17, v36, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v21, v37, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v22, v38, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v23, v35, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v49, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v50, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v51, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v39, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v48, v23 |
| ; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17 |
| ; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v21 |
| ; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v22 |
| ; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v23 |
| ; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v2 |
| ; GFX11-NEXT: v_perm_b32 v0, v21, v17, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v1, v22, v1, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v2, v23, v4, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v3, v24, v3, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v4, v26, v18, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v5, v16, v5, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v6, v20, v6, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v7, v19, v7, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v8, v27, v8, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v9, v25, v9, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v10, v68, v10, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v11, v70, v11, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v12, v69, v12, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v13, v67, v13, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v14, v66, v14, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v15, v64, v15, 0x5040100 |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <64 x i8> %a, splat (i8 3) |
| %a2 = bitcast <64 x i8> %a1 to <32 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <64 x i8> %a to <32 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <32 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x i16> %phi |
| } |
| |
| define <32 x bfloat> @bitcast_v32f16_to_v32bf16(<32 x half> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32f16_to_v32bf16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v32, v0 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v33, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v34, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v35, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v36, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v37, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v38, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v39, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v48, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v49, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v50, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v51, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v52, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v53, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v54, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v55, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v40, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v41, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v42, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v43, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v44, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v45, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v46, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v47, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v56, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v57, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v58, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v59, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v60, v28 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v61, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v62, v30 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v63 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v63, v31 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB50_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v34 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v36 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v48 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v55 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v40 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v42 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v44 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v46 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v58 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v59 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v60 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v61 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v62 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v63 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: .LBB50_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB50_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v63 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v62 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v61 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v60 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v59 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v58 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v57 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v56 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v47 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v46 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v45 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v44 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v43 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v42 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v41 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v40 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v55 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v53 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v32 |
| ; GCN-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30 |
| ; GCN-NEXT: v_add_f32_e32 v31, 0x38000000, v31 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v31, v31 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v30, v30 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v28, v28 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v32, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v33, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v34, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v35, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v36, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v37, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v38, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v39, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v48, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v49, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v50, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v51, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v52, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v53, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v54, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v55, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v30 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v26 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v21 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v18 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v34 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v36 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v48 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v55 |
| ; GCN-NEXT: .LBB50_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32f16_to_v32bf16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB50_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v17, 0x200 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v15 |
| ; VI-NEXT: v_add_f16_sdwa v15, v15, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v15, v19, v15 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v14 |
| ; VI-NEXT: v_add_f16_sdwa v14, v14, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v14, v19, v14 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v13 |
| ; VI-NEXT: v_add_f16_sdwa v13, v13, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v13, v19, v13 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v12 |
| ; VI-NEXT: v_add_f16_sdwa v12, v12, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v12, v19, v12 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v11 |
| ; VI-NEXT: v_add_f16_sdwa v11, v11, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v11, v19, v11 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v10 |
| ; VI-NEXT: v_add_f16_sdwa v10, v10, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v10, v19, v10 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v9 |
| ; VI-NEXT: v_add_f16_sdwa v9, v9, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v9, v19, v9 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v8 |
| ; VI-NEXT: v_add_f16_sdwa v8, v8, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v8, v19, v8 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v7 |
| ; VI-NEXT: v_add_f16_sdwa v7, v7, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v7, v19, v7 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v6 |
| ; VI-NEXT: v_add_f16_sdwa v6, v6, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v6, v19, v6 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v5 |
| ; VI-NEXT: v_add_f16_sdwa v5, v5, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v5, v19, v5 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v4 |
| ; VI-NEXT: v_add_f16_sdwa v4, v4, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v16, 0x200, v0 |
| ; VI-NEXT: v_add_f16_sdwa v0, v0, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v18, 0x200, v1 |
| ; VI-NEXT: v_add_f16_sdwa v1, v1, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v4, v19, v4 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v2 |
| ; VI-NEXT: v_add_f16_sdwa v2, v2, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_sdwa v17, v3, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, 0x200, v3 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v17 |
| ; VI-NEXT: v_or_b32_e32 v2, v19, v2 |
| ; VI-NEXT: v_or_b32_e32 v1, v18, v1 |
| ; VI-NEXT: v_or_b32_e32 v0, v16, v0 |
| ; VI-NEXT: .LBB50_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32f16_to_v32bf16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB50_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: s_movk_i32 s6, 0x200 |
| ; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB50_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32f16_to_v32bf16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB50_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-NEXT: .LBB50_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <32 x half> %a1 to <32 x bfloat> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x half> %a to <32 x bfloat> |
| br label %end |
| |
| end: |
| %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x bfloat> %phi |
| } |
| |
| define <32 x half> @bitcast_v32bf16_to_v32f16(<32 x bfloat> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32bf16_to_v32f16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v0 |
| ; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v1 |
| ; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v2 |
| ; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v3 |
| ; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v4 |
| ; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v5 |
| ; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v6 |
| ; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v7 |
| ; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v8 |
| ; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v9 |
| ; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v10 |
| ; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v11 |
| ; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v12 |
| ; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v13 |
| ; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v14 |
| ; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v15 |
| ; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v16 |
| ; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v17 |
| ; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v18 |
| ; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v19 |
| ; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v20 |
| ; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v21 |
| ; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v22 |
| ; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v23 |
| ; GCN-NEXT: v_mul_f32_e32 v56, 1.0, v24 |
| ; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v25 |
| ; GCN-NEXT: v_mul_f32_e32 v58, 1.0, v26 |
| ; GCN-NEXT: v_mul_f32_e32 v59, 1.0, v27 |
| ; GCN-NEXT: v_mul_f32_e32 v60, 1.0, v28 |
| ; GCN-NEXT: v_mul_f32_e32 v61, 1.0, v29 |
| ; GCN-NEXT: v_mul_f32_e32 v62, 1.0, v30 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v63 |
| ; GCN-NEXT: v_mul_f32_e32 v63, 1.0, v31 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB51_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v33 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v34 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v35 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v36 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v37 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v38 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v39 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v48 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v49 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v50 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v51 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v52 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v54 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v55 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v40 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v41 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v42 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v43 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v44 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v45 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v46 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v47 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v56 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v57 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v58 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v59 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v60 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v61 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v62 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v63 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v20 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v22 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v24 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v26 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v28 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v29 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v30 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: .LBB51_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB51_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v63 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v62 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v61 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v60 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v59 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v58 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v57 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v56 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v47 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v46 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v45 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v44 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v43 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v42 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v41 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v40 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v55 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v54 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v53 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v52 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v51 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v50 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v49 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v48 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v39 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v38 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v37 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v36 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v35 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v34 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v33 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v32 |
| ; GCN-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x40c00000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v27, 0x40c00000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x40c00000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x40c00000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v30, 0x40c00000, v30 |
| ; GCN-NEXT: v_add_f32_e32 v31, 0x40c00000, v31 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v0 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v33, 16, v1 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v34, 16, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v35, 16, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v38, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v39, 16, v7 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v48, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v9 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v53, 16, v13 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v54, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v55, 16, v15 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v17 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v18 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v19 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v20 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v21 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v22 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v23 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v24 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v25 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v26 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v27 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v28 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v29 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v30 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v31 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v10 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v12 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v55 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v53 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v39 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v33 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v32 |
| ; GCN-NEXT: .LBB51_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32bf16_to_v32f16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB51_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v16 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 0x7fff, v17 |
| ; VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; VI-NEXT: v_bfe_u32 v17, v0, 16, 1 |
| ; VI-NEXT: s_movk_i32 s6, 0x7fff |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, v17, v0 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, s6, v17 |
| ; VI-NEXT: v_or_b32_e32 v18, 0x400000, v0 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v0, v0 |
| ; VI-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v1 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v1, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v1 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v1, v18, v19, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v2 |
| ; VI-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; VI-NEXT: v_bfe_u32 v19, v18, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v19, vcc, v19, v18 |
| ; VI-NEXT: v_add_u32_e32 v19, vcc, s6, v19 |
| ; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; VI-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v18, v18 |
| ; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc |
| ; VI-NEXT: v_bfe_u32 v19, v2, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v19, vcc, v19, v2 |
| ; VI-NEXT: v_add_u32_e32 v19, vcc, s6, v19 |
| ; VI-NEXT: v_or_b32_e32 v20, 0x400000, v2 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v2, v19, v20, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v19, 16, v3 |
| ; VI-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 |
| ; VI-NEXT: v_bfe_u32 v20, v19, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v20, vcc, v20, v19 |
| ; VI-NEXT: v_add_u32_e32 v20, vcc, s6, v20 |
| ; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; VI-NEXT: v_or_b32_e32 v21, 0x400000, v19 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v19, v19 |
| ; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v19, v20, v21, vcc |
| ; VI-NEXT: v_bfe_u32 v20, v3, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v20, vcc, v20, v3 |
| ; VI-NEXT: v_add_u32_e32 v20, vcc, s6, v20 |
| ; VI-NEXT: v_or_b32_e32 v21, 0x400000, v3 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v4 |
| ; VI-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 |
| ; VI-NEXT: v_bfe_u32 v21, v20, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v21, vcc, v21, v20 |
| ; VI-NEXT: v_add_u32_e32 v21, vcc, s6, v21 |
| ; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; VI-NEXT: v_or_b32_e32 v22, 0x400000, v20 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v20, v20 |
| ; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v20, v21, v22, vcc |
| ; VI-NEXT: v_bfe_u32 v21, v4, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v21, vcc, v21, v4 |
| ; VI-NEXT: v_add_u32_e32 v21, vcc, s6, v21 |
| ; VI-NEXT: v_or_b32_e32 v22, 0x400000, v4 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v4, v21, v22, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v5 |
| ; VI-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 |
| ; VI-NEXT: v_bfe_u32 v22, v21, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v22, vcc, v22, v21 |
| ; VI-NEXT: v_add_u32_e32 v22, vcc, s6, v22 |
| ; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; VI-NEXT: v_or_b32_e32 v23, 0x400000, v21 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v21, v21 |
| ; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v21, v22, v23, vcc |
| ; VI-NEXT: v_bfe_u32 v22, v5, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v22, vcc, v22, v5 |
| ; VI-NEXT: v_add_u32_e32 v22, vcc, s6, v22 |
| ; VI-NEXT: v_or_b32_e32 v23, 0x400000, v5 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v5, v22, v23, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v6 |
| ; VI-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; VI-NEXT: v_bfe_u32 v23, v22, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v23, vcc, v23, v22 |
| ; VI-NEXT: v_add_u32_e32 v23, vcc, s6, v23 |
| ; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; VI-NEXT: v_or_b32_e32 v24, 0x400000, v22 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v22, v22 |
| ; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v22, v23, v24, vcc |
| ; VI-NEXT: v_bfe_u32 v23, v6, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v23, vcc, v23, v6 |
| ; VI-NEXT: v_add_u32_e32 v23, vcc, s6, v23 |
| ; VI-NEXT: v_or_b32_e32 v24, 0x400000, v6 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v6, v23, v24, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v7 |
| ; VI-NEXT: v_add_f32_e32 v23, 0x40c00000, v23 |
| ; VI-NEXT: v_bfe_u32 v24, v23, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v24, vcc, v24, v23 |
| ; VI-NEXT: v_add_u32_e32 v24, vcc, s6, v24 |
| ; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; VI-NEXT: v_or_b32_e32 v25, 0x400000, v23 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v23, v23 |
| ; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v23, v24, v25, vcc |
| ; VI-NEXT: v_bfe_u32 v24, v7, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v24, vcc, v24, v7 |
| ; VI-NEXT: v_add_u32_e32 v24, vcc, s6, v24 |
| ; VI-NEXT: v_or_b32_e32 v25, 0x400000, v7 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v7, v24, v25, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v8 |
| ; VI-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 |
| ; VI-NEXT: v_bfe_u32 v25, v24, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v25, vcc, v25, v24 |
| ; VI-NEXT: v_add_u32_e32 v25, vcc, s6, v25 |
| ; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; VI-NEXT: v_or_b32_e32 v26, 0x400000, v24 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v24, v24 |
| ; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v24, v25, v26, vcc |
| ; VI-NEXT: v_bfe_u32 v25, v8, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v25, vcc, v25, v8 |
| ; VI-NEXT: v_add_u32_e32 v25, vcc, s6, v25 |
| ; VI-NEXT: v_or_b32_e32 v26, 0x400000, v8 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v8, v25, v26, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v9 |
| ; VI-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 |
| ; VI-NEXT: v_bfe_u32 v26, v25, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v26, vcc, v26, v25 |
| ; VI-NEXT: v_add_u32_e32 v26, vcc, s6, v26 |
| ; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; VI-NEXT: v_or_b32_e32 v27, 0x400000, v25 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v25, v25 |
| ; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v25, v26, v27, vcc |
| ; VI-NEXT: v_bfe_u32 v26, v9, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v26, vcc, v26, v9 |
| ; VI-NEXT: v_add_u32_e32 v26, vcc, s6, v26 |
| ; VI-NEXT: v_or_b32_e32 v27, 0x400000, v9 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v9, v26, v27, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v10 |
| ; VI-NEXT: v_add_f32_e32 v26, 0x40c00000, v26 |
| ; VI-NEXT: v_bfe_u32 v27, v26, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v27, vcc, v27, v26 |
| ; VI-NEXT: v_add_u32_e32 v27, vcc, s6, v27 |
| ; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; VI-NEXT: v_or_b32_e32 v28, 0x400000, v26 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v26, v26 |
| ; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v26, v27, v28, vcc |
| ; VI-NEXT: v_bfe_u32 v27, v10, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v27, vcc, v27, v10 |
| ; VI-NEXT: v_add_u32_e32 v27, vcc, s6, v27 |
| ; VI-NEXT: v_or_b32_e32 v28, 0x400000, v10 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v10, v27, v28, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v11 |
| ; VI-NEXT: v_add_f32_e32 v27, 0x40c00000, v27 |
| ; VI-NEXT: v_bfe_u32 v28, v27, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v28, vcc, v28, v27 |
| ; VI-NEXT: v_add_u32_e32 v28, vcc, s6, v28 |
| ; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; VI-NEXT: v_or_b32_e32 v29, 0x400000, v27 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v27, v27 |
| ; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v27, v28, v29, vcc |
| ; VI-NEXT: v_bfe_u32 v28, v11, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v28, vcc, v28, v11 |
| ; VI-NEXT: v_add_u32_e32 v28, vcc, s6, v28 |
| ; VI-NEXT: v_or_b32_e32 v29, 0x400000, v11 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v11, v28, v29, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v12 |
| ; VI-NEXT: v_add_f32_e32 v28, 0x40c00000, v28 |
| ; VI-NEXT: v_bfe_u32 v29, v28, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v29, vcc, v29, v28 |
| ; VI-NEXT: v_add_u32_e32 v29, vcc, s6, v29 |
| ; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; VI-NEXT: v_or_b32_e32 v30, 0x400000, v28 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v28, v28 |
| ; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v28, v29, v30, vcc |
| ; VI-NEXT: v_bfe_u32 v29, v12, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v29, vcc, v29, v12 |
| ; VI-NEXT: v_add_u32_e32 v29, vcc, s6, v29 |
| ; VI-NEXT: v_or_b32_e32 v30, 0x400000, v12 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v12, v29, v30, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v13 |
| ; VI-NEXT: v_add_f32_e32 v29, 0x40c00000, v29 |
| ; VI-NEXT: v_bfe_u32 v30, v29, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v30, vcc, v30, v29 |
| ; VI-NEXT: v_add_u32_e32 v30, vcc, s6, v30 |
| ; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; VI-NEXT: v_or_b32_e32 v31, 0x400000, v29 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v29, v29 |
| ; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v29, v30, v31, vcc |
| ; VI-NEXT: v_bfe_u32 v30, v13, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v30, vcc, v30, v13 |
| ; VI-NEXT: v_add_u32_e32 v30, vcc, s6, v30 |
| ; VI-NEXT: v_or_b32_e32 v31, 0x400000, v13 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v13, v30, v31, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v14 |
| ; VI-NEXT: v_add_f32_e32 v30, 0x40c00000, v30 |
| ; VI-NEXT: v_bfe_u32 v31, v30, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v31, vcc, v31, v30 |
| ; VI-NEXT: v_add_u32_e32 v31, vcc, s6, v31 |
| ; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; VI-NEXT: v_or_b32_e32 v32, 0x400000, v30 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v30, v30 |
| ; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v30, v31, v32, vcc |
| ; VI-NEXT: v_bfe_u32 v31, v14, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v31, vcc, v31, v14 |
| ; VI-NEXT: v_add_u32_e32 v31, vcc, s6, v31 |
| ; VI-NEXT: v_or_b32_e32 v32, 0x400000, v14 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v14, v31, v32, vcc |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v15 |
| ; VI-NEXT: v_add_f32_e32 v31, 0x40c00000, v31 |
| ; VI-NEXT: v_bfe_u32 v32, v31, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v32, vcc, v32, v31 |
| ; VI-NEXT: v_add_u32_e32 v32, vcc, s6, v32 |
| ; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; VI-NEXT: v_or_b32_e32 v33, 0x400000, v31 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v31, v31 |
| ; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc |
| ; VI-NEXT: v_bfe_u32 v32, v15, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v32, vcc, v32, v15 |
| ; VI-NEXT: v_add_u32_e32 v32, vcc, s6, v32 |
| ; VI-NEXT: v_or_b32_e32 v33, 0x400000, v15 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v15, v32, v33, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 |
| ; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; VI-NEXT: v_alignbit_b32 v15, v15, v31, 16 |
| ; VI-NEXT: v_alignbit_b32 v14, v14, v30, 16 |
| ; VI-NEXT: v_alignbit_b32 v13, v13, v29, 16 |
| ; VI-NEXT: v_alignbit_b32 v12, v12, v28, 16 |
| ; VI-NEXT: v_alignbit_b32 v11, v11, v27, 16 |
| ; VI-NEXT: v_alignbit_b32 v10, v10, v26, 16 |
| ; VI-NEXT: v_alignbit_b32 v9, v9, v25, 16 |
| ; VI-NEXT: v_alignbit_b32 v8, v8, v24, 16 |
| ; VI-NEXT: v_alignbit_b32 v7, v7, v23, 16 |
| ; VI-NEXT: v_alignbit_b32 v6, v6, v22, 16 |
| ; VI-NEXT: v_alignbit_b32 v5, v5, v21, 16 |
| ; VI-NEXT: v_alignbit_b32 v4, v4, v20, 16 |
| ; VI-NEXT: v_alignbit_b32 v3, v3, v19, 16 |
| ; VI-NEXT: v_alignbit_b32 v2, v2, v18, 16 |
| ; VI-NEXT: v_alignbit_b32 v1, v1, v17, 16 |
| ; VI-NEXT: v_alignbit_b32 v0, v0, v16, 16 |
| ; VI-NEXT: .LBB51_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32bf16_to_v32f16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB51_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1 |
| ; GFX9-NEXT: s_movk_i32 s6, 0x7fff |
| ; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc |
| ; GFX9-NEXT: v_bfe_u32 v17, v0, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v17, v17, v0, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v0 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v18, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; GFX9-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX9-NEXT: v_add3_u32 v18, v18, v17, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; GFX9-NEXT: v_bfe_u32 v18, v1, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v18, v18, v1, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v19, 0x400000, v1 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v1, v18, v19, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v18, 16, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GFX9-NEXT: v_bfe_u32 v19, v18, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX9-NEXT: v_add3_u32 v19, v19, v18, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc |
| ; GFX9-NEXT: v_bfe_u32 v19, v2, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v19, v19, v2, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v2 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v2, v19, v20, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v19, 16, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 |
| ; GFX9-NEXT: v_bfe_u32 v20, v19, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX9-NEXT: v_add3_u32 v20, v20, v19, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v21, 0x400000, v19 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v19, v20, v21, vcc |
| ; GFX9-NEXT: v_bfe_u32 v20, v3, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v20, v20, v3, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v21, 0x400000, v3 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v20, 16, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 |
| ; GFX9-NEXT: v_bfe_u32 v21, v20, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX9-NEXT: v_add3_u32 v21, v21, v20, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v22, 0x400000, v20 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v20, v21, v22, vcc |
| ; GFX9-NEXT: v_bfe_u32 v21, v4, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v21, v21, v4, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v22, 0x400000, v4 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v4, v21, v22, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v21, 16, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 |
| ; GFX9-NEXT: v_bfe_u32 v22, v21, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX9-NEXT: v_add3_u32 v22, v22, v21, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v23, 0x400000, v21 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v21, v22, v23, vcc |
| ; GFX9-NEXT: v_bfe_u32 v22, v5, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v22, v22, v5, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v23, 0x400000, v5 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v5, v22, v23, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v22, 16, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; GFX9-NEXT: v_bfe_u32 v23, v22, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX9-NEXT: v_add3_u32 v23, v23, v22, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v24, 0x400000, v22 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v22, v23, v24, vcc |
| ; GFX9-NEXT: v_bfe_u32 v23, v6, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v23, v23, v6, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v24, 0x400000, v6 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v6, v23, v24, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v23, 16, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v23, 0x40c00000, v23 |
| ; GFX9-NEXT: v_bfe_u32 v24, v23, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GFX9-NEXT: v_add3_u32 v24, v24, v23, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v25, 0x400000, v23 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v23, v24, v25, vcc |
| ; GFX9-NEXT: v_bfe_u32 v24, v7, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v24, v24, v7, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v25, 0x400000, v7 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v7, v24, v25, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v24, 16, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 |
| ; GFX9-NEXT: v_bfe_u32 v25, v24, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX9-NEXT: v_add3_u32 v25, v25, v24, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v26, 0x400000, v24 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v24, v25, v26, vcc |
| ; GFX9-NEXT: v_bfe_u32 v25, v8, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v25, v25, v8, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v26, 0x400000, v8 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v8, v25, v26, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v25, 16, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 |
| ; GFX9-NEXT: v_bfe_u32 v26, v25, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX9-NEXT: v_add3_u32 v26, v26, v25, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v27, 0x400000, v25 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v25, v26, v27, vcc |
| ; GFX9-NEXT: v_bfe_u32 v26, v9, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v26, v26, v9, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v27, 0x400000, v9 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v9, v26, v27, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v26, 16, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v26, 0x40c00000, v26 |
| ; GFX9-NEXT: v_bfe_u32 v27, v26, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX9-NEXT: v_add3_u32 v27, v27, v26, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v28, 0x400000, v26 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v26, v27, v28, vcc |
| ; GFX9-NEXT: v_bfe_u32 v27, v10, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v27, v27, v10, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v28, 0x400000, v10 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v10, v27, v28, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v27, 16, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v27, 0x40c00000, v27 |
| ; GFX9-NEXT: v_bfe_u32 v28, v27, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; GFX9-NEXT: v_add3_u32 v28, v28, v27, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v29, 0x400000, v27 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v27, v28, v29, vcc |
| ; GFX9-NEXT: v_bfe_u32 v28, v11, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v28, v28, v11, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v29, 0x400000, v11 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v11, v28, v29, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v28, 16, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v28, 0x40c00000, v28 |
| ; GFX9-NEXT: v_bfe_u32 v29, v28, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX9-NEXT: v_add3_u32 v29, v29, v28, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v30, 0x400000, v28 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v28, v29, v30, vcc |
| ; GFX9-NEXT: v_bfe_u32 v29, v12, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v29, v29, v12, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v30, 0x400000, v12 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v12, v29, v30, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v29, 16, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v29, 0x40c00000, v29 |
| ; GFX9-NEXT: v_bfe_u32 v30, v29, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX9-NEXT: v_add3_u32 v30, v30, v29, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v31, 0x400000, v29 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v29, v30, v31, vcc |
| ; GFX9-NEXT: v_bfe_u32 v30, v13, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v30, v30, v13, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v31, 0x400000, v13 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v13, v30, v31, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v30, 16, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v30, 0x40c00000, v30 |
| ; GFX9-NEXT: v_bfe_u32 v31, v30, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX9-NEXT: v_add3_u32 v31, v31, v30, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v30 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v30, v31, v32, vcc |
| ; GFX9-NEXT: v_bfe_u32 v31, v14, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v31, v31, v14, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v14 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v14, v31, v32, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v31, 16, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v31, 0x40c00000, v31 |
| ; GFX9-NEXT: v_bfe_u32 v32, v31, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX9-NEXT: v_add3_u32 v32, v32, v31, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v31 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31 |
| ; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc |
| ; GFX9-NEXT: v_bfe_u32 v32, v15, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v32, v32, v15, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v15 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v15, v32, v33, vcc |
| ; GFX9-NEXT: s_mov_b32 s6, 0x7060302 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v31, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v30, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v29, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v28, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v11, v27, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v10, v26, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v9, v25, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v8, v24, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v7, v23, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v6, v22, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v5, v21, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v4, v20, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v3, v19, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v2, v18, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v1, v17, s6 |
| ; GFX9-NEXT: v_perm_b32 v0, v0, v16, s6 |
| ; GFX9-NEXT: .LBB51_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32bf16_to_v32f16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v16 |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB51_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v1 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v0 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v11 |
| ; GFX11-NEXT: v_dual_add_f32 v24, 0x40c00000, v24 :: v_dual_lshlrev_b32 v25, 16, v6 |
| ; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v19, v16, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v16 |
| ; GFX11-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v17, 0x7fff |
| ; GFX11-NEXT: v_add3_u32 v19, v19, v16, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX11-NEXT: v_dual_add_f32 v26, 0x40c00000, v26 :: v_dual_lshlrev_b32 v27, 16, v8 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v19, v22, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 |
| ; GFX11-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v22, 16, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX11-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX11-NEXT: v_add_f32_e32 v27, 0x40c00000, v27 |
| ; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v10 |
| ; GFX11-NEXT: v_bfe_u32 v20, v0, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v0 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 |
| ; GFX11-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GFX11-NEXT: v_dual_add_f32 v28, 0x40c00000, v28 :: v_dual_add_f32 v29, 0x40c00000, v29 |
| ; GFX11-NEXT: v_add3_u32 v20, v20, v0, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX11-NEXT: v_dual_add_f32 v30, 0x40c00000, v30 :: v_dual_lshlrev_b32 v31, 16, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v0, v20, v23 :: v_dual_lshlrev_b32 v23, 16, v4 |
| ; GFX11-NEXT: v_bfe_u32 v20, v1, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX11-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_add_f32 v23, 0x40c00000, v23 |
| ; GFX11-NEXT: v_perm_b32 v0, v0, v16, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v21, v19, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v19, v20, v1, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX11-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v31, 0x40c00000, v31 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v1, v19, v20 :: v_dual_add_f32 v18, 0x40c00000, v18 |
| ; GFX11-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX11-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; GFX11-NEXT: v_bfe_u32 v21, v18, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v33, 0x400000, v31 |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v18, 0x7fff |
| ; GFX11-NEXT: v_bfe_u32 v21, v2, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v34, v12, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v1, v1, v17, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v18, v19, v20 :: v_dual_and_b32 v7, 0xffff0000, v7 |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v2, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v2 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 |
| ; GFX11-NEXT: v_bfe_u32 v21, v22, 16, 1 |
| ; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v2, v19, v20 :: v_dual_and_b32 v11, 0xffff0000, v11 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v19, v21, v22, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v22 |
| ; GFX11-NEXT: v_bfe_u32 v21, v3, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 |
| ; GFX11-NEXT: v_bfe_u32 v22, v23, 16, 1 |
| ; GFX11-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GFX11-NEXT: v_perm_b32 v2, v2, v18, 0x7060302 |
| ; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v19, v19, v20, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v20, v21, v3, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v3 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v32, 0x400000, v11 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v20, v22, v23, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v23 |
| ; GFX11-NEXT: v_bfe_u32 v22, v4, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 |
| ; GFX11-NEXT: v_bfe_u32 v23, v24, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v3, v3, v19, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v20, v20, v21, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v21, v22, v4, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v4 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v4, v21, v22, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v21, v23, v24, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v24 |
| ; GFX11-NEXT: v_bfe_u32 v23, v5, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24 |
| ; GFX11-NEXT: v_bfe_u32 v24, v25, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v4, v4, v20, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v21, v21, v22, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v22, v23, v5, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v5 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v5, v22, v23, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v22, v24, v25, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v25 |
| ; GFX11-NEXT: v_bfe_u32 v24, v6, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25 |
| ; GFX11-NEXT: v_bfe_u32 v25, v26, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v5, v5, v21, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v22, v22, v23, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v23, v24, v6, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v6 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v6, v23, v24, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v23, v25, v26, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v26 |
| ; GFX11-NEXT: v_bfe_u32 v25, v7, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26 |
| ; GFX11-NEXT: v_bfe_u32 v26, v27, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v6, v6, v22, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v23, v23, v24, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v24, v25, v7, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v7 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v7, v24, v25, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v24, v26, v27, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v27 |
| ; GFX11-NEXT: v_bfe_u32 v26, v8, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27 |
| ; GFX11-NEXT: v_bfe_u32 v27, v28, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v7, v7, v23, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v24, v24, v25, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v25, v26, v8, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v8 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v8, v25, v26, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v25, v27, v28, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v26, 0x400000, v28 |
| ; GFX11-NEXT: v_bfe_u32 v27, v9, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28 |
| ; GFX11-NEXT: v_bfe_u32 v28, v29, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v8, v8, v24, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v25, v25, v26, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v26, v27, v9, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v27, 0x400000, v9 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v9, v26, v27, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v26, v28, v29, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v27, 0x400000, v29 |
| ; GFX11-NEXT: v_bfe_u32 v28, v10, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29 |
| ; GFX11-NEXT: v_bfe_u32 v29, v30, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v9, v9, v25, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v26, v26, v27, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v27, v28, v10, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v28, 0x400000, v10 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v10, v27, v28, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v27, v29, v30, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v28, 0x400000, v30 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30 |
| ; GFX11-NEXT: v_bfe_u32 v30, v31, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v29, v11, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v27, v27, v28 :: v_dual_lshlrev_b32 v28, 16, v13 |
| ; GFX11-NEXT: v_add3_u32 v30, v30, v31, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX11-NEXT: v_add3_u32 v31, v34, v12, 0x7fff |
| ; GFX11-NEXT: v_add_f32_e32 v28, 0x40c00000, v28 |
| ; GFX11-NEXT: v_add3_u32 v29, v29, v11, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v30, v30, v33, vcc_lo |
| ; GFX11-NEXT: v_or_b32_e32 v33, 0x400000, v12 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 |
| ; GFX11-NEXT: v_bfe_u32 v35, v28, 16, 1 |
| ; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v28 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v12, v31, v33, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v34, v35, v28, 0x7fff |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v35, 16, v14 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX11-NEXT: v_bfe_u32 v37, v13, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v10, v10, v26, 0x7060302 |
| ; GFX11-NEXT: v_dual_add_f32 v31, 0x40c00000, v35 :: v_dual_cndmask_b32 v28, v34, v36 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v15 |
| ; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX11-NEXT: v_add3_u32 v33, v37, v13, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v35, v31, 16, 1 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX11-NEXT: v_add_f32_e32 v34, 0x40c00000, v34 |
| ; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v31 |
| ; GFX11-NEXT: v_bfe_u32 v38, v14, 16, 1 |
| ; GFX11-NEXT: v_add3_u32 v35, v35, v31, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 |
| ; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX11-NEXT: v_bfe_u32 v39, v34, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v34 |
| ; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v13 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v31, v35, v37, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v37, v38, v14, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v14 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14 |
| ; GFX11-NEXT: v_bfe_u32 v35, v15, 16, 1 |
| ; GFX11-NEXT: v_add3_u32 v39, v39, v34, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v49, 0x400000, v15 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v14, v37, v38, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34 |
| ; GFX11-NEXT: v_add3_u32 v35, v35, v15, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v14, v14, v31, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v34, v39, v48, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v15, v35, v49, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_perm_b32 v15, v15, v34, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v13, v33, v36, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 |
| ; GFX11-NEXT: v_perm_b32 v12, v12, v30, 0x7060302 |
| ; GFX11-NEXT: v_perm_b32 v13, v13, v28, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v11, v29, v32, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) |
| ; GFX11-NEXT: v_perm_b32 v11, v11, v27, 0x7060302 |
| ; GFX11-NEXT: .LBB51_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x bfloat> %a, splat (bfloat 0xR40C0) |
| %a2 = bitcast <32 x bfloat> %a1 to <32 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x bfloat> %a to <32 x half> |
| br label %end |
| |
| end: |
| %phi = phi <32 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x half> %phi |
| } |
| |
| define <64 x i8> @bitcast_v32f16_to_v64i8(<32 x half> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32f16_to_v64i8: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v38, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v34, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v31, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v32, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v49, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v35, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v50, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v36, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v6, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v9, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v51, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v37, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v5, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v13, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v52, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v4, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v54, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v15, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v3, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v24, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v2, v28 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v27, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v28, v30 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v29 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v48 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v1, v33 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v39 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; kill: killed $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; kill: killed $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; kill: killed $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB52_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 16, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v1 |
| ; GCN-NEXT: v_bfe_u32 v33, v31, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v33, v8, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v33, v6, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v33, v5, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v33, v4, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v33, v3, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v33, v2, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_or_b32_e32 v57, v34, v10 |
| ; GCN-NEXT: v_or_b32_e32 v56, v32, v11 |
| ; GCN-NEXT: v_or_b32_e32 v47, v35, v12 |
| ; GCN-NEXT: v_or_b32_e32 v45, v7, v14 |
| ; GCN-NEXT: v_or_b32_e32 v41, v36, v20 |
| ; GCN-NEXT: v_or_b32_e32 v53, v9, v23 |
| ; GCN-NEXT: v_or_b32_e32 v48, v37, v25 |
| ; GCN-NEXT: v_or_b32_e32 v39, v13, v26 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v33, v16, v29 |
| ; GCN-NEXT: v_or_b32_e32 v30, v18, v30 |
| ; GCN-NEXT: v_or_b32_e32 v25, v15, v38 |
| ; GCN-NEXT: v_or_b32_e32 v23, v22, v49 |
| ; GCN-NEXT: v_or_b32_e32 v20, v17, v24 |
| ; GCN-NEXT: v_or_b32_e32 v14, v27, v50 |
| ; GCN-NEXT: v_or_b32_e32 v12, v21, v28 |
| ; GCN-NEXT: v_or_b32_e32 v10, v19, v51 |
| ; GCN-NEXT: v_alignbit_b32 v42, v56, v57, 24 |
| ; GCN-NEXT: v_alignbit_b32 v43, v56, v57, 16 |
| ; GCN-NEXT: v_alignbit_b32 v61, v56, v57, 8 |
| ; GCN-NEXT: v_alignbit_b32 v55, v45, v47, 24 |
| ; GCN-NEXT: v_alignbit_b32 v40, v45, v47, 16 |
| ; GCN-NEXT: v_alignbit_b32 v60, v45, v47, 8 |
| ; GCN-NEXT: v_alignbit_b32 v7, v53, v41, 24 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v53, v41, 16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v59, v53, v41, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v39, v48, 24 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v39, v48, 16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v46, v39, v48, 8 |
| ; GCN-NEXT: v_alignbit_b32 v29, v30, v33, 24 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v30, v33, 16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v26, v30, v33, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v23, v25, 24 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v23, v25, 16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v23, v25, 8 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v14, v20, 24 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v14, v20, 16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v14, v20, 8 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v10, v12, 24 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v10, v12, 16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v10, v12, 8 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 8, v56 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v45 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v62, 8, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v58, 8, v39 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v30 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v23 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v14 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v10 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v7, v1, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: .LBB52_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB52_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v10, v28 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v12, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v14, v24 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v27 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v20, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v22 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v22, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v16, v16 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v4, v4 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v51 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v24, v37 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v26, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v6, v6 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v49 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v28, v35 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v8, v8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v30, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v31 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v32, v32 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x38000000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x38000000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x38000000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x38000000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v26, 0x38000000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v27, 0x38000000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v28, 0x38000000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v29, 0x38000000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v30, 0x38000000, v30 |
| ; GCN-NEXT: v_add_f32_e32 v31, 0x38000000, v31 |
| ; GCN-NEXT: v_add_f32_e32 v32, 0x38000000, v32 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v33, v12 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v12, v14 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v14, v17 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v17, v19 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v19, v20 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v20, v22 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v22, v23 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v24, v24 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v23, v25 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v26, v26 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v25, v27 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v27, v28 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v28, v29 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v29, v30 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v31, v31 |
| ; GCN-NEXT: v_cvt_f16_f32_e32 v32, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v3 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v37, 16, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v31 |
| ; GCN-NEXT: v_bfe_u32 v12, v31, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v12, v8, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v12, v6, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v12, v5, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v12, v4, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v12, v3, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v12, v2, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v11, v10 |
| ; GCN-NEXT: v_or_b32_e32 v10, v33, v30 |
| ; GCN-NEXT: v_or_b32_e32 v20, v14, v34 |
| ; GCN-NEXT: v_or_b32_e32 v14, v17, v35 |
| ; GCN-NEXT: v_or_b32_e32 v25, v15, v19 |
| ; GCN-NEXT: v_or_b32_e32 v23, v21, v36 |
| ; GCN-NEXT: v_or_b32_e32 v33, v16, v37 |
| ; GCN-NEXT: v_or_b32_e32 v30, v18, v38 |
| ; GCN-NEXT: v_or_b32_e32 v48, v24, v22 |
| ; GCN-NEXT: v_or_b32_e32 v39, v13, v39 |
| ; GCN-NEXT: v_or_b32_e32 v41, v26, v49 |
| ; GCN-NEXT: v_or_b32_e32 v53, v9, v50 |
| ; GCN-NEXT: v_or_b32_e32 v47, v27, v51 |
| ; GCN-NEXT: v_or_b32_e32 v45, v7, v52 |
| ; GCN-NEXT: v_or_b32_e32 v57, v29, v28 |
| ; GCN-NEXT: v_or_b32_e32 v56, v32, v54 |
| ; GCN-NEXT: v_alignbit_b32 v42, v56, v57, 24 |
| ; GCN-NEXT: v_alignbit_b32 v43, v56, v57, 16 |
| ; GCN-NEXT: v_alignbit_b32 v61, v56, v57, 8 |
| ; GCN-NEXT: v_alignbit_b32 v55, v45, v47, 24 |
| ; GCN-NEXT: v_alignbit_b32 v40, v45, v47, 16 |
| ; GCN-NEXT: v_alignbit_b32 v60, v45, v47, 8 |
| ; GCN-NEXT: v_alignbit_b32 v7, v53, v41, 24 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v53, v41, 16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v59, v53, v41, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v39, v48, 24 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v39, v48, 16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v46, v39, v48, 8 |
| ; GCN-NEXT: v_alignbit_b32 v29, v30, v33, 24 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v30, v33, 16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v26, v30, v33, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v23, v25, 24 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v23, v25, 16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v23, v25, 8 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v14, v20, 24 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v14, v20, 16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v14, v20, 8 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v10, v12, 24 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v10, v12, 16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v7, v10, v12, 8 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 8, v56 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v45 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v62, 8, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v58, 8, v39 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v44, 8, v30 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v23 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v14 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v10 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_bfe_u32 v7, v1, 8, 8 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill |
| ; GCN-NEXT: .LBB52_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v61 |
| ; GCN-NEXT: v_or_b32_e32 v13, v7, v9 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v11 |
| ; GCN-NEXT: v_or_b32_e32 v15, v7, v9 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v60 |
| ; GCN-NEXT: v_or_b32_e32 v16, v7, v9 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v63 |
| ; GCN-NEXT: v_or_b32_e32 v17, v7, v9 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v59 |
| ; GCN-NEXT: v_or_b32_e32 v19, v7, v9 |
| ; GCN-NEXT: v_and_b32_e32 v37, 0xff, v43 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v38, 24, v42 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v31 |
| ; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v34, 24, v7 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v62 |
| ; GCN-NEXT: v_or_b32_e32 v21, v7, v9 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 4, v0 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v40 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v36, 24, v55 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v48 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v46 |
| ; GCN-NEXT: v_or_b32_e32 v22, v9, v11 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 8, v0 |
| ; GCN-NEXT: v_and_b32_e32 v48, 0xff, v8 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 24, v8 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xff, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v58 |
| ; GCN-NEXT: v_or_b32_e32 v24, v8, v11 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 12, v0 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v39, 0xff, v11 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 24, v11 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 8, v26 |
| ; GCN-NEXT: v_or_b32_e32 v26, v11, v18 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 16, v0 |
| ; GCN-NEXT: v_and_b32_e32 v51, 0xff, v6 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v31, 24, v6 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v30 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 8, v44 |
| ; GCN-NEXT: v_or_b32_e32 v28, v6, v18 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 20, v0 |
| ; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v54, 0xff, v18 |
| ; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 24, v18 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v25 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 8, v25 |
| ; GCN-NEXT: v_or_b32_e32 v25, v18, v25 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 24, v0 |
| ; GCN-NEXT: v_and_b32_e32 v42, 0xff, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 24, v5 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v23 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v23, 8, v23 |
| ; GCN-NEXT: v_or_b32_e32 v23, v5, v23 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0 |
| ; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v44, 0xff, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 24, v29 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 8, v29 |
| ; GCN-NEXT: v_or_b32_e32 v29, v20, v29 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 32, v0 |
| ; GCN-NEXT: v_and_b32_e32 v46, 0xff, v4 |
| ; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 24, v4 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v14 |
| ; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v14 |
| ; GCN-NEXT: v_or_b32_e32 v14, v4, v14 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 36, v0 |
| ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v47, 0xff, v53 |
| ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 24, v53 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v53 |
| ; GCN-NEXT: v_or_b32_e32 v53, v12, v53 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 40, v0 |
| ; GCN-NEXT: v_and_b32_e32 v56, 0xff, v3 |
| ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v40, 24, v3 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v10 |
| ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 8, v10 |
| ; GCN-NEXT: v_or_b32_e32 v41, v3, v10 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 44, v0 |
| ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v57, 0xff, v10 |
| ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v43, 24, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v37 |
| ; GCN-NEXT: v_or_b32_e32 v37, v38, v10 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 48, v0 |
| ; GCN-NEXT: v_and_b32_e32 v58, 0xff, v2 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v38, 24, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v32 |
| ; GCN-NEXT: v_or_b32_e32 v34, v34, v2 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 52, v0 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v59, 0xff, v32 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 24, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v35 |
| ; GCN-NEXT: v_or_b32_e32 v35, v36, v32 |
| ; GCN-NEXT: v_add_i32_e32 v32, vcc, 56, v0 |
| ; GCN-NEXT: v_and_b32_e32 v36, 0xff, v1 |
| ; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v60, 24, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v48 |
| ; GCN-NEXT: v_or_b32_e32 v48, v49, v1 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 60, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v42 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v44 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v46 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v58 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v59 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GCN-NEXT: v_or_b32_e32 v27, v27, v39 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff, v21 |
| ; GCN-NEXT: v_or_b32_e32 v31, v31, v49 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22 |
| ; GCN-NEXT: v_or_b32_e32 v30, v30, v51 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24 |
| ; GCN-NEXT: v_or_b32_e32 v33, v33, v54 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26 |
| ; GCN-NEXT: v_or_b32_e32 v39, v50, v42 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28 |
| ; GCN-NEXT: v_or_b32_e32 v49, v52, v44 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25 |
| ; GCN-NEXT: v_or_b32_e32 v50, v55, v46 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23 |
| ; GCN-NEXT: v_or_b32_e32 v51, v40, v47 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_or_b32_e32 v52, v43, v56 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GCN-NEXT: v_or_b32_e32 v38, v38, v57 |
| ; GCN-NEXT: v_and_b32_e32 v53, 0xffff, v53 |
| ; GCN-NEXT: v_or_b32_e32 v54, v45, v58 |
| ; GCN-NEXT: v_and_b32_e32 v55, 0xffff, v41 |
| ; GCN-NEXT: v_or_b32_e32 v36, v60, v36 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v37 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v34 |
| ; GCN-NEXT: v_or_b32_e32 v16, v16, v35 |
| ; GCN-NEXT: v_or_b32_e32 v17, v17, v48 |
| ; GCN-NEXT: v_or_b32_e32 v19, v19, v27 |
| ; GCN-NEXT: v_or_b32_e32 v21, v21, v31 |
| ; GCN-NEXT: v_or_b32_e32 v22, v22, v30 |
| ; GCN-NEXT: v_or_b32_e32 v24, v24, v33 |
| ; GCN-NEXT: v_or_b32_e32 v26, v26, v39 |
| ; GCN-NEXT: v_or_b32_e32 v27, v28, v49 |
| ; GCN-NEXT: v_or_b32_e32 v25, v25, v50 |
| ; GCN-NEXT: v_or_b32_e32 v23, v23, v51 |
| ; GCN-NEXT: v_or_b32_e32 v28, v29, v52 |
| ; GCN-NEXT: v_or_b32_e32 v14, v14, v38 |
| ; GCN-NEXT: v_or_b32_e32 v29, v53, v54 |
| ; GCN-NEXT: v_or_b32_e32 v30, v55, v36 |
| ; GCN-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v15, v7, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v16, v9, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v17, v8, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v19, v11, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v21, v6, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v22, v18, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v24, v5, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v26, v20, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v27, v4, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v25, v12, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v23, v3, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v28, v10, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v14, v2, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v29, v32, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v30, v1, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32f16_to_v64i8: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v1 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr23 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB52_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v15 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v14 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_lshrrev_b64 v[23:24], 24, v[7:8] |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4] |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 24, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 24, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v46, 24, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 8, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v44, 8, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v59, 24, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v47, 8, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v61, 24, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v58, 8, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v60, 8, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v54, 24, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v62, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v63, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v40, 24, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 8, v2 |
| ; VI-NEXT: v_lshrrev_b64 v[26:27], 24, v[5:6] |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 8, v1 |
| ; VI-NEXT: .LBB52_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB52_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v17, 0x200 |
| ; VI-NEXT: v_add_f16_sdwa v51, v2, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v51 |
| ; VI-NEXT: v_add_f16_e32 v2, 0x200, v2 |
| ; VI-NEXT: v_add_f16_sdwa v53, v1, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v29, v2, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v53 |
| ; VI-NEXT: v_add_f16_e32 v1, 0x200, v1 |
| ; VI-NEXT: v_add_f16_sdwa v49, v4, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v28, v1, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v49 |
| ; VI-NEXT: v_add_f16_e32 v4, 0x200, v4 |
| ; VI-NEXT: v_add_f16_sdwa v52, v3, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v55, v4, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v52 |
| ; VI-NEXT: v_add_f16_e32 v3, 0x200, v3 |
| ; VI-NEXT: v_add_f16_sdwa v38, v6, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v54, v3, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v38 |
| ; VI-NEXT: v_add_f16_e32 v6, 0x200, v6 |
| ; VI-NEXT: v_add_f16_sdwa v50, v5, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v26, v6, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v50 |
| ; VI-NEXT: v_add_f16_e32 v5, 0x200, v5 |
| ; VI-NEXT: v_add_f16_sdwa v36, v8, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v25, v5, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v36 |
| ; VI-NEXT: v_add_f16_e32 v8, 0x200, v8 |
| ; VI-NEXT: v_add_f16_sdwa v48, v7, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v23, v8, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v48 |
| ; VI-NEXT: v_add_f16_e32 v7, 0x200, v7 |
| ; VI-NEXT: v_add_f16_sdwa v34, v10, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v22, v7, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v34 |
| ; VI-NEXT: v_add_f16_e32 v10, 0x200, v10 |
| ; VI-NEXT: v_add_f16_sdwa v39, v9, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v47, v10, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v39 |
| ; VI-NEXT: v_add_f16_e32 v9, 0x200, v9 |
| ; VI-NEXT: v_add_f16_sdwa v32, v12, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v46, v9, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v32 |
| ; VI-NEXT: v_add_f16_e32 v12, 0x200, v12 |
| ; VI-NEXT: v_add_f16_sdwa v37, v11, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v21, v12, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v37 |
| ; VI-NEXT: v_add_f16_e32 v11, 0x200, v11 |
| ; VI-NEXT: v_add_f16_sdwa v31, v14, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v20, v11, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v31 |
| ; VI-NEXT: v_add_f16_e32 v14, 0x200, v14 |
| ; VI-NEXT: v_add_f16_sdwa v35, v13, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v41, v14, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v35 |
| ; VI-NEXT: v_add_f16_e32 v13, 0x200, v13 |
| ; VI-NEXT: v_add_f16_sdwa v30, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_sdwa v33, v15, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v40, v13, v18 |
| ; VI-NEXT: v_lshlrev_b32_e32 v18, 16, v30 |
| ; VI-NEXT: v_add_f16_e32 v16, 0x200, v16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v33 |
| ; VI-NEXT: v_add_f16_e32 v15, 0x200, v15 |
| ; VI-NEXT: v_or_b32_e32 v43, v16, v18 |
| ; VI-NEXT: v_or_b32_e32 v42, v15, v17 |
| ; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[42:43] |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[40:41] |
| ; VI-NEXT: v_lshrrev_b32_e32 v19, 8, v43 |
| ; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v19, 8, v42 |
| ; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v19, 8, v41 |
| ; VI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v20 |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[20:21] |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 8, v21 |
| ; VI-NEXT: v_lshrrev_b32_e32 v44, 8, v47 |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[46:47] |
| ; VI-NEXT: v_lshrrev_b32_e32 v47, 8, v23 |
| ; VI-NEXT: v_lshrrev_b64 v[23:24], 24, v[22:23] |
| ; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v22 |
| ; VI-NEXT: v_lshrrev_b32_e32 v58, 8, v26 |
| ; VI-NEXT: v_lshrrev_b32_e32 v60, 8, v25 |
| ; VI-NEXT: v_lshrrev_b64 v[26:27], 24, v[25:26] |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[54:55] |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[28:29] |
| ; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v40 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 8, v46 |
| ; VI-NEXT: v_lshrrev_b32_e32 v62, 8, v55 |
| ; VI-NEXT: v_lshrrev_b32_e32 v63, 8, v54 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 8, v29 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 8, v28 |
| ; VI-NEXT: v_bfe_u32 v28, v30, 8, 8 |
| ; VI-NEXT: v_bfe_u32 v29, v31, 8, 8 |
| ; VI-NEXT: v_bfe_u32 v46, v32, 8, 8 |
| ; VI-NEXT: v_bfe_u32 v57, v34, 8, 8 |
| ; VI-NEXT: v_bfe_u32 v59, v36, 8, 8 |
| ; VI-NEXT: v_bfe_u32 v61, v38, 8, 8 |
| ; VI-NEXT: v_bfe_u32 v54, v49, 8, 8 |
| ; VI-NEXT: v_bfe_u32 v40, v51, 8, 8 |
| ; VI-NEXT: .LBB52_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v22 |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v55 |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v24 |
| ; VI-NEXT: v_or_b32_sdwa v22, v53, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v40 |
| ; VI-NEXT: v_or_b32_sdwa v1, v51, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v63 |
| ; VI-NEXT: v_or_b32_sdwa v1, v52, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v62 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v54 |
| ; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v49, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v60 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v26 |
| ; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v50, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v58 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61 |
| ; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v56 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v23 |
| ; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v47 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v59 |
| ; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v36, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v45 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v39, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v44 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v57 |
| ; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v34, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v43 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; VI-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v37, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v42 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v46 |
| ; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v32, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; VI-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v29 |
| ; VI-NEXT: v_or_b32_sdwa v2, v31, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v28 |
| ; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 60, v0 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32f16_to_v64i8: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB52_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 24, v16 |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_mov_b32_e32 v26, v23 |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: .LBB52_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB52_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: s_movk_i32 s6, 0x200 |
| ; GFX9-NEXT: v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 8, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: .LBB52_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v29 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v23 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v23, v27, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28 |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v62 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v22 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v60 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v61, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v59 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v57 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v56 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v46 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v43 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v21 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v55 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v54 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v50 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v49 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v39 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v37 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v32 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56 |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v26 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60 |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32f16_to_v64i8: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB52_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB52_2: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB52_4 |
| ; GFX11-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB52_4: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v96, 0xff, v96 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v86, 0xff, v86 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v87 |
| ; GFX11-NEXT: v_or_b32_e32 v24, v96, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v85, 8, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v25 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v86, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v84, 8, v84 |
| ; GFX11-NEXT: v_and_b32_e32 v83, 0xff, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v82 |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v80, 8, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v71, 8, v71 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v24 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v81 |
| ; GFX11-NEXT: v_and_b32_e32 v70, 0xff, v70 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v83, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v25, v80 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v71 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v70, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v22 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v66 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v65 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v55 |
| ; GFX11-NEXT: v_lshlrev_b16 v54, 8, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v53, 8, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v23, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v25, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v52 |
| ; GFX11-NEXT: v_lshlrev_b16 v20, 8, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v50 |
| ; GFX11-NEXT: v_lshlrev_b16 v49, 8, v49 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v48, 8, v48 |
| ; GFX11-NEXT: v_and_b32_e32 v39, 0xff, v39 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v23, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v25, v49 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v48 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v39, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v38 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v37 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v35 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v34 |
| ; GFX11-NEXT: v_lshlrev_b16 v18, 8, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_lshlrev_b16 v69, 8, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v68, 0xff, v68 |
| ; GFX11-NEXT: v_lshlrev_b16 v67, 8, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v23, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v32 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v30 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v29 |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v26, 8, v26 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v69 |
| ; GFX11-NEXT: v_or_b32_e32 v67, v68, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v23, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v25, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v67 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; GFX11-NEXT: s_clause 0x3 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off |
| ; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <32 x half> %a1 to <64 x i8> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x half> %a to <64 x i8> |
| br label %end |
| |
| end: |
| %phi = phi <64 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <64 x i8> %phi |
| } |
| |
| define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v64i8_to_v32f16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:112 |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:104 |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:100 |
| ; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:96 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:92 |
| ; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:88 |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:84 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:80 |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:76 |
| ; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:72 |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:68 |
| ; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:64 |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 |
| ; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:56 |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 |
| ; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:48 |
| ; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:8 |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:132 |
| ; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:128 |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:124 |
| ; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:120 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v39 |
| ; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:40 |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:36 |
| ; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:32 |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:28 |
| ; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:24 |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 |
| ; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v1 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v3 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v5 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v7 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v9 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v11 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v13 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v15 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v17 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v19 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v21 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v23 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v25 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v27 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v29 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v38 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v37 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt vmcnt(14) expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v52 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v51 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v50 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v39 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v36 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v35 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v34 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v33 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v32 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v30 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 8, v28 |
| ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v44, 8, v26 |
| ; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:44 |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 |
| ; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:116 |
| ; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:108 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 8, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v46, 8, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 8, v48 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB53_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v2 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v2 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v4 |
| ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v6 |
| ; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v22 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v43 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v60 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v58 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v57 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v56 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v42 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v40 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xff, v63 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v62 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v61 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xff, v41 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v59 |
| ; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v25, v25, v26 |
| ; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v26, v27, v26 |
| ; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v27, v29, v27 |
| ; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v28, v31, v28 |
| ; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v29, v32, v29 |
| ; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v30, v33, v30 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v31, v34, v31 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v4, v32 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v32, v5, v32 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v6, v6, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v34, v7, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v8, v8, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v36, v9, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v10, v10, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v38, v11, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v12, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v48, v13, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v14, v14, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v50, v15, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v16, v16, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v52, v17, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v18, v18, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v54, v19, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v20, v20, v5 |
| ; GCN-NEXT: v_or_b32_e32 v40, v21, v44 |
| ; GCN-NEXT: v_or_b32_e32 v22, v22, v45 |
| ; GCN-NEXT: v_or_b32_e32 v41, v23, v46 |
| ; GCN-NEXT: v_or_b32_e32 v24, v24, v47 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v55, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v39, v2 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v35, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v26 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v51, v27 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v28 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v33, v29 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v30 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v37, v31 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v4 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v49, v32 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v6 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v53, v34 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v8 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v32, v36 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v10 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v34, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v12 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v36, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v14 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v38, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v16 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v48, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v18 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v50, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v20 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v52, v40 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v22 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v54, v41 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v24 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; kill: killed $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; kill: killed $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; kill: killed $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; kill: killed $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; kill: killed $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: .LBB53_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB53_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 3, v59 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GCN-NEXT: v_or_b32_e32 v1, v47, v1 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v26 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GCN-NEXT: v_or_b32_e32 v3, v46, v3 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v28 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GCN-NEXT: v_or_b32_e32 v5, v45, v5 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v41 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GCN-NEXT: v_or_b32_e32 v7, v44, v7 |
| ; GCN-NEXT: s_movk_i32 s6, 0x300 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v24 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 3, v61 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 3, v62 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v63 |
| ; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v40 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v42 |
| ; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v30 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v56 |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v57 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v58 |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v60 |
| ; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v43 |
| ; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v28 |
| ; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v29 |
| ; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v30 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v31 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v22 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xff, v29 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v31 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v9, v32, v9 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v11, v32, v11 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v13, v32, v13 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v15, v32, v15 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v17, v32, v17 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v19, v32, v19 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v32, v21 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v23, v32, v23 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v24, v32, v24 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v25, v32, v25 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v26, v32, v26 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v27, v32, v27 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v28, v32, v28 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v29, v32, v29 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v30, v32, v30 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v31, v32, v31 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v22, v32, v22 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v20, v32, v20 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v18, v32, v18 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v16, v32, v16 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v14, v32, v14 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v32, v12 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v10, v32, v10 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v8, v32, v8 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v6, v32, v6 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v32, v4 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v2, v32, v2 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v0, v32, v0 |
| ; GCN-NEXT: v_add_i32_e32 v40, vcc, 0x300, v1 |
| ; GCN-NEXT: v_add_i32_e32 v54, vcc, s6, v3 |
| ; GCN-NEXT: v_add_i32_e32 v41, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v52, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v42, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v50, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v43, vcc, s6, v13 |
| ; GCN-NEXT: v_add_i32_e32 v48, vcc, s6, v15 |
| ; GCN-NEXT: v_add_i32_e32 v44, vcc, s6, v17 |
| ; GCN-NEXT: v_add_i32_e32 v38, vcc, s6, v19 |
| ; GCN-NEXT: v_add_i32_e32 v21, vcc, s6, v21 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, s6, v23 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v24 |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, s6, v25 |
| ; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v26 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, s6, v27 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, s6, v28 |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, s6, v29 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v30 |
| ; GCN-NEXT: v_add_i32_e32 v27, vcc, s6, v31 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v22 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v20 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v18 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v16 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v14 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v2 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v55, v0 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v39, v4 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v3, v3 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v35, v8 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v5, v5 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v51, v12 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v7, v7 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v33, v16 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v9, v9 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v37, v20 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v11, v11 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v49, v27 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v13, v13 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v53, v26 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v32, v25 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v34, v24 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v36, v23 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v38, v38 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v23, v44 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v48, v48 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v25, v43 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v50, v50 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v27, v42 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v52, v52 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v29, v41 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v54, v54 |
| ; GCN-NEXT: v_cvt_f32_f16_e32 v31, v40 |
| ; GCN-NEXT: .LBB53_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_mov_b32_e32 v0, v55 |
| ; GCN-NEXT: v_mov_b32_e32 v2, v39 |
| ; GCN-NEXT: v_mov_b32_e32 v4, v35 |
| ; GCN-NEXT: v_mov_b32_e32 v6, v51 |
| ; GCN-NEXT: v_mov_b32_e32 v8, v33 |
| ; GCN-NEXT: v_mov_b32_e32 v10, v37 |
| ; GCN-NEXT: v_mov_b32_e32 v12, v49 |
| ; GCN-NEXT: v_mov_b32_e32 v14, v53 |
| ; GCN-NEXT: v_mov_b32_e32 v16, v32 |
| ; GCN-NEXT: v_mov_b32_e32 v18, v34 |
| ; GCN-NEXT: v_mov_b32_e32 v20, v36 |
| ; GCN-NEXT: v_mov_b32_e32 v22, v38 |
| ; GCN-NEXT: v_mov_b32_e32 v24, v48 |
| ; GCN-NEXT: s_waitcnt vmcnt(1) |
| ; GCN-NEXT: v_mov_b32_e32 v26, v50 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_mov_b32_e32 v28, v52 |
| ; GCN-NEXT: v_mov_b32_e32 v30, v54 |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v64i8_to_v32f16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:132 |
| ; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64 |
| ; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72 |
| ; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80 |
| ; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88 |
| ; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:96 |
| ; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:104 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:112 |
| ; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120 |
| ; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:128 |
| ; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:20 |
| ; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:12 |
| ; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:100 |
| ; VI-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:68 |
| ; VI-NEXT: buffer_load_ushort v55, off, s[0:3], s32 offset:36 |
| ; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:4 |
| ; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:116 |
| ; VI-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:124 |
| ; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v1 |
| ; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v3 |
| ; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v5 |
| ; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v7 |
| ; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v11 |
| ; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v13 |
| ; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v15 |
| ; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v17 |
| ; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v19 |
| ; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v23 |
| ; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v25 |
| ; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v27 |
| ; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v29 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24 |
| ; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v0 |
| ; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v2 |
| ; VI-NEXT: v_lshlrev_b16_e32 v47, 8, v4 |
| ; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v6 |
| ; VI-NEXT: v_lshlrev_b16_e32 v57, 8, v8 |
| ; VI-NEXT: v_lshlrev_b16_e32 v58, 8, v10 |
| ; VI-NEXT: v_lshlrev_b16_e32 v59, 8, v12 |
| ; VI-NEXT: v_lshlrev_b16_e32 v60, 8, v14 |
| ; VI-NEXT: v_lshlrev_b16_e32 v62, 8, v16 |
| ; VI-NEXT: v_lshlrev_b16_e32 v63, 8, v18 |
| ; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v20 |
| ; VI-NEXT: s_waitcnt vmcnt(13) |
| ; VI-NEXT: v_lshlrev_b16_e32 v24, 8, v22 |
| ; VI-NEXT: s_waitcnt vmcnt(12) |
| ; VI-NEXT: v_lshlrev_b16_e32 v18, 8, v26 |
| ; VI-NEXT: s_waitcnt vmcnt(11) |
| ; VI-NEXT: v_lshlrev_b16_e32 v26, 8, v28 |
| ; VI-NEXT: s_waitcnt vmcnt(10) |
| ; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v30 |
| ; VI-NEXT: s_waitcnt vmcnt(9) |
| ; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v31 |
| ; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:108 |
| ; VI-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:92 |
| ; VI-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:84 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:76 |
| ; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:60 |
| ; VI-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:52 |
| ; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:44 |
| ; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:28 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v32 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB53_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: v_or_b32_sdwa v9, v39, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v10, v20, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v28, v59 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v22, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v30, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v38, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr39 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: ; implicit-def: $vgpr30 |
| ; VI-NEXT: ; implicit-def: $vgpr35 |
| ; VI-NEXT: ; implicit-def: $vgpr31 |
| ; VI-NEXT: ; implicit-def: $vgpr38 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: ; implicit-def: $vgpr33 |
| ; VI-NEXT: ; implicit-def: $vgpr32 |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(7) |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr23 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v49, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v48, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v55, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v36, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v43, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v37, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v44, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v61, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr49 |
| ; VI-NEXT: ; implicit-def: $vgpr48 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr36 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr37 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr34 |
| ; VI-NEXT: .LBB53_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB53_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v38 |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v44 |
| ; VI-NEXT: v_or_b32_sdwa v0, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v14, v26, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v35 |
| ; VI-NEXT: v_mov_b32_e32 v3, 0x300 |
| ; VI-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v18, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v37 |
| ; VI-NEXT: v_or_b32_sdwa v24, v24, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v30 |
| ; VI-NEXT: v_or_b32_sdwa v0, v16, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v12, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v43 |
| ; VI-NEXT: v_or_b32_sdwa v16, v63, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v22 |
| ; VI-NEXT: v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v11, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v36 |
| ; VI-NEXT: v_or_b32_sdwa v22, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v28 |
| ; VI-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v10, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v55 |
| ; VI-NEXT: v_or_b32_sdwa v28, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v9, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v48 |
| ; VI-NEXT: v_or_b32_sdwa v20, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v39 |
| ; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v8, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v49 |
| ; VI-NEXT: v_or_b32_sdwa v30, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: v_add_u16_sdwa v13, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v61 |
| ; VI-NEXT: v_or_b32_sdwa v15, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v31 |
| ; VI-NEXT: v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v26, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v16, 0x300, v16 |
| ; VI-NEXT: v_or_b32_e32 v12, v16, v12 |
| ; VI-NEXT: v_add_u16_e32 v16, 0x300, v24 |
| ; VI-NEXT: v_add_u16_e32 v14, 0x300, v14 |
| ; VI-NEXT: v_add_u16_e32 v15, 0x300, v15 |
| ; VI-NEXT: v_or_b32_e32 v13, v16, v13 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v26 |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v18 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v7, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v29, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v6, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v25, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v5, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v21, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v17 |
| ; VI-NEXT: v_or_b32_e32 v4, v17, v4 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v21 |
| ; VI-NEXT: v_or_b32_e32 v5, v17, v5 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v25 |
| ; VI-NEXT: v_or_b32_e32 v6, v17, v6 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v29 |
| ; VI-NEXT: v_or_b32_e32 v7, v17, v7 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v30 |
| ; VI-NEXT: v_or_b32_e32 v8, v17, v8 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v20 |
| ; VI-NEXT: v_or_b32_e32 v9, v17, v9 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v28 |
| ; VI-NEXT: v_or_b32_e32 v10, v17, v10 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v22 |
| ; VI-NEXT: v_or_b32_e32 v11, v17, v11 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v19, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v23, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v2, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v27, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v1, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v31, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v3 |
| ; VI-NEXT: v_or_b32_e32 v0, v3, v0 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v31 |
| ; VI-NEXT: v_or_b32_e32 v1, v3, v1 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v27 |
| ; VI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v23 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v19 |
| ; VI-NEXT: .LBB53_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v64i8_to_v32f16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:132 |
| ; GFX9-NEXT: s_nop 0 |
| ; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64 |
| ; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72 |
| ; GFX9-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80 |
| ; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:88 |
| ; GFX9-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:96 |
| ; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:104 |
| ; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:112 |
| ; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:120 |
| ; GFX9-NEXT: buffer_load_ushort v33, off, s[0:3], s32 offset:128 |
| ; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:20 |
| ; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:12 |
| ; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:100 |
| ; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:68 |
| ; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:36 |
| ; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v1 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v3 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v5 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v7 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v9 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v11 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v13 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v15 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v17 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v19 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v21 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v23 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v25 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v27 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v29 |
| ; GFX9-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:124 |
| ; GFX9-NEXT: s_waitcnt vmcnt(24) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v22 |
| ; GFX9-NEXT: s_waitcnt vmcnt(23) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(22) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(21) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v4 |
| ; GFX9-NEXT: s_waitcnt vmcnt(20) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v57, 8, v6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(19) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v8 |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v59, 8, v10 |
| ; GFX9-NEXT: s_waitcnt vmcnt(17) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v58, 8, v12 |
| ; GFX9-NEXT: s_waitcnt vmcnt(16) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v61, 8, v14 |
| ; GFX9-NEXT: s_waitcnt vmcnt(15) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v60, 8, v16 |
| ; GFX9-NEXT: s_waitcnt vmcnt(14) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v18 |
| ; GFX9-NEXT: s_waitcnt vmcnt(13) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v20 |
| ; GFX9-NEXT: s_waitcnt vmcnt(12) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v30, 8, v24 |
| ; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:116 |
| ; GFX9-NEXT: s_waitcnt vmcnt(11) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28 |
| ; GFX9-NEXT: s_waitcnt vmcnt(10) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v31 |
| ; GFX9-NEXT: s_waitcnt vmcnt(9) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v31, 8, v32 |
| ; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:108 |
| ; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:92 |
| ; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:84 |
| ; GFX9-NEXT: buffer_load_ushort v34, off, s[0:3], s32 offset:76 |
| ; GFX9-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:60 |
| ; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:52 |
| ; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:44 |
| ; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v26, 8, v26 |
| ; GFX9-NEXT: s_waitcnt vmcnt(16) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB53_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(23) |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v39, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(9) |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v62, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v63, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v35, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v34, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v16, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v32, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v16, v38, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v0, v1, v0, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v3, v2, s6 |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v8, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6 |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6 |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6 |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v51, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v8, v9, v8, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v52, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v9, v10, v9, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v42, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v10, v11, v10, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v37, v61 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v11, v12, v11, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v44, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v12, v13, v12, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v28, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v13, v14, v13, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v47, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v14, v15, v14, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v18, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v15, v16, v15, s6 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: .LBB53_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB53_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v18 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v28 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v30, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v16 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v26, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v34 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v22, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v37 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v61, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v35 |
| ; GFX9-NEXT: v_add_u16_e32 v2, 3, v47 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v24, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v24, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v42 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v63 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v26, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v52 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v62 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v28, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v51 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v39 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v30, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v32 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v20, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v3 |
| ; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v2 |
| ; GFX9-NEXT: v_add_u16_e32 v1, 3, v38 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v1 |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v8, v30, v8, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v28, v9, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v26, v10, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v24, v11, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v22, v12, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v16, v13, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v20, v14, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v18, v15, s6 |
| ; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v31, 3, v31 |
| ; GFX9-NEXT: v_or_b32_sdwa v31, v48, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v31, 0x300, v31 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v7, v27, v7, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v6, v23, v6, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v5, v19, v5, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v17, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v4, v17, v4, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v21, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v3, v21, v3, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v55, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v25, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v2, v25, v2, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v29, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v1, v29, v1, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0 |
| ; GFX9-NEXT: v_perm_b32 v0, v31, v0, s6 |
| ; GFX9-NEXT: .LBB53_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v64i8_to_v32f16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_dual_mov_b32 v32, v14 :: v_dual_mov_b32 v31, v8 |
| ; GFX11-NEXT: v_dual_mov_b32 v34, v12 :: v_dual_mov_b32 v35, v10 |
| ; GFX11-NEXT: v_dual_mov_b32 v38, v6 :: v_dual_mov_b32 v33, v4 |
| ; GFX11-NEXT: v_dual_mov_b32 v37, v2 :: v_dual_mov_b32 v36, v0 |
| ; GFX11-NEXT: s_clause 0x1f |
| ; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:128 |
| ; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:124 |
| ; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:120 |
| ; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:116 |
| ; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:112 |
| ; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:108 |
| ; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:104 |
| ; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:100 |
| ; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:96 |
| ; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:92 |
| ; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:88 |
| ; GFX11-NEXT: scratch_load_b32 v12, off, s32 offset:132 |
| ; GFX11-NEXT: scratch_load_u16 v14, off, s32 |
| ; GFX11-NEXT: scratch_load_u16 v96, off, s32 offset:8 |
| ; GFX11-NEXT: scratch_load_u16 v100, off, s32 offset:16 |
| ; GFX11-NEXT: scratch_load_u16 v101, off, s32 offset:24 |
| ; GFX11-NEXT: scratch_load_u16 v102, off, s32 offset:32 |
| ; GFX11-NEXT: scratch_load_u16 v103, off, s32 offset:40 |
| ; GFX11-NEXT: scratch_load_u16 v113, off, s32 offset:48 |
| ; GFX11-NEXT: scratch_load_u16 v114, off, s32 offset:56 |
| ; GFX11-NEXT: scratch_load_u16 v115, off, s32 offset:64 |
| ; GFX11-NEXT: scratch_load_u16 v116, off, s32 offset:72 |
| ; GFX11-NEXT: scratch_load_u16 v117, off, s32 offset:80 |
| ; GFX11-NEXT: scratch_load_u16 v99, off, s32 offset:84 |
| ; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:76 |
| ; GFX11-NEXT: scratch_load_u16 v97, off, s32 offset:68 |
| ; GFX11-NEXT: scratch_load_u16 v82, off, s32 offset:60 |
| ; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:52 |
| ; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:44 |
| ; GFX11-NEXT: scratch_load_u16 v87, off, s32 offset:36 |
| ; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:28 |
| ; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:20 |
| ; GFX11-NEXT: s_clause 0x1 |
| ; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:12 |
| ; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:4 |
| ; GFX11-NEXT: v_lshlrev_b16 v50, 8, v1 |
| ; GFX11-NEXT: v_lshlrev_b16 v51, 8, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v49, 8, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v39, 8, v7 |
| ; GFX11-NEXT: v_lshlrev_b16 v54, 8, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v48, 8, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v55, 8, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v52, 8, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v53, 8, v17 |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v19 |
| ; GFX11-NEXT: v_lshlrev_b16 v83, 8, v21 |
| ; GFX11-NEXT: v_lshlrev_b16 v81, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v27, 8, v29 |
| ; GFX11-NEXT: s_waitcnt vmcnt(33) |
| ; GFX11-NEXT: v_lshlrev_b16 v119, 8, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(31) |
| ; GFX11-NEXT: v_lshlrev_b16 v118, 8, v2 |
| ; GFX11-NEXT: s_waitcnt vmcnt(22) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12 |
| ; GFX11-NEXT: s_waitcnt vmcnt(21) |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v14 |
| ; GFX11-NEXT: s_waitcnt vmcnt(20) |
| ; GFX11-NEXT: v_lshlrev_b16 v98, 8, v96 |
| ; GFX11-NEXT: s_waitcnt vmcnt(19) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v100 |
| ; GFX11-NEXT: s_waitcnt vmcnt(18) |
| ; GFX11-NEXT: v_lshlrev_b16 v96, 8, v101 |
| ; GFX11-NEXT: s_waitcnt vmcnt(17) |
| ; GFX11-NEXT: v_lshlrev_b16 v29, 8, v102 |
| ; GFX11-NEXT: s_waitcnt vmcnt(16) |
| ; GFX11-NEXT: v_lshlrev_b16 v112, 8, v103 |
| ; GFX11-NEXT: s_waitcnt vmcnt(15) |
| ; GFX11-NEXT: v_lshlrev_b16 v103, 8, v113 |
| ; GFX11-NEXT: s_waitcnt vmcnt(14) |
| ; GFX11-NEXT: v_lshlrev_b16 v101, 8, v114 |
| ; GFX11-NEXT: s_waitcnt vmcnt(13) |
| ; GFX11-NEXT: v_lshlrev_b16 v100, 8, v115 |
| ; GFX11-NEXT: s_waitcnt vmcnt(12) |
| ; GFX11-NEXT: v_lshlrev_b16 v114, 8, v116 |
| ; GFX11-NEXT: s_waitcnt vmcnt(11) |
| ; GFX11-NEXT: v_lshlrev_b16 v102, 8, v117 |
| ; GFX11-NEXT: v_lshlrev_b16 v117, 8, v10 |
| ; GFX11-NEXT: v_lshlrev_b16 v113, 8, v8 |
| ; GFX11-NEXT: v_lshlrev_b16 v116, 8, v6 |
| ; GFX11-NEXT: v_lshlrev_b16 v115, 8, v4 |
| ; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB53_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB53_4 |
| ; GFX11-NEXT: .LBB53_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB53_3: ; %cmp.false |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v37 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v38 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v35 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v50 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v34 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18 |
| ; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x5040100 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v3, v39 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v4, v54 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v5, v48 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v6, v55 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v32 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v49 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v52 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v81 |
| ; GFX11-NEXT: v_perm_b32 v1, v1, v2, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v2, v4, v3, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v3, v6, v5, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v4, v8, v7, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v5, v10, v9, 0x5040100 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v30 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v71 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v65 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v68 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v87 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v27 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v98 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v25 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v96 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v29 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v112 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v103 |
| ; GFX11-NEXT: v_perm_b32 v6, v7, v6, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v7, v9, v8, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v8, v11, v10, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v9, v13, v12, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v10, v15, v14, 0x5040100 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v86 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v82 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v97 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v99 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v66 |
| ; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v70 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v101 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v100 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v114 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v102 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v117 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v113 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v17, v116 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v18, v115 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v19, v118 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v20, v119 |
| ; GFX11-NEXT: v_perm_b32 v11, v12, v11, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v12, v14, v13, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v13, v16, v15, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v14, v18, v17, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v15, v20, v19, 0x5040100 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr16 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr97 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr99 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr98 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr112 |
| ; GFX11-NEXT: ; implicit-def: $vgpr103 |
| ; GFX11-NEXT: ; implicit-def: $vgpr101 |
| ; GFX11-NEXT: ; implicit-def: $vgpr100 |
| ; GFX11-NEXT: ; implicit-def: $vgpr114 |
| ; GFX11-NEXT: ; implicit-def: $vgpr102 |
| ; GFX11-NEXT: ; implicit-def: $vgpr117 |
| ; GFX11-NEXT: ; implicit-def: $vgpr113 |
| ; GFX11-NEXT: ; implicit-def: $vgpr116 |
| ; GFX11-NEXT: ; implicit-def: $vgpr115 |
| ; GFX11-NEXT: ; implicit-def: $vgpr118 |
| ; GFX11-NEXT: ; implicit-def: $vgpr119 |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB53_2 |
| ; GFX11-NEXT: .LBB53_4: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u16 v0, v70, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v67, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v69, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v66, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v5, v64, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: s_waitcnt vmcnt(8) |
| ; GFX11-NEXT: v_add_nc_u16 v6, v97, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v118, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v119, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v99, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v64, 0x300, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v116, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v115, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v113, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v66, 0x300, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v114, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v117, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v85, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v67, 0x300, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(6) |
| ; GFX11-NEXT: v_add_nc_u16 v0, v86, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v82, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: s_waitcnt vmcnt(4) |
| ; GFX11-NEXT: v_add_nc_u16 v3, v87, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v84, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v102, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v101, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v100, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v69, 0x300, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v112, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v103, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v70, 0x300, v1 |
| ; GFX11-NEXT: s_waitcnt vmcnt(2) |
| ; GFX11-NEXT: v_add_nc_u16 v1, v80, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v68, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v68, 0x300, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_add_nc_u16 v0, v71, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v65, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v28, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v96, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v29, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v98, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v25, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v27, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v0, v30, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v24, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v26, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v20, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v22, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v23, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v21, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v19, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v83, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v81, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v16, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v0, v34, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v31, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v18, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v32, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v53, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v55, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v54, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v17, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v52, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v33, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v17, v36, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v21, v37, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v22, v38, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v23, v35, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v49, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v50, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v51, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v39, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v48, v23 |
| ; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17 |
| ; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v21 |
| ; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v22 |
| ; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v23 |
| ; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v2 |
| ; GFX11-NEXT: v_perm_b32 v0, v21, v17, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v1, v22, v1, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v2, v23, v4, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v3, v24, v3, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v4, v26, v18, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v5, v16, v5, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v6, v20, v6, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v7, v19, v7, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v8, v27, v8, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v9, v25, v9, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v10, v68, v10, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v11, v70, v11, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v12, v69, v12, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v13, v67, v13, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v14, v66, v14, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v15, v64, v15, 0x5040100 |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <64 x i8> %a, splat (i8 3) |
| %a2 = bitcast <64 x i8> %a1 to <32 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <64 x i8> %a to <32 x half> |
| br label %end |
| |
| end: |
| %phi = phi <32 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x half> %phi |
| } |
| |
| define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v32bf16_to_v64i8: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:8 |
| ; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v2 |
| ; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v1 |
| ; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v4 |
| ; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v3 |
| ; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v6 |
| ; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v5 |
| ; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v8 |
| ; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7 |
| ; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v10 |
| ; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v9 |
| ; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12 |
| ; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v11 |
| ; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v14 |
| ; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v13 |
| ; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v16 |
| ; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15 |
| ; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v18 |
| ; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17 |
| ; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v20 |
| ; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19 |
| ; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v22 |
| ; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v21 |
| ; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v24 |
| ; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v23 |
| ; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v26 |
| ; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v25 |
| ; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28 |
| ; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v27 |
| ; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v30 |
| ; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v29 |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v39 |
| ; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v31 |
| ; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v33 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; kill: killed $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB54_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v51 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v35 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v52 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v37 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v54 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v40 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v50 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v42 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v18 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v43 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v55 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v51, 16, v44 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v28 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v52, 16, v46 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v29 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v35 |
| ; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 24, v37 |
| ; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 24, v12 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 24, v50 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 24, v18 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 24, v55 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 24, v28 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 24, v29 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v57, v5, v34, 16 |
| ; GCN-NEXT: v_alignbit_b32 v56, v14, v32, 16 |
| ; GCN-NEXT: v_alignbit_b32 v47, v11, v36, 16 |
| ; GCN-NEXT: v_alignbit_b32 v45, v10, v7, 16 |
| ; GCN-NEXT: v_alignbit_b32 v41, v13, v38, 16 |
| ; GCN-NEXT: v_alignbit_b32 v53, v8, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v48, v16, v49, 16 |
| ; GCN-NEXT: v_alignbit_b32 v39, v6, v15, 16 |
| ; GCN-NEXT: v_alignbit_b32 v33, v21, v17, 16 |
| ; GCN-NEXT: v_alignbit_b32 v31, v4, v19, 16 |
| ; GCN-NEXT: v_alignbit_b32 v26, v24, v20, 16 |
| ; GCN-NEXT: v_alignbit_b32 v24, v3, v22, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v51, v23, 16 |
| ; GCN-NEXT: v_alignbit_b32 v16, v2, v25, 16 |
| ; GCN-NEXT: v_alignbit_b32 v13, v52, v27, 16 |
| ; GCN-NEXT: v_alignbit_b32 v11, v1, v30, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v56, v57, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v56, v57, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v61, v56, v57, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v45, v47, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v45, v47, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v60, v45, v47, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v53, v41, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v53, v41, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v59, v53, v41, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v39, v48, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v39, v48, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v39, v48, 8 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v31, v33, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v31, v33, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v31, v33, 8 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v24, v26, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v24, v26, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v24, v26, 8 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v16, v21, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v16, v21, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v16, v21, 8 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v11, v13, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v11, v13, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v11, v13, 8 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v56 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v45 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v62, 8, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v58, 8, v39 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v31 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v24 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v11 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: .LBB54_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB54_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v27 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v46 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v30 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v29 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v23 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v44 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v25 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v28 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v20 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v43 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v22 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v55 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v42 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v49 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v40 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v50 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v38 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v54 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v36 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v52 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v37 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v34 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v51 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v32 |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v35 |
| ; GCN-NEXT: v_add_f32_e32 v33, 0x40c00000, v1 |
| ; GCN-NEXT: v_add_f32_e32 v1, 0x40c00000, v2 |
| ; GCN-NEXT: v_add_f32_e32 v34, 0x40c00000, v3 |
| ; GCN-NEXT: v_add_f32_e32 v35, 0x40c00000, v4 |
| ; GCN-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GCN-NEXT: v_add_f32_e32 v2, 0x40c00000, v6 |
| ; GCN-NEXT: v_add_f32_e32 v36, 0x40c00000, v8 |
| ; GCN-NEXT: v_add_f32_e32 v37, 0x40c00000, v10 |
| ; GCN-NEXT: v_add_f32_e32 v38, 0x40c00000, v11 |
| ; GCN-NEXT: v_add_f32_e32 v3, 0x40c00000, v13 |
| ; GCN-NEXT: v_add_f32_e32 v39, 0x40c00000, v14 |
| ; GCN-NEXT: v_add_f32_e32 v11, 0x40c00000, v16 |
| ; GCN-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; GCN-NEXT: v_add_f32_e32 v4, 0x40c00000, v20 |
| ; GCN-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 |
| ; GCN-NEXT: v_add_f32_e32 v13, 0x40c00000, v18 |
| ; GCN-NEXT: v_add_f32_e32 v18, 0x40c00000, v21 |
| ; GCN-NEXT: v_add_f32_e32 v6, 0x40c00000, v22 |
| ; GCN-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GCN-NEXT: v_add_f32_e32 v16, 0x40c00000, v23 |
| ; GCN-NEXT: v_add_f32_e32 v20, 0x40c00000, v24 |
| ; GCN-NEXT: v_add_f32_e32 v8, 0x40c00000, v25 |
| ; GCN-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GCN-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GCN-NEXT: v_add_f32_e32 v22, 0x40c00000, v26 |
| ; GCN-NEXT: v_add_f32_e32 v10, 0x40c00000, v27 |
| ; GCN-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GCN-NEXT: v_add_f32_e32 v21, 0x40c00000, v28 |
| ; GCN-NEXT: v_add_f32_e32 v23, 0x40c00000, v29 |
| ; GCN-NEXT: v_add_f32_e32 v14, 0x40c00000, v30 |
| ; GCN-NEXT: v_add_f32_e32 v25, 0x40c00000, v31 |
| ; GCN-NEXT: v_add_f32_e32 v24, 0x40c00000, v32 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v1 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v35 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v2 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v37 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v3 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v11 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v4 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v13 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v6 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v16 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v32, 16, v8 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v12 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v49, 16, v10 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v21 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v50, 16, v14 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v24 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v24, 24, v24 |
| ; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v21, 24, v21 |
| ; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 24, v12 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 24, v16 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v12, 24, v13 |
| ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 24, v11 |
| ; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 24, v37 |
| ; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v11, 24, v35 |
| ; GCN-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v13, v26, v33, 16 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v11, v1, v34, 16 |
| ; GCN-NEXT: v_alignbit_b32 v21, v27, v5, 16 |
| ; GCN-NEXT: v_alignbit_b32 v16, v2, v36, 16 |
| ; GCN-NEXT: v_alignbit_b32 v26, v28, v38, 16 |
| ; GCN-NEXT: v_alignbit_b32 v24, v3, v39, 16 |
| ; GCN-NEXT: v_alignbit_b32 v33, v29, v17, 16 |
| ; GCN-NEXT: v_alignbit_b32 v31, v4, v19, 16 |
| ; GCN-NEXT: v_alignbit_b32 v48, v30, v18, 16 |
| ; GCN-NEXT: v_alignbit_b32 v39, v6, v15, 16 |
| ; GCN-NEXT: v_alignbit_b32 v41, v32, v20, 16 |
| ; GCN-NEXT: v_alignbit_b32 v53, v8, v9, 16 |
| ; GCN-NEXT: v_alignbit_b32 v47, v49, v22, 16 |
| ; GCN-NEXT: v_alignbit_b32 v45, v10, v7, 16 |
| ; GCN-NEXT: v_alignbit_b32 v57, v50, v23, 16 |
| ; GCN-NEXT: v_alignbit_b32 v56, v14, v25, 16 |
| ; GCN-NEXT: v_alignbit_b32 v5, v56, v57, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v56, v57, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v61, v56, v57, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v45, v47, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v45, v47, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v60, v45, v47, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v53, v41, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v53, v41, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_alignbit_b32 v59, v53, v41, 8 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v39, v48, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v39, v48, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v39, v48, 8 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v31, v33, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v31, v33, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v31, v33, 8 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v24, v26, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v24, v26, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v24, v26, 8 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v16, v21, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v16, v21, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v16, v21, 8 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v11, v13, 24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v11, v13, 16 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_alignbit_b32 v5, v11, v13, 8 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v5, 8, v56 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v63, 8, v45 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v62, 8, v53 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v58, 8, v39 |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v31 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v24 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v16 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshrrev_b32_e32 v7, 8, v11 |
| ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill |
| ; GCN-NEXT: .LBB54_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v61 |
| ; GCN-NEXT: v_or_b32_e32 v12, v7, v9 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 8, v5 |
| ; GCN-NEXT: v_or_b32_e32 v15, v7, v5 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 8, v60 |
| ; GCN-NEXT: v_or_b32_e32 v17, v5, v7 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 8, v63 |
| ; GCN-NEXT: v_or_b32_e32 v18, v5, v7 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v41 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 8, v59 |
| ; GCN-NEXT: v_or_b32_e32 v19, v5, v7 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v36, 0xff, v5 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v5 |
| ; GCN-NEXT: v_and_b32_e32 v30, 0xff, v14 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 24, v5 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v53 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 8, v62 |
| ; GCN-NEXT: v_or_b32_e32 v20, v5, v7 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 4, v0 |
| ; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v7 |
| ; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v35, 24, v7 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v48 |
| ; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 8, v9 |
| ; GCN-NEXT: v_or_b32_e32 v22, v7, v9 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 8, v0 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xff, v10 |
| ; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v48, 24, v9 |
| ; GCN-NEXT: v_and_b32_e32 v9, 0xff, v39 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 8, v58 |
| ; GCN-NEXT: v_or_b32_e32 v23, v9, v10 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 12, v0 |
| ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v49, 0xff, v10 |
| ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 24, v10 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v33 |
| ; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v14 |
| ; GCN-NEXT: v_or_b32_e32 v25, v10, v14 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 16, v0 |
| ; GCN-NEXT: v_and_b32_e32 v51, 0xff, v8 |
| ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 24, v8 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xff, v31 |
| ; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v14 |
| ; GCN-NEXT: v_or_b32_e32 v28, v8, v14 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, 20, v0 |
| ; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v54, 0xff, v14 |
| ; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v31, 24, v14 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xff, v26 |
| ; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 8, v26 |
| ; GCN-NEXT: v_or_b32_e32 v26, v14, v26 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, 24, v0 |
| ; GCN-NEXT: v_and_b32_e32 v42, 0xff, v6 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 24, v6 |
| ; GCN-NEXT: v_and_b32_e32 v6, 0xff, v24 |
| ; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 8, v24 |
| ; GCN-NEXT: v_or_b32_e32 v24, v6, v24 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 28, v0 |
| ; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v44, 0xff, v33 |
| ; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 24, v33 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xff, v21 |
| ; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 8, v33 |
| ; GCN-NEXT: v_or_b32_e32 v33, v21, v33 |
| ; GCN-NEXT: v_add_i32_e32 v21, vcc, 32, v0 |
| ; GCN-NEXT: v_and_b32_e32 v46, 0xff, v4 |
| ; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 24, v4 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v16 |
| ; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 8, v16 |
| ; GCN-NEXT: v_or_b32_e32 v16, v4, v16 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 36, v0 |
| ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v47, 0xff, v53 |
| ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 24, v53 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v53, 8, v53 |
| ; GCN-NEXT: v_or_b32_e32 v53, v13, v53 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, 40, v0 |
| ; GCN-NEXT: v_and_b32_e32 v56, 0xff, v3 |
| ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v40, 24, v3 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v11 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v11 |
| ; GCN-NEXT: v_or_b32_e32 v41, v3, v11 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 44, v0 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v57, 0xff, v11 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v43, 24, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v36 |
| ; GCN-NEXT: v_or_b32_e32 v36, v37, v11 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, 48, v0 |
| ; GCN-NEXT: v_and_b32_e32 v58, 0xff, v2 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v37, 24, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v30 |
| ; GCN-NEXT: v_or_b32_e32 v32, v32, v2 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, 52, v0 |
| ; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v59, 0xff, v30 |
| ; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 24, v30 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v34 |
| ; GCN-NEXT: v_or_b32_e32 v34, v35, v30 |
| ; GCN-NEXT: v_add_i32_e32 v30, vcc, 56, v0 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v1 |
| ; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v60, 24, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v38 |
| ; GCN-NEXT: v_or_b32_e32 v38, v48, v1 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, 60, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v51 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v42 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v44 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v46 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v56 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v56, 16, v57 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v57, 16, v58 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v58, 16, v59 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GCN-NEXT: v_or_b32_e32 v27, v27, v48 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xffff, v20 |
| ; GCN-NEXT: v_or_b32_e32 v29, v29, v49 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22 |
| ; GCN-NEXT: v_or_b32_e32 v31, v31, v51 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff, v23 |
| ; GCN-NEXT: v_or_b32_e32 v39, v39, v54 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25 |
| ; GCN-NEXT: v_or_b32_e32 v48, v50, v42 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28 |
| ; GCN-NEXT: v_or_b32_e32 v49, v52, v44 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xffff, v26 |
| ; GCN-NEXT: v_or_b32_e32 v50, v55, v46 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xffff, v24 |
| ; GCN-NEXT: v_or_b32_e32 v51, v40, v47 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xffff, v33 |
| ; GCN-NEXT: v_or_b32_e32 v52, v43, v56 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GCN-NEXT: v_or_b32_e32 v37, v37, v57 |
| ; GCN-NEXT: v_and_b32_e32 v53, 0xffff, v53 |
| ; GCN-NEXT: v_or_b32_e32 v54, v45, v58 |
| ; GCN-NEXT: v_and_b32_e32 v55, 0xffff, v41 |
| ; GCN-NEXT: v_or_b32_e32 v35, v60, v35 |
| ; GCN-NEXT: v_or_b32_e32 v12, v12, v36 |
| ; GCN-NEXT: v_or_b32_e32 v15, v15, v32 |
| ; GCN-NEXT: v_or_b32_e32 v17, v17, v34 |
| ; GCN-NEXT: v_or_b32_e32 v18, v18, v38 |
| ; GCN-NEXT: v_or_b32_e32 v19, v19, v27 |
| ; GCN-NEXT: v_or_b32_e32 v20, v20, v29 |
| ; GCN-NEXT: v_or_b32_e32 v22, v22, v31 |
| ; GCN-NEXT: v_or_b32_e32 v23, v23, v39 |
| ; GCN-NEXT: v_or_b32_e32 v25, v25, v48 |
| ; GCN-NEXT: v_or_b32_e32 v27, v28, v49 |
| ; GCN-NEXT: v_or_b32_e32 v26, v26, v50 |
| ; GCN-NEXT: v_or_b32_e32 v24, v24, v51 |
| ; GCN-NEXT: v_or_b32_e32 v28, v33, v52 |
| ; GCN-NEXT: v_or_b32_e32 v16, v16, v37 |
| ; GCN-NEXT: v_or_b32_e32 v29, v53, v54 |
| ; GCN-NEXT: v_or_b32_e32 v31, v55, v35 |
| ; GCN-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v15, v5, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v17, v7, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v18, v9, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v19, v10, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v20, v8, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v22, v14, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v23, v6, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v25, v21, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v27, v4, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v26, v13, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v24, v3, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v28, v11, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v16, v2, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v29, v30, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_store_dword v31, v1, s[0:3], 0 offen |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v32bf16_to_v64i8: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr49 |
| ; VI-NEXT: ; implicit-def: $vgpr48 |
| ; VI-NEXT: ; implicit-def: $vgpr39 |
| ; VI-NEXT: ; implicit-def: $vgpr38 |
| ; VI-NEXT: ; implicit-def: $vgpr37 |
| ; VI-NEXT: ; implicit-def: $vgpr36 |
| ; VI-NEXT: ; implicit-def: $vgpr35 |
| ; VI-NEXT: ; implicit-def: $vgpr34 |
| ; VI-NEXT: ; implicit-def: $vgpr33 |
| ; VI-NEXT: ; implicit-def: $vgpr32 |
| ; VI-NEXT: ; implicit-def: $vgpr31 |
| ; VI-NEXT: ; implicit-def: $vgpr30 |
| ; VI-NEXT: ; kill: killed $vgpr17 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB54_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v16 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v17, 8, v16 |
| ; VI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8] |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 24, v16 |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; VI-NEXT: v_mov_b32_e32 v26, v22 |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4] |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2] |
| ; VI-NEXT: .LBB54_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB54_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v2 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, 0x7fff, v18 |
| ; VI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v2, 16, 1 |
| ; VI-NEXT: s_movk_i32 s6, 0x7fff |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v2 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v2 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; VI-NEXT: v_cndmask_b32_e32 v2, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 |
| ; VI-NEXT: v_alignbit_b32 v2, v2, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v1 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v1, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v1 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; VI-NEXT: v_cndmask_b32_e32 v1, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1 |
| ; VI-NEXT: v_alignbit_b32 v1, v1, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v4 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v4, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v4 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v4 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; VI-NEXT: v_cndmask_b32_e32 v4, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v4 |
| ; VI-NEXT: v_alignbit_b32 v4, v4, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v3 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v3, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v3 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v3 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; VI-NEXT: v_cndmask_b32_e32 v3, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3 |
| ; VI-NEXT: v_alignbit_b32 v3, v3, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v6 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v6, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v6 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v6 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; VI-NEXT: v_cndmask_b32_e32 v6, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v6 |
| ; VI-NEXT: v_alignbit_b32 v6, v6, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v5 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v5, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v5 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v5 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; VI-NEXT: v_cndmask_b32_e32 v5, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5 |
| ; VI-NEXT: v_alignbit_b32 v5, v5, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v8 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v8, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v8 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v8 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; VI-NEXT: v_cndmask_b32_e32 v8, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v8 |
| ; VI-NEXT: v_alignbit_b32 v8, v8, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v7 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v7, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v7 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v7 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; VI-NEXT: v_cndmask_b32_e32 v7, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v7, 16, v7 |
| ; VI-NEXT: v_alignbit_b32 v7, v7, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v10 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v10, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v10 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v10 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; VI-NEXT: v_cndmask_b32_e32 v10, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v10 |
| ; VI-NEXT: v_alignbit_b32 v10, v10, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v9 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v9, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v9 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v9 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; VI-NEXT: v_cndmask_b32_e32 v9, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v9 |
| ; VI-NEXT: v_alignbit_b32 v9, v9, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v12 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v12, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v12 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v12 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; VI-NEXT: v_cndmask_b32_e32 v12, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v12 |
| ; VI-NEXT: v_alignbit_b32 v12, v12, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v11 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v11, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v11 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v11 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; VI-NEXT: v_cndmask_b32_e32 v11, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v11 |
| ; VI-NEXT: v_alignbit_b32 v11, v11, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v14 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v14, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v14 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v14 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; VI-NEXT: v_cndmask_b32_e32 v14, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v14 |
| ; VI-NEXT: v_alignbit_b32 v14, v14, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v13 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v13, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v13 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v13 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; VI-NEXT: v_cndmask_b32_e32 v13, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v13 |
| ; VI-NEXT: v_alignbit_b32 v13, v13, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v16 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v16, 0xffff0000, v16 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v16, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v16 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v16 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; VI-NEXT: v_cndmask_b32_e32 v16, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v16 |
| ; VI-NEXT: v_alignbit_b32 v16, v16, v17, 16 |
| ; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v15 |
| ; VI-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; VI-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v17 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; VI-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; VI-NEXT: v_bfe_u32 v18, v15, 16, 1 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, v18, v15 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, s6, v18 |
| ; VI-NEXT: v_or_b32_e32 v19, 0x400000, v15 |
| ; VI-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; VI-NEXT: v_cndmask_b32_e32 v15, v18, v19, vcc |
| ; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v15 |
| ; VI-NEXT: v_alignbit_b32 v15, v15, v17, 16 |
| ; VI-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; VI-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; VI-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; VI-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; VI-NEXT: v_lshrrev_b64 v[21:22], 24, v[3:4] |
| ; VI-NEXT: v_lshrrev_b64 v[24:25], 24, v[7:8] |
| ; VI-NEXT: v_lshrrev_b64 v[22:23], 24, v[1:2] |
| ; VI-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 24, v16 |
| ; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 8, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 8, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 24, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 8, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 8, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 24, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 8, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v49, 8, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v50, 24, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v52, 8, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v55, 24, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v41, 8, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v43, 8, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v44, 24, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v46, 8, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v56, 8, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v57, 24, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v59, 8, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v61, 8, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v62, 24, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v63, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 8, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 8, v1 |
| ; VI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; VI-NEXT: .LBB54_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v29 |
| ; VI-NEXT: v_lshlrev_b16_e32 v22, 8, v22 |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v22, v27, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v28 |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v62 |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v63, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 4, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v61 |
| ; VI-NEXT: v_or_b32_sdwa v1, v60, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 8, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v59 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v57 |
| ; VI-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v58, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v56 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; VI-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v46 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v44 |
| ; VI-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 20, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v43 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v24 |
| ; VI-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 24, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v55 |
| ; VI-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 28, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v54 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; VI-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v52 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v50 |
| ; VI-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v51, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 36, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v49 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; VI-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v48, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 40, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v39 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v37 |
| ; VI-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v38, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 44, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v36 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; VI-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v35, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v34 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v32 |
| ; VI-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v33, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 52, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v31 |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; VI-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v30, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_add_u32_e32 v2, vcc, 56, v0 |
| ; VI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; VI-NEXT: v_lshlrev_b16_e32 v2, 8, v26 |
| ; VI-NEXT: v_add_u32_e32 v0, vcc, 60, v0 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_lshlrev_b16_e32 v1, 8, v1 |
| ; VI-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v32bf16_to_v64i8: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; kill: killed $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB54_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v16 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v14 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v12 |
| ; GFX9-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[3:4] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 24, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 24, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 8, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 24, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 8, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 24, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 8, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 24, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 24, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 8, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 24, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 8, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 8, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 8, v1 |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[5:6] |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[1:2] |
| ; GFX9-NEXT: .LBB54_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB54_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v2 |
| ; GFX9-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 |
| ; GFX9-NEXT: v_bfe_u32 v18, v17, 16, 1 |
| ; GFX9-NEXT: s_movk_i32 s6, 0x7fff |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX9-NEXT: v_add3_u32 v18, v18, v17, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v19, 0x400000, v17 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17 |
| ; GFX9-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v17, v18, v19, vcc |
| ; GFX9-NEXT: v_bfe_u32 v18, v2, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v18, v18, v2, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v19, 0x400000, v2 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v2, v18, v19, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v18, 16, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 |
| ; GFX9-NEXT: v_bfe_u32 v19, v18, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX9-NEXT: v_add3_u32 v19, v19, v18, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v18 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v18, v19, v20, vcc |
| ; GFX9-NEXT: v_bfe_u32 v19, v1, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v19, v19, v1, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v1 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; GFX9-NEXT: s_mov_b32 s7, 0x7060302 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v1, v19, v20, vcc |
| ; GFX9-NEXT: v_perm_b32 v27, v1, v18, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v4 |
| ; GFX9-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 |
| ; GFX9-NEXT: v_bfe_u32 v19, v1, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX9-NEXT: v_add3_u32 v19, v19, v1, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v1 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 |
| ; GFX9-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v1, v19, v20, vcc |
| ; GFX9-NEXT: v_bfe_u32 v19, v4, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v19, v19, v4, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v20, 0x400000, v4 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v4, v19, v20, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v19, 16, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 |
| ; GFX9-NEXT: v_bfe_u32 v20, v19, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX9-NEXT: v_add3_u32 v20, v20, v19, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v21, 0x400000, v19 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v19, v20, v21, vcc |
| ; GFX9-NEXT: v_bfe_u32 v20, v3, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v20, v20, v3, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v21, 0x400000, v3 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc |
| ; GFX9-NEXT: v_perm_b32 v29, v3, v19, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v6 |
| ; GFX9-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX9-NEXT: v_bfe_u32 v20, v3, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX9-NEXT: v_add3_u32 v20, v20, v3, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v21, 0x400000, v3 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 |
| ; GFX9-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc |
| ; GFX9-NEXT: v_bfe_u32 v20, v6, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v20, v20, v6, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v21, 0x400000, v6 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v6, v20, v21, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v20, 16, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 |
| ; GFX9-NEXT: v_bfe_u32 v21, v20, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX9-NEXT: v_add3_u32 v21, v21, v20, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v22, 0x400000, v20 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v20, v21, v22, vcc |
| ; GFX9-NEXT: v_bfe_u32 v21, v5, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v21, v21, v5, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v22, 0x400000, v5 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v5, v21, v22, vcc |
| ; GFX9-NEXT: v_perm_b32 v31, v5, v20, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v8 |
| ; GFX9-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 |
| ; GFX9-NEXT: v_bfe_u32 v21, v5, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX9-NEXT: v_add3_u32 v21, v21, v5, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v22, 0x400000, v5 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 |
| ; GFX9-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v5, v21, v22, vcc |
| ; GFX9-NEXT: v_bfe_u32 v21, v8, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v21, v21, v8, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v22, 0x400000, v8 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v8, v21, v22, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v21, 16, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 |
| ; GFX9-NEXT: v_bfe_u32 v22, v21, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GFX9-NEXT: v_add3_u32 v22, v22, v21, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v23, 0x400000, v21 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v21, v22, v23, vcc |
| ; GFX9-NEXT: v_bfe_u32 v22, v7, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v22, v22, v7, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v23, 0x400000, v7 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v7, v22, v23, vcc |
| ; GFX9-NEXT: v_perm_b32 v33, v7, v21, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v7, 16, v10 |
| ; GFX9-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GFX9-NEXT: v_bfe_u32 v22, v7, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX9-NEXT: v_add3_u32 v22, v22, v7, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v23, 0x400000, v7 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7 |
| ; GFX9-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v7, v22, v23, vcc |
| ; GFX9-NEXT: v_bfe_u32 v22, v10, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v22, v22, v10, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v23, 0x400000, v10 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v10, v22, v23, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v22, 16, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; GFX9-NEXT: v_bfe_u32 v23, v22, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX9-NEXT: v_add3_u32 v23, v23, v22, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v24, 0x400000, v22 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v22, v23, v24, vcc |
| ; GFX9-NEXT: v_bfe_u32 v23, v9, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v23, v23, v9, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v24, 0x400000, v9 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v9, v23, v24, vcc |
| ; GFX9-NEXT: v_perm_b32 v35, v9, v22, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v12 |
| ; GFX9-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GFX9-NEXT: v_bfe_u32 v23, v9, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX9-NEXT: v_add3_u32 v23, v23, v9, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v24, 0x400000, v9 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9 |
| ; GFX9-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v9, v23, v24, vcc |
| ; GFX9-NEXT: v_bfe_u32 v23, v12, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v23, v23, v12, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v24, 0x400000, v12 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v12, v23, v24, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v23, 16, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v23, 0x40c00000, v23 |
| ; GFX9-NEXT: v_bfe_u32 v24, v23, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; GFX9-NEXT: v_add3_u32 v24, v24, v23, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v25, 0x400000, v23 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v23, v24, v25, vcc |
| ; GFX9-NEXT: v_bfe_u32 v24, v11, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v24, v24, v11, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v25, 0x400000, v11 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v11, v24, v25, vcc |
| ; GFX9-NEXT: v_perm_b32 v37, v11, v23, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v11, 16, v14 |
| ; GFX9-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX9-NEXT: v_bfe_u32 v24, v11, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX9-NEXT: v_add3_u32 v24, v24, v11, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v25, 0x400000, v11 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11 |
| ; GFX9-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v11, v24, v25, vcc |
| ; GFX9-NEXT: v_bfe_u32 v24, v14, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v24, v24, v14, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v25, 0x400000, v14 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v14, v24, v25, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v24, 16, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 |
| ; GFX9-NEXT: v_bfe_u32 v25, v24, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX9-NEXT: v_add3_u32 v25, v25, v24, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v26, 0x400000, v24 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v24, v25, v26, vcc |
| ; GFX9-NEXT: v_bfe_u32 v25, v13, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v25, v25, v13, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v26, 0x400000, v13 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v13, v25, v26, vcc |
| ; GFX9-NEXT: v_perm_b32 v48, v13, v24, s7 |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v13, 16, v16 |
| ; GFX9-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX9-NEXT: v_bfe_u32 v25, v13, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v16, 0xffff0000, v16 |
| ; GFX9-NEXT: v_add3_u32 v25, v25, v13, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v26, 0x400000, v13 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13 |
| ; GFX9-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v13, v25, v26, vcc |
| ; GFX9-NEXT: v_bfe_u32 v25, v16, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v25, v25, v16, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v26, 0x400000, v16 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v16, v25, v26, vcc |
| ; GFX9-NEXT: v_lshlrev_b32_e32 v25, 16, v15 |
| ; GFX9-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 |
| ; GFX9-NEXT: v_bfe_u32 v26, v25, 16, 1 |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX9-NEXT: v_add3_u32 v26, v26, v25, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v39, 0x400000, v25 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25 |
| ; GFX9-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v25, v26, v39, vcc |
| ; GFX9-NEXT: v_bfe_u32 v26, v15, 16, 1 |
| ; GFX9-NEXT: v_add3_u32 v26, v26, v15, s6 |
| ; GFX9-NEXT: v_or_b32_e32 v39, 0x400000, v15 |
| ; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15 |
| ; GFX9-NEXT: v_cndmask_b32_e32 v15, v26, v39, vcc |
| ; GFX9-NEXT: v_perm_b32 v51, v16, v13, s7 |
| ; GFX9-NEXT: v_perm_b32 v50, v15, v25, s7 |
| ; GFX9-NEXT: v_perm_b32 v28, v2, v17, s7 |
| ; GFX9-NEXT: v_perm_b32 v30, v4, v1, s7 |
| ; GFX9-NEXT: v_perm_b32 v49, v14, v11, s7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v4, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b64 v[17:18], 24, v[50:51] |
| ; GFX9-NEXT: v_perm_b32 v32, v6, v3, s7 |
| ; GFX9-NEXT: v_perm_b32 v38, v12, v9, s7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v3, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b64 v[18:19], 24, v[48:49] |
| ; GFX9-NEXT: v_perm_b32 v34, v8, v5, s7 |
| ; GFX9-NEXT: v_perm_b32 v36, v10, v7, s7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v8, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v5, 16, v20 |
| ; GFX9-NEXT: v_lshrrev_b64 v[19:20], 24, v[37:38] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v7, 16, v21 |
| ; GFX9-NEXT: v_lshrrev_b64 v[20:21], 24, v[35:36] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v22 |
| ; GFX9-NEXT: v_lshrrev_b64 v[21:22], 24, v[33:34] |
| ; GFX9-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v23 |
| ; GFX9-NEXT: v_lshrrev_b64 v[22:23], 24, v[29:30] |
| ; GFX9-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v25 |
| ; GFX9-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v24 |
| ; GFX9-NEXT: v_lshrrev_b64 v[25:26], 24, v[31:32] |
| ; GFX9-NEXT: v_lshrrev_b64 v[23:24], 24, v[27:28] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v54, 24, v51 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v55, 8, v51 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v50 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v41, 8, v50 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v43, 24, v49 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v49, 8, v49 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v48 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v48, 8, v48 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 24, v38 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 8, v38 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v37 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 8, v37 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 24, v36 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 8, v36 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v35 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 8, v35 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 24, v34 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 8, v34 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v33 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 8, v33 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 24, v32 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 8, v32 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v31 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 8, v31 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 24, v30 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 8, v30 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v29 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 8, v29 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v50, 24, v28 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v51, 8, v28 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 8, v27 |
| ; GFX9-NEXT: .LBB54_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v27 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v23 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v23, v28, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v51 |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v50 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v22 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v29 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v39, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:8 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v30 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v63 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:12 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v31 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v25 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v62, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v32 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v61 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v42, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:20 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v33 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v21 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v7, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v60, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:24 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v34 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v59 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v8, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v53, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:28 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v35 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v20 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v57, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:32 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v56 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v10, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v52, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:36 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v37 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v19 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v11, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v47, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:40 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v38 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v46 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:44 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v48 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v18 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v13, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v44, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:48 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v49 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v43 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:52 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v41 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v17 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v15, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v40, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:56 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v1, 8, v55 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v2, 8, v54 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:60 |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v32bf16_to_v64i8: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: ; implicit-def: $vgpr97 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB54_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 24, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 8, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 8, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 24, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 8, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 8, v13 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 24, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 16, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 8, v11 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 24, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 8, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 24, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 8, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v7 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 24, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 8, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 16, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 8, v5 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 24, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v4 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 24, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 8, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v1 |
| ; GFX11-NEXT: .LBB54_2: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB54_4 |
| ; GFX11-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v2 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v18, 0x40c00000, v18 |
| ; GFX11-NEXT: v_add_f32_e32 v19, 0x40c00000, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v1, v17, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 |
| ; GFX11-NEXT: v_bfe_u32 v22, v18, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v24, v19, 16, 1 |
| ; GFX11-NEXT: v_add3_u32 v1, v1, v17, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v18 |
| ; GFX11-NEXT: v_add3_u32 v22, v22, v18, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v1, v1, v20 :: v_dual_add_f32 v2, 0x40c00000, v2 |
| ; GFX11-NEXT: v_add3_u32 v20, v24, v19, 0x7fff |
| ; GFX11-NEXT: v_bfe_u32 v21, v2, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v2 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v2, 0x7fff |
| ; GFX11-NEXT: v_dual_cndmask_b32 v2, v21, v23 :: v_dual_lshlrev_b32 v21, 16, v4 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v27, v2, v1, 0x7060302 |
| ; GFX11-NEXT: v_dual_add_f32 v18, 0x40c00000, v21 :: v_dual_lshlrev_b32 v21, 16, v3 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v17, v22, v17, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 |
| ; GFX11-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v18 |
| ; GFX11-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v19, v20, v23, vcc_lo |
| ; GFX11-NEXT: v_bfe_u32 v20, v18, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v4 |
| ; GFX11-NEXT: v_bfe_u32 v24, v21, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v26, v19, v17, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v19, v4, 16, 1 |
| ; GFX11-NEXT: v_add3_u32 v20, v20, v18, 0x7fff |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v81, 16, v2 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v17 |
| ; GFX11-NEXT: v_add3_u32 v19, v19, v4, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v18, v20, v22, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v21 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 24, v27 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v26 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v4, v19, v23 :: v_dual_lshlrev_b32 v23, 16, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6 |
| ; GFX11-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 |
| ; GFX11-NEXT: v_add3_u32 v19, v24, v21, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21 |
| ; GFX11-NEXT: v_perm_b32 v29, v4, v18, 0x7060302 |
| ; GFX11-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 |
| ; GFX11-NEXT: v_bfe_u32 v20, v3, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v3 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v19, v19, v22 :: v_dual_lshlrev_b32 v22, 16, v5 |
| ; GFX11-NEXT: v_add_f32_e32 v21, 0x40c00000, v23 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_add3_u32 v20, v20, v3, 0x7fff |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v68, 16, v4 |
| ; GFX11-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v21 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v18 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v3, v20, v24, vcc_lo |
| ; GFX11-NEXT: v_bfe_u32 v20, v21, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v6 |
| ; GFX11-NEXT: v_bfe_u32 v25, v22, 16, 1 |
| ; GFX11-NEXT: v_perm_b32 v28, v3, v19, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v3, v6, 16, 1 |
| ; GFX11-NEXT: v_add3_u32 v20, v20, v21, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v87, 16, v28 |
| ; GFX11-NEXT: v_add3_u32 v3, v3, v6, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v20, v20, v23 :: v_dual_add_f32 v5, 0x40c00000, v5 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 |
| ; GFX11-NEXT: v_add3_u32 v6, v25, v22, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v23, 0x400000, v22 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v3, v3, v24, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v8 |
| ; GFX11-NEXT: v_bfe_u32 v21, v5, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v5 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v5, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v22, v6, v23, vcc_lo |
| ; GFX11-NEXT: v_add_f32_e32 v6, 0x40c00000, v24 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v7 |
| ; GFX11-NEXT: v_perm_b32 v86, v3, v20, 0x7060302 |
| ; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v6 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v5, v21, v25 :: v_dual_add_f32 v8, 0x40c00000, v8 |
| ; GFX11-NEXT: v_bfe_u32 v21, v6, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v65, 16, v3 |
| ; GFX11-NEXT: v_perm_b32 v85, v5, v22, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v5, v8, 16, 1 |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v6, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v19 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v84, 16, v85 |
| ; GFX11-NEXT: v_add3_u32 v5, v5, v8, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v6, v21, v24, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v5, v5, v25, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v10 |
| ; GFX11-NEXT: v_add_f32_e32 v23, 0x40c00000, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 |
| ; GFX11-NEXT: v_perm_b32 v83, v5, v6, 0x7060302 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v5 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_bfe_u32 v30, v23, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v23 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 |
| ; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v22 |
| ; GFX11-NEXT: v_add3_u32 v8, v30, v23, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_dual_cndmask_b32 v23, v8, v24 :: v_dual_lshlrev_b32 v24, 16, v9 |
| ; GFX11-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 |
| ; GFX11-NEXT: v_add_f32_e32 v8, 0x40c00000, v25 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-NEXT: v_bfe_u32 v31, v24, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v8 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 |
| ; GFX11-NEXT: v_bfe_u32 v21, v7, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v30, 0x400000, v7 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v7, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v7, v21, v30, vcc_lo |
| ; GFX11-NEXT: v_bfe_u32 v21, v8, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 |
| ; GFX11-NEXT: v_or_b32_e32 v30, 0x400000, v10 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_perm_b32 v82, v7, v23, 0x7060302 |
| ; GFX11-NEXT: v_bfe_u32 v7, v10, 16, 1 |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v8, 0x7fff |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v71, 16, v82 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_add3_u32 v7, v7, v10, 0x7fff |
| ; GFX11-NEXT: v_cndmask_b32_e32 v8, v21, v25, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 |
| ; GFX11-NEXT: v_add3_u32 v10, v31, v24, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v24 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v80, 8, v82 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v7, v7, v30 :: v_dual_lshlrev_b32 v30, 16, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v70, v7, v8, 0x7060302 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v7 |
| ; GFX11-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v24, v10, v25, vcc_lo |
| ; GFX11-NEXT: v_add_f32_e32 v10, 0x40c00000, v30 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v7, 16, v23 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v64, 8, v70 |
| ; GFX11-NEXT: v_bfe_u32 v30, v12, 16, 1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-NEXT: v_add3_u32 v30, v30, v12, 0x7fff |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9 |
| ; GFX11-NEXT: v_add_f32_e32 v9, 0x40c00000, v9 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_bfe_u32 v21, v9, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v31, 0x400000, v9 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 |
| ; GFX11-NEXT: v_add3_u32 v21, v21, v9, 0x7fff |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v9, v21, v31, vcc_lo |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v11 |
| ; GFX11-NEXT: v_bfe_u32 v25, v10, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 |
| ; GFX11-NEXT: v_perm_b32 v69, v9, v24, 0x7060302 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v9, 0x40c00000, v21 |
| ; GFX11-NEXT: v_add3_u32 v21, v25, v10, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v10 |
| ; GFX11-NEXT: v_or_b32_e32 v31, 0x400000, v12 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11 |
| ; GFX11-NEXT: v_bfe_u32 v32, v9, 16, 1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v66, 16, v69 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v10, v21, v25, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 |
| ; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v9 |
| ; GFX11-NEXT: v_add3_u32 v21, v32, v9, 0x7fff |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v67, 8, v69 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v12, v30, v31, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v14 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14 |
| ; GFX11-NEXT: v_add_f32_e32 v11, 0x40c00000, v11 |
| ; GFX11-NEXT: v_perm_b32 v55, v12, v10, 0x7060302 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v9, v21, v25, vcc_lo |
| ; GFX11-NEXT: v_add_f32_e32 v21, 0x40c00000, v30 |
| ; GFX11-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 |
| ; GFX11-NEXT: v_bfe_u32 v31, v11, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v30, 0x400000, v11 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 |
| ; GFX11-NEXT: v_bfe_u32 v32, v21, 16, 1 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 8, v55 |
| ; GFX11-NEXT: v_add3_u32 v25, v31, v11, 0x7fff |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_cndmask_b32_e32 v11, v25, v30, vcc_lo |
| ; GFX11-NEXT: v_add_f32_e32 v25, 0x40c00000, v31 |
| ; GFX11-NEXT: v_add3_u32 v30, v32, v21, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v31, 0x400000, v21 |
| ; GFX11-NEXT: v_bfe_u32 v32, v14, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21 |
| ; GFX11-NEXT: v_bfe_u32 v33, v25, 16, 1 |
| ; GFX11-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 |
| ; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v25 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v21, v30, v31, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v30, v32, v14, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v31, 0x400000, v14 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14 |
| ; GFX11-NEXT: v_add3_u32 v32, v33, v25, 0x7fff |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v33, 16, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v16 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v14, v30, v31, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX11-NEXT: v_add_f32_e32 v30, 0x40c00000, v33 |
| ; GFX11-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 |
| ; GFX11-NEXT: v_perm_b32 v54, v11, v9, 0x7060302 |
| ; GFX11-NEXT: v_perm_b32 v50, v14, v21, 0x7060302 |
| ; GFX11-NEXT: v_dual_cndmask_b32 v25, v32, v34 :: v_dual_lshlrev_b32 v34, 16, v15 |
| ; GFX11-NEXT: v_bfe_u32 v35, v13, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v32, 0x400000, v13 |
| ; GFX11-NEXT: v_bfe_u32 v33, v30, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 |
| ; GFX11-NEXT: v_add3_u32 v31, v35, v13, 0x7fff |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v11, 16, v9 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v9, 16, v24 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 8, v50 |
| ; GFX11-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v13, v31, v32, vcc_lo |
| ; GFX11-NEXT: v_add_f32_e32 v31, 0x40c00000, v34 |
| ; GFX11-NEXT: v_add3_u32 v32, v33, v30, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v33, 0x400000, v30 |
| ; GFX11-NEXT: v_bfe_u32 v34, v16, 16, 1 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30 |
| ; GFX11-NEXT: v_bfe_u32 v35, v31, 16, 1 |
| ; GFX11-NEXT: v_bfe_u32 v30, v15, 16, 1 |
| ; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v31 |
| ; GFX11-NEXT: v_or_b32_e32 v37, 0x400000, v15 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v33, vcc_lo |
| ; GFX11-NEXT: v_add3_u32 v33, v34, v16, 0x7fff |
| ; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v16 |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 |
| ; GFX11-NEXT: v_add3_u32 v35, v35, v31, 0x7fff |
| ; GFX11-NEXT: v_add3_u32 v30, v30, v15, 0x7fff |
| ; GFX11-NEXT: v_perm_b32 v49, v13, v25, 0x7060302 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v13, 16, v25 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v16, v33, v34, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v12 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v12, 16, v10 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v8 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v8, 16, v6 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v31, v35, v36, vcc_lo |
| ; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v20 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v49 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 16, v54 |
| ; GFX11-NEXT: v_cndmask_b32_e32 v15, v30, v37, vcc_lo |
| ; GFX11-NEXT: v_perm_b32 v37, v16, v32, 0x7060302 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v30, 16, v16 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v16, 16, v32 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-NEXT: v_perm_b32 v36, v15, v31, 0x7060302 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v15, 16, v31 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 16, v14 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v14, 16, v21 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 24, v37 |
| ; GFX11-NEXT: v_lshrrev_b64 v[17:18], 24, v[36:37] |
| ; GFX11-NEXT: v_lshrrev_b64 v[18:19], 24, v[49:50] |
| ; GFX11-NEXT: v_lshrrev_b64 v[19:20], 24, v[54:55] |
| ; GFX11-NEXT: v_lshrrev_b64 v[20:21], 24, v[69:70] |
| ; GFX11-NEXT: v_lshrrev_b64 v[21:22], 24, v[82:83] |
| ; GFX11-NEXT: v_lshrrev_b64 v[22:23], 24, v[85:86] |
| ; GFX11-NEXT: v_lshrrev_b64 v[23:24], 24, v[28:29] |
| ; GFX11-NEXT: v_lshrrev_b64 v[24:25], 24, v[26:27] |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 8, v37 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v36 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 8, v36 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 24, v50 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 8, v49 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 24, v55 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 8, v54 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 24, v70 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v69, 24, v83 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v70, 8, v83 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v82, 24, v86 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v83, 8, v86 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v85, 8, v85 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v86, 24, v29 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v29, 8, v29 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v28, 8, v28 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v27, 8, v27 |
| ; GFX11-NEXT: v_lshrrev_b32_e32 v25, 8, v26 |
| ; GFX11-NEXT: .LBB54_4: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v97 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v27 |
| ; GFX11-NEXT: v_or_b32_e32 v24, v26, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v26, 0xff, v81 |
| ; GFX11-NEXT: v_lshlrev_b16 v27, 8, v96 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX11-NEXT: v_lshlrev_b16 v28, 8, v28 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v25 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v26, v27 |
| ; GFX11-NEXT: v_and_b32_e32 v81, 0xff, v87 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v24 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v29 |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v28 |
| ; GFX11-NEXT: v_lshlrev_b16 v26, 8, v86 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v24 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v68 |
| ; GFX11-NEXT: v_lshlrev_b16 v27, 8, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v84 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v6 |
| ; GFX11-NEXT: v_lshlrev_b16 v29, 8, v83 |
| ; GFX11-NEXT: v_and_b32_e32 v65, 0xff, v65 |
| ; GFX11-NEXT: v_lshlrev_b16 v68, 8, v82 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v81, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v25, v25, v26 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v27 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v28, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v29 |
| ; GFX11-NEXT: v_or_b32_e32 v26, v65, v68 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v3, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v4, v25 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v5, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v71 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v8 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v70 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v52 |
| ; GFX11-NEXT: v_lshlrev_b16 v26, 8, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v27, 8, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v23, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v25, v26 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v27 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v66 |
| ; GFX11-NEXT: v_lshlrev_b16 v20, 8, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v38 |
| ; GFX11-NEXT: v_lshlrev_b16 v26, 8, v55 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v27, 8, v54 |
| ; GFX11-NEXT: v_and_b32_e32 v28, 0xff, v53 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v23, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v25, v26 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v27 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v28, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v33 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v50 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v49 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v48 |
| ; GFX11-NEXT: v_lshlrev_b16 v18, 8, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v14 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v39 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v23, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v31 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v37 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v22, 8, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v35 |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GFX11-NEXT: v_lshlrev_b16 v24, 8, v34 |
| ; GFX11-NEXT: v_and_b32_e32 v25, 0xff, v30 |
| ; GFX11-NEXT: v_lshlrev_b16 v26, 8, v32 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v20, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v23, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v24 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v25, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v18 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v20 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; GFX11-NEXT: s_clause 0x3 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off |
| ; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 |
| ; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = fadd <32 x bfloat> %a, splat (bfloat 0xR40C0) |
| %a2 = bitcast <32 x bfloat> %a1 to <64 x i8> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <32 x bfloat> %a to <64 x i8> |
| br label %end |
| |
| end: |
| %phi = phi <64 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <64 x i8> %phi |
| } |
| |
| define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) { |
| ; GCN-LABEL: bitcast_v64i8_to_v32bf16: |
| ; GCN: ; %bb.0: |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(4) |
| ; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:96 |
| ; GCN-NEXT: s_waitcnt expcnt(2) |
| ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:80 |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:76 |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:72 |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:68 |
| ; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:64 |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:60 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:56 |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:52 |
| ; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:48 |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:44 |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:36 |
| ; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:32 |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:24 |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:20 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16 |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v5 |
| ; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v7 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v11 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v13 |
| ; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v15 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v19 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v21 |
| ; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v23 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v27 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v29 |
| ; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:132 |
| ; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:128 |
| ; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:124 |
| ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:120 |
| ; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:116 |
| ; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:112 |
| ; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:108 |
| ; GCN-NEXT: s_waitcnt vmcnt(6) |
| ; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v3 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v32 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v31 |
| ; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v30 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v26 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v25 |
| ; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(1) |
| ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v24 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v10 |
| ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:84 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 8, v3 |
| ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:92 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v17 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:104 |
| ; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v13 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 8, v11 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v3, 24, v7 |
| ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GCN-NEXT: ; implicit-def: $vgpr43 |
| ; GCN-NEXT: ; implicit-def: $vgpr35 |
| ; GCN-NEXT: ; implicit-def: $vgpr49 |
| ; GCN-NEXT: ; implicit-def: $vgpr3 |
| ; GCN-NEXT: ; implicit-def: $vgpr33 |
| ; GCN-NEXT: ; implicit-def: $vgpr5 |
| ; GCN-NEXT: ; implicit-def: $vgpr39 |
| ; GCN-NEXT: ; implicit-def: $vgpr7 |
| ; GCN-NEXT: ; implicit-def: $vgpr51 |
| ; GCN-NEXT: ; implicit-def: $vgpr55 |
| ; GCN-NEXT: ; implicit-def: $vgpr41 |
| ; GCN-NEXT: ; implicit-def: $vgpr11 |
| ; GCN-NEXT: ; implicit-def: $vgpr32 |
| ; GCN-NEXT: ; implicit-def: $vgpr13 |
| ; GCN-NEXT: ; implicit-def: $vgpr34 |
| ; GCN-NEXT: ; implicit-def: $vgpr15 |
| ; GCN-NEXT: ; implicit-def: $vgpr36 |
| ; GCN-NEXT: ; implicit-def: $vgpr37 |
| ; GCN-NEXT: ; implicit-def: $vgpr38 |
| ; GCN-NEXT: ; implicit-def: $vgpr19 |
| ; GCN-NEXT: ; implicit-def: $vgpr48 |
| ; GCN-NEXT: ; implicit-def: $vgpr21 |
| ; GCN-NEXT: ; implicit-def: $vgpr50 |
| ; GCN-NEXT: ; implicit-def: $vgpr23 |
| ; GCN-NEXT: ; implicit-def: $vgpr52 |
| ; GCN-NEXT: ; implicit-def: $vgpr53 |
| ; GCN-NEXT: ; implicit-def: $vgpr54 |
| ; GCN-NEXT: ; implicit-def: $vgpr27 |
| ; GCN-NEXT: ; implicit-def: $vgpr40 |
| ; GCN-NEXT: ; implicit-def: $vgpr29 |
| ; GCN-NEXT: ; implicit-def: $vgpr42 |
| ; GCN-NEXT: ; implicit-def: $vgpr31 |
| ; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB55_2 |
| ; GCN-NEXT: ; %bb.1: ; %cmp.false |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v4 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v2 |
| ; GCN-NEXT: v_and_b32_e32 v4, 0xff, v12 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v2, v4, v2 |
| ; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v20 |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v11, v5, v2 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v28 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v5, v2 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v5, 24, v1 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xff, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v0 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 24, v9 |
| ; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v1 |
| ; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v1 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 24, v1 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xff, v2 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v2 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xff, v2 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 24, v2 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xff, v2 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_and_b32_e32 v32, 0xff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v33, 0xff, v62 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 24, v61 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v60 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v56 |
| ; GCN-NEXT: v_and_b32_e32 v36, 0xff, v44 |
| ; GCN-NEXT: v_and_b32_e32 v37, 0xff, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 24, v45 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xff, v63 |
| ; GCN-NEXT: v_and_b32_e32 v39, 0xff, v46 |
| ; GCN-NEXT: v_and_b32_e32 v48, 0xff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v49, 0xff, v59 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 24, v58 |
| ; GCN-NEXT: v_and_b32_e32 v50, 0xff, v57 |
| ; GCN-NEXT: v_and_b32_e32 v51, 0xff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v10, 24, v17 |
| ; GCN-NEXT: v_and_b32_e32 v52, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v53, 0xff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v54, 0xff, v18 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v55, 16, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v15 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 16, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v19, 16, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v28 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v31 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v32 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v33 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v34 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v23, v35, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v36 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v29, 16, v38 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v31, v39, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v48 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v44, 16, v49 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v45, 16, v50 |
| ; GCN-NEXT: v_or_b32_e32 v26, v51, v26 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v46, 16, v52 |
| ; GCN-NEXT: v_or_b32_e32 v30, v53, v30 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v47, 16, v54 |
| ; GCN-NEXT: v_or_b32_e32 v43, v5, v55 |
| ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v35, v5, v40 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v3 |
| ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v3, v3, v41 |
| ; GCN-NEXT: v_or_b32_e32 v33, v0, v21 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v5, v0, v22 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v0 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v7, v0, v7 |
| ; GCN-NEXT: v_or_b32_e32 v51, v1, v19 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v55, v0, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v11 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v11, v0, v13 |
| ; GCN-NEXT: v_or_b32_e32 v32, v4, v15 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v13, v0, v18 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v12 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v15, v0, v14 |
| ; GCN-NEXT: v_or_b32_e32 v36, v8, v16 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v37, v0, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v23 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v19, v0, v27 |
| ; GCN-NEXT: v_or_b32_e32 v48, v9, v28 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v0, v29 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v31 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v23, v0, v42 |
| ; GCN-NEXT: v_or_b32_e32 v52, v2, v44 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v53, v0, v45 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v26 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v27, v0, v25 |
| ; GCN-NEXT: v_or_b32_e32 v40, v10, v24 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v29, v0, v46 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v30 |
| ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v31, v0, v47 |
| ; GCN-NEXT: ; implicit-def: $vgpr0 |
| ; GCN-NEXT: ; implicit-def: $vgpr1 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr4 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr8 |
| ; GCN-NEXT: ; implicit-def: $vgpr9 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr12 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr16 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr20 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr28 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr62 |
| ; GCN-NEXT: ; implicit-def: $vgpr61 |
| ; GCN-NEXT: ; implicit-def: $vgpr60 |
| ; GCN-NEXT: ; implicit-def: $vgpr56 |
| ; GCN-NEXT: ; implicit-def: $vgpr44 |
| ; GCN-NEXT: ; implicit-def: $vgpr47 |
| ; GCN-NEXT: ; implicit-def: $vgpr45 |
| ; GCN-NEXT: ; implicit-def: $vgpr63 |
| ; GCN-NEXT: ; implicit-def: $vgpr46 |
| ; GCN-NEXT: ; implicit-def: $vgpr6 |
| ; GCN-NEXT: ; implicit-def: $vgpr59 |
| ; GCN-NEXT: ; implicit-def: $vgpr58 |
| ; GCN-NEXT: ; implicit-def: $vgpr57 |
| ; GCN-NEXT: ; implicit-def: $vgpr10 |
| ; GCN-NEXT: ; implicit-def: $vgpr25 |
| ; GCN-NEXT: ; implicit-def: $vgpr24 |
| ; GCN-NEXT: ; implicit-def: $vgpr17 |
| ; GCN-NEXT: ; implicit-def: $vgpr22 |
| ; GCN-NEXT: ; implicit-def: $vgpr14 |
| ; GCN-NEXT: ; implicit-def: $vgpr18 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; kill: killed $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr26 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; kill: killed $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; kill: killed $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr30 |
| ; GCN-NEXT: ; implicit-def: $vgpr2 |
| ; GCN-NEXT: ; kill: killed $vgpr2 |
| ; GCN-NEXT: .LBB55_2: ; %Flow |
| ; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GCN-NEXT: s_cbranch_execz .LBB55_4 |
| ; GCN-NEXT: ; %bb.3: ; %cmp.true |
| ; GCN-NEXT: s_waitcnt expcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 3, v14 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GCN-NEXT: v_or_b32_e32 v3, v30, v3 |
| ; GCN-NEXT: s_waitcnt vmcnt(2) |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, 3, v24 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v7, 8, v17 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xff, v5 |
| ; GCN-NEXT: v_or_b32_e32 v5, v7, v5 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, 3, v10 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xff, v7 |
| ; GCN-NEXT: v_or_b32_e32 v7, v26, v7 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, 3, v59 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v58 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xff, v10 |
| ; GCN-NEXT: v_or_b32_e32 v10, v11, v10 |
| ; GCN-NEXT: s_movk_i32 s7, 0x300 |
| ; GCN-NEXT: v_add_i32_e32 v34, vcc, 3, v18 |
| ; GCN-NEXT: s_mov_b32 s6, 0x3000000 |
| ; GCN-NEXT: v_add_i32_e32 v35, vcc, 3, v22 |
| ; GCN-NEXT: v_add_i32_e32 v36, vcc, 3, v25 |
| ; GCN-NEXT: v_add_i32_e32 v38, vcc, 3, v57 |
| ; GCN-NEXT: v_add_i32_e32 v15, vcc, 3, v46 |
| ; GCN-NEXT: v_add_i32_e32 v17, vcc, 3, v6 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, 3, v47 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v14, 8, v45 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, 3, v63 |
| ; GCN-NEXT: v_add_i32_e32 v21, vcc, 3, v56 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, 3, v44 |
| ; GCN-NEXT: v_add_i32_e32 v23, vcc, 3, v62 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v13, 8, v61 |
| ; GCN-NEXT: v_add_i32_e32 v24, vcc, 3, v60 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, 3, v28 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v26, vcc, 3, v2 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v27, vcc, 3, v2 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v11, 8, v2 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, 3, v2 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, 3, v20 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v29, vcc, 3, v2 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_lshlrev_b32_e32 v2, 8, v2 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v30, vcc, 3, v6 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v31, vcc, 3, v6 |
| ; GCN-NEXT: v_add_i32_e32 v32, vcc, 3, v8 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v8, 8, v9 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, 3, v6 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; GCN-NEXT: v_add_i32_e32 v33, vcc, 3, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v0, 8, v1 |
| ; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_add_i32_e32 v37, vcc, 3, v1 |
| ; GCN-NEXT: v_and_b32_e32 v1, 0xff, v34 |
| ; GCN-NEXT: v_and_b32_e32 v34, 0xff, v35 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xff, v36 |
| ; GCN-NEXT: v_and_b32_e32 v36, 0xff, v38 |
| ; GCN-NEXT: v_and_b32_e32 v38, 0xff, v15 |
| ; GCN-NEXT: v_and_b32_e32 v39, 0xff, v17 |
| ; GCN-NEXT: v_and_b32_e32 v48, 0xff, v18 |
| ; GCN-NEXT: v_and_b32_e32 v49, 0xff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v50, 0xff, v21 |
| ; GCN-NEXT: v_and_b32_e32 v51, 0xff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v52, 0xff, v23 |
| ; GCN-NEXT: v_and_b32_e32 v53, 0xff, v24 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v26, 0xff, v26 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xff, v27 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v54, 0xff, v20 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xff, v29 |
| ; GCN-NEXT: v_and_b32_e32 v16, 0xff, v16 |
| ; GCN-NEXT: v_and_b32_e32 v17, 0xff, v30 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v18, 0xff, v31 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xff, v32 |
| ; GCN-NEXT: v_and_b32_e32 v20, 0xff, v9 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xff, v4 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xff, v6 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xff, v33 |
| ; GCN-NEXT: v_and_b32_e32 v24, 0xff, v37 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v34 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v35 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v36 |
| ; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v29, v29, v38 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v39 |
| ; GCN-NEXT: v_or_b32_e32 v14, v14, v48 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v31, 16, v49 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v32, v32, v50 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v51 |
| ; GCN-NEXT: v_or_b32_e32 v13, v13, v52 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v53 |
| ; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v25, v35, v25 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; GCN-NEXT: v_or_b32_e32 v11, v11, v27 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v27, 16, v28 |
| ; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v28, v28, v54 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; GCN-NEXT: v_or_b32_e32 v2, v2, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v17 |
| ; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v12, v17, v12 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v17, 16, v18 |
| ; GCN-NEXT: v_or_b32_e32 v8, v8, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v20 |
| ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v19, v19, v21 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v22 |
| ; GCN-NEXT: v_or_b32_e32 v0, v0, v23 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v21, 16, v24 |
| ; GCN-NEXT: v_add_i32_e32 v3, vcc, 0x300, v3 |
| ; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v9, v22, v9 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, s7, v5 |
| ; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v6, v22, v6 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, s7, v7 |
| ; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v4, v22, v4 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, s7, v10 |
| ; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v1, v22, v1 |
| ; GCN-NEXT: v_add_i32_e32 v22, vcc, s7, v29 |
| ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v23, v23, v30 |
| ; GCN-NEXT: v_add_i32_e32 v14, vcc, s7, v14 |
| ; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v24, v24, v31 |
| ; GCN-NEXT: v_add_i32_e32 v29, vcc, s7, v32 |
| ; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v30, v30, v33 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, s7, v13 |
| ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v31, v31, v34 |
| ; GCN-NEXT: v_add_i32_e32 v25, vcc, s7, v25 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v26, v32, v26 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s7, v11 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v27, v32, v27 |
| ; GCN-NEXT: v_add_i32_e32 v28, vcc, s7, v28 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v15, v32, v15 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s7, v2 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v16, v32, v16 |
| ; GCN-NEXT: v_add_i32_e32 v12, vcc, s7, v12 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v17, v32, v17 |
| ; GCN-NEXT: v_add_i32_e32 v8, vcc, s7, v8 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v18, v32, v18 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, s7, v19 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v20, v32, v20 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, s7, v0 |
| ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) |
| ; GCN-NEXT: v_or_b32_e32 v21, v32, v21 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GCN-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GCN-NEXT: v_and_b32_e32 v22, 0xffff, v22 |
| ; GCN-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff, v29 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GCN-NEXT: v_and_b32_e32 v25, 0xffff, v25 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GCN-NEXT: v_and_b32_e32 v28, 0xffff, v28 |
| ; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GCN-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GCN-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GCN-NEXT: v_or_b32_e32 v3, v9, v3 |
| ; GCN-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; GCN-NEXT: v_or_b32_e32 v4, v4, v7 |
| ; GCN-NEXT: v_or_b32_e32 v1, v1, v10 |
| ; GCN-NEXT: v_or_b32_e32 v6, v23, v22 |
| ; GCN-NEXT: v_or_b32_e32 v7, v24, v14 |
| ; GCN-NEXT: v_or_b32_e32 v9, v30, v29 |
| ; GCN-NEXT: v_or_b32_e32 v10, v31, v13 |
| ; GCN-NEXT: v_or_b32_e32 v13, v26, v25 |
| ; GCN-NEXT: v_or_b32_e32 v11, v27, v11 |
| ; GCN-NEXT: v_or_b32_e32 v14, v15, v28 |
| ; GCN-NEXT: v_or_b32_e32 v2, v16, v2 |
| ; GCN-NEXT: v_or_b32_e32 v12, v17, v12 |
| ; GCN-NEXT: v_or_b32_e32 v8, v18, v8 |
| ; GCN-NEXT: v_or_b32_e32 v15, v20, v19 |
| ; GCN-NEXT: v_or_b32_e32 v0, v21, v0 |
| ; GCN-NEXT: v_add_i32_e32 v16, vcc, s6, v3 |
| ; GCN-NEXT: v_add_i32_e32 v17, vcc, s6, v5 |
| ; GCN-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; GCN-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; GCN-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; GCN-NEXT: v_add_i32_e32 v18, vcc, s6, v7 |
| ; GCN-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; GCN-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; GCN-NEXT: v_add_i32_e32 v19, vcc, s6, v13 |
| ; GCN-NEXT: v_add_i32_e32 v20, vcc, s6, v11 |
| ; GCN-NEXT: v_add_i32_e32 v13, vcc, s6, v14 |
| ; GCN-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; GCN-NEXT: v_add_i32_e32 v11, vcc, s6, v12 |
| ; GCN-NEXT: v_add_i32_e32 v7, vcc, s6, v8 |
| ; GCN-NEXT: v_add_i32_e32 v5, vcc, s6, v15 |
| ; GCN-NEXT: v_add_i32_e32 v0, vcc, s6, v0 |
| ; GCN-NEXT: v_and_b32_e32 v35, 0xffff0000, v0 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v43, 16, v0 |
| ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v5 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v49, 16, v5 |
| ; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v7 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v33, 16, v7 |
| ; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v11 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v39, 16, v11 |
| ; GCN-NEXT: v_and_b32_e32 v55, 0xffff0000, v2 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v51, 16, v2 |
| ; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v13 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v41, 16, v13 |
| ; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v20 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v32, 16, v20 |
| ; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v19 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v34, 16, v19 |
| ; GCN-NEXT: v_and_b32_e32 v37, 0xffff0000, v10 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v10 |
| ; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v9 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v38, 16, v9 |
| ; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v18 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v48, 16, v18 |
| ; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v6 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v50, 16, v6 |
| ; GCN-NEXT: v_and_b32_e32 v53, 0xffff0000, v1 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v52, 16, v1 |
| ; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v4 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v54, 16, v4 |
| ; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v17 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v40, 16, v17 |
| ; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v16 |
| ; GCN-NEXT: v_lshlrev_b32_e32 v42, 16, v16 |
| ; GCN-NEXT: .LBB55_4: ; %end |
| ; GCN-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GCN-NEXT: v_mov_b32_e32 v0, v43 |
| ; GCN-NEXT: v_mov_b32_e32 v1, v35 |
| ; GCN-NEXT: v_mov_b32_e32 v2, v49 |
| ; GCN-NEXT: v_mov_b32_e32 v4, v33 |
| ; GCN-NEXT: v_mov_b32_e32 v6, v39 |
| ; GCN-NEXT: v_mov_b32_e32 v8, v51 |
| ; GCN-NEXT: v_mov_b32_e32 v9, v55 |
| ; GCN-NEXT: s_waitcnt vmcnt(6) |
| ; GCN-NEXT: v_mov_b32_e32 v10, v41 |
| ; GCN-NEXT: v_mov_b32_e32 v12, v32 |
| ; GCN-NEXT: v_mov_b32_e32 v14, v34 |
| ; GCN-NEXT: v_mov_b32_e32 v16, v36 |
| ; GCN-NEXT: s_waitcnt vmcnt(3) |
| ; GCN-NEXT: v_mov_b32_e32 v17, v37 |
| ; GCN-NEXT: v_mov_b32_e32 v18, v38 |
| ; GCN-NEXT: v_mov_b32_e32 v20, v48 |
| ; GCN-NEXT: v_mov_b32_e32 v22, v50 |
| ; GCN-NEXT: s_waitcnt vmcnt(2) |
| ; GCN-NEXT: v_mov_b32_e32 v24, v52 |
| ; GCN-NEXT: v_mov_b32_e32 v25, v53 |
| ; GCN-NEXT: v_mov_b32_e32 v26, v54 |
| ; GCN-NEXT: v_mov_b32_e32 v28, v40 |
| ; GCN-NEXT: v_mov_b32_e32 v30, v42 |
| ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; GCN-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v64i8_to_v32bf16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:132 |
| ; VI-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; VI-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; VI-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; VI-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; VI-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; VI-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; VI-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; VI-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; VI-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64 |
| ; VI-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72 |
| ; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80 |
| ; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:88 |
| ; VI-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:96 |
| ; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:104 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:112 |
| ; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:120 |
| ; VI-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:128 |
| ; VI-NEXT: buffer_load_ushort v48, off, s[0:3], s32 offset:20 |
| ; VI-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:12 |
| ; VI-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:100 |
| ; VI-NEXT: buffer_load_ushort v43, off, s[0:3], s32 offset:68 |
| ; VI-NEXT: buffer_load_ushort v55, off, s[0:3], s32 offset:36 |
| ; VI-NEXT: buffer_load_ushort v49, off, s[0:3], s32 offset:4 |
| ; VI-NEXT: buffer_load_ushort v61, off, s[0:3], s32 offset:116 |
| ; VI-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:124 |
| ; VI-NEXT: v_lshlrev_b16_e32 v50, 8, v1 |
| ; VI-NEXT: v_lshlrev_b16_e32 v51, 8, v3 |
| ; VI-NEXT: v_lshlrev_b16_e32 v52, 8, v5 |
| ; VI-NEXT: v_lshlrev_b16_e32 v53, 8, v7 |
| ; VI-NEXT: v_lshlrev_b16_e32 v54, 8, v9 |
| ; VI-NEXT: v_lshlrev_b16_e32 v40, 8, v11 |
| ; VI-NEXT: v_lshlrev_b16_e32 v41, 8, v13 |
| ; VI-NEXT: v_lshlrev_b16_e32 v42, 8, v15 |
| ; VI-NEXT: v_lshlrev_b16_e32 v17, 8, v17 |
| ; VI-NEXT: v_lshlrev_b16_e32 v19, 8, v19 |
| ; VI-NEXT: v_lshlrev_b16_e32 v21, 8, v21 |
| ; VI-NEXT: v_lshlrev_b16_e32 v23, 8, v23 |
| ; VI-NEXT: v_lshlrev_b16_e32 v25, 8, v25 |
| ; VI-NEXT: v_lshlrev_b16_e32 v27, 8, v27 |
| ; VI-NEXT: v_lshlrev_b16_e32 v29, 8, v29 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24 |
| ; VI-NEXT: v_lshlrev_b16_e32 v45, 8, v0 |
| ; VI-NEXT: v_lshlrev_b16_e32 v46, 8, v2 |
| ; VI-NEXT: v_lshlrev_b16_e32 v47, 8, v4 |
| ; VI-NEXT: v_lshlrev_b16_e32 v56, 8, v6 |
| ; VI-NEXT: v_lshlrev_b16_e32 v57, 8, v8 |
| ; VI-NEXT: v_lshlrev_b16_e32 v58, 8, v10 |
| ; VI-NEXT: v_lshlrev_b16_e32 v59, 8, v12 |
| ; VI-NEXT: v_lshlrev_b16_e32 v60, 8, v14 |
| ; VI-NEXT: v_lshlrev_b16_e32 v62, 8, v16 |
| ; VI-NEXT: v_lshlrev_b16_e32 v63, 8, v18 |
| ; VI-NEXT: v_lshlrev_b16_e32 v16, 8, v20 |
| ; VI-NEXT: s_waitcnt vmcnt(13) |
| ; VI-NEXT: v_lshlrev_b16_e32 v24, 8, v22 |
| ; VI-NEXT: s_waitcnt vmcnt(12) |
| ; VI-NEXT: v_lshlrev_b16_e32 v18, 8, v26 |
| ; VI-NEXT: s_waitcnt vmcnt(11) |
| ; VI-NEXT: v_lshlrev_b16_e32 v26, 8, v28 |
| ; VI-NEXT: s_waitcnt vmcnt(10) |
| ; VI-NEXT: v_lshlrev_b16_e32 v33, 8, v30 |
| ; VI-NEXT: s_waitcnt vmcnt(9) |
| ; VI-NEXT: v_lshlrev_b16_e32 v34, 8, v31 |
| ; VI-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:108 |
| ; VI-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:92 |
| ; VI-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:84 |
| ; VI-NEXT: buffer_load_ushort v30, off, s[0:3], s32 offset:76 |
| ; VI-NEXT: buffer_load_ushort v22, off, s[0:3], s32 offset:60 |
| ; VI-NEXT: buffer_load_ushort v36, off, s[0:3], s32 offset:52 |
| ; VI-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:44 |
| ; VI-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:28 |
| ; VI-NEXT: s_waitcnt vmcnt(14) |
| ; VI-NEXT: v_lshlrev_b16_e32 v32, 8, v32 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB55_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: v_or_b32_sdwa v9, v39, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v10, v20, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v28, v59 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v22, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v30, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v38, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr39 |
| ; VI-NEXT: ; implicit-def: $vgpr20 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr22 |
| ; VI-NEXT: ; implicit-def: $vgpr30 |
| ; VI-NEXT: ; implicit-def: $vgpr35 |
| ; VI-NEXT: ; implicit-def: $vgpr31 |
| ; VI-NEXT: ; implicit-def: $vgpr38 |
| ; VI-NEXT: ; implicit-def: $vgpr47 |
| ; VI-NEXT: ; implicit-def: $vgpr57 |
| ; VI-NEXT: ; implicit-def: $vgpr59 |
| ; VI-NEXT: ; implicit-def: $vgpr62 |
| ; VI-NEXT: ; implicit-def: $vgpr18 |
| ; VI-NEXT: ; implicit-def: $vgpr33 |
| ; VI-NEXT: ; implicit-def: $vgpr32 |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(7) |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr50 |
| ; VI-NEXT: ; implicit-def: $vgpr51 |
| ; VI-NEXT: ; implicit-def: $vgpr52 |
| ; VI-NEXT: ; implicit-def: $vgpr53 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr19 |
| ; VI-NEXT: ; implicit-def: $vgpr23 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr45 |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr54 |
| ; VI-NEXT: ; implicit-def: $vgpr40 |
| ; VI-NEXT: ; implicit-def: $vgpr42 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr41 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr17 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr21 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: ; implicit-def: $vgpr25 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v49, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v48, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v55, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v36, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v43, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v37, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v44, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v61, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; kill: killed $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr49 |
| ; VI-NEXT: ; implicit-def: $vgpr48 |
| ; VI-NEXT: ; implicit-def: $vgpr55 |
| ; VI-NEXT: ; implicit-def: $vgpr36 |
| ; VI-NEXT: ; implicit-def: $vgpr43 |
| ; VI-NEXT: ; implicit-def: $vgpr37 |
| ; VI-NEXT: ; implicit-def: $vgpr44 |
| ; VI-NEXT: ; implicit-def: $vgpr61 |
| ; VI-NEXT: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr46 |
| ; VI-NEXT: ; implicit-def: $vgpr56 |
| ; VI-NEXT: ; implicit-def: $vgpr58 |
| ; VI-NEXT: ; implicit-def: $vgpr60 |
| ; VI-NEXT: ; implicit-def: $vgpr63 |
| ; VI-NEXT: ; implicit-def: $vgpr16 |
| ; VI-NEXT: ; implicit-def: $vgpr24 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; VI-NEXT: ; implicit-def: $vgpr34 |
| ; VI-NEXT: .LBB55_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB55_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: s_waitcnt vmcnt(8) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v38 |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v44 |
| ; VI-NEXT: v_or_b32_sdwa v0, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_or_b32_sdwa v14, v26, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(6) |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v35 |
| ; VI-NEXT: v_mov_b32_e32 v3, 0x300 |
| ; VI-NEXT: v_or_b32_sdwa v2, v18, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v18, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(5) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v37 |
| ; VI-NEXT: v_or_b32_sdwa v24, v24, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(4) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v30 |
| ; VI-NEXT: v_or_b32_sdwa v0, v16, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v12, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v43 |
| ; VI-NEXT: v_or_b32_sdwa v16, v63, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(3) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v22 |
| ; VI-NEXT: v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v11, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: s_waitcnt vmcnt(2) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v36 |
| ; VI-NEXT: v_or_b32_sdwa v22, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(1) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v28 |
| ; VI-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v10, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v55 |
| ; VI-NEXT: v_or_b32_sdwa v28, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v9, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v48 |
| ; VI-NEXT: v_or_b32_sdwa v20, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v39 |
| ; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v8, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v49 |
| ; VI-NEXT: v_or_b32_sdwa v30, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; VI-NEXT: v_add_u16_sdwa v13, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v61 |
| ; VI-NEXT: v_or_b32_sdwa v15, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v31 |
| ; VI-NEXT: v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v26, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_u16_e32 v16, 0x300, v16 |
| ; VI-NEXT: v_or_b32_e32 v12, v16, v12 |
| ; VI-NEXT: v_add_u16_e32 v16, 0x300, v24 |
| ; VI-NEXT: v_add_u16_e32 v14, 0x300, v14 |
| ; VI-NEXT: v_add_u16_e32 v15, 0x300, v15 |
| ; VI-NEXT: v_or_b32_e32 v13, v16, v13 |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v26 |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v18 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v7, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v29, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v6, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v25, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v5, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v21, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v17 |
| ; VI-NEXT: v_or_b32_e32 v4, v17, v4 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v21 |
| ; VI-NEXT: v_or_b32_e32 v5, v17, v5 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v25 |
| ; VI-NEXT: v_or_b32_e32 v6, v17, v6 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v29 |
| ; VI-NEXT: v_or_b32_e32 v7, v17, v7 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v30 |
| ; VI-NEXT: v_or_b32_e32 v8, v17, v8 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v20 |
| ; VI-NEXT: v_or_b32_e32 v9, v17, v9 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v28 |
| ; VI-NEXT: v_or_b32_e32 v10, v17, v10 |
| ; VI-NEXT: v_add_u16_e32 v17, 0x300, v22 |
| ; VI-NEXT: v_or_b32_e32 v11, v17, v11 |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v19, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v23, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v2, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v27, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v1, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v31, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v3 |
| ; VI-NEXT: v_or_b32_e32 v0, v3, v0 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v31 |
| ; VI-NEXT: v_or_b32_e32 v1, v3, v1 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v27 |
| ; VI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; VI-NEXT: v_add_u16_e32 v3, 0x300, v23 |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v19 |
| ; VI-NEXT: .LBB55_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v64i8_to_v32bf16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill |
| ; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:132 |
| ; GFX9-NEXT: s_nop 0 |
| ; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 |
| ; GFX9-NEXT: buffer_load_ushort v2, off, s[0:3], s32 offset:8 |
| ; GFX9-NEXT: buffer_load_ushort v4, off, s[0:3], s32 offset:16 |
| ; GFX9-NEXT: buffer_load_ushort v6, off, s[0:3], s32 offset:24 |
| ; GFX9-NEXT: buffer_load_ushort v8, off, s[0:3], s32 offset:32 |
| ; GFX9-NEXT: buffer_load_ushort v10, off, s[0:3], s32 offset:40 |
| ; GFX9-NEXT: buffer_load_ushort v12, off, s[0:3], s32 offset:48 |
| ; GFX9-NEXT: buffer_load_ushort v14, off, s[0:3], s32 offset:56 |
| ; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:64 |
| ; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:72 |
| ; GFX9-NEXT: buffer_load_ushort v20, off, s[0:3], s32 offset:80 |
| ; GFX9-NEXT: buffer_load_ushort v24, off, s[0:3], s32 offset:88 |
| ; GFX9-NEXT: buffer_load_ushort v26, off, s[0:3], s32 offset:96 |
| ; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:104 |
| ; GFX9-NEXT: buffer_load_ushort v31, off, s[0:3], s32 offset:112 |
| ; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:120 |
| ; GFX9-NEXT: buffer_load_ushort v33, off, s[0:3], s32 offset:128 |
| ; GFX9-NEXT: buffer_load_ushort v52, off, s[0:3], s32 offset:20 |
| ; GFX9-NEXT: buffer_load_ushort v39, off, s[0:3], s32 offset:12 |
| ; GFX9-NEXT: buffer_load_ushort v47, off, s[0:3], s32 offset:100 |
| ; GFX9-NEXT: buffer_load_ushort v44, off, s[0:3], s32 offset:68 |
| ; GFX9-NEXT: buffer_load_ushort v42, off, s[0:3], s32 offset:36 |
| ; GFX9-NEXT: buffer_load_ushort v51, off, s[0:3], s32 offset:4 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v49, 8, v1 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v48, 8, v3 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v53, 8, v5 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v50, 8, v7 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v55, 8, v9 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v54, 8, v11 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v41, 8, v13 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v40, 8, v15 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v43, 8, v17 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v17, 8, v19 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v21, 8, v21 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v19, 8, v23 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v25, 8, v25 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v23, 8, v27 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v29, 8, v29 |
| ; GFX9-NEXT: buffer_load_ushort v38, off, s[0:3], s32 offset:124 |
| ; GFX9-NEXT: s_waitcnt vmcnt(24) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v22 |
| ; GFX9-NEXT: s_waitcnt vmcnt(23) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v27, 8, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(22) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v46, 8, v2 |
| ; GFX9-NEXT: s_waitcnt vmcnt(21) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v45, 8, v4 |
| ; GFX9-NEXT: s_waitcnt vmcnt(20) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v57, 8, v6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(19) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v56, 8, v8 |
| ; GFX9-NEXT: s_waitcnt vmcnt(18) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v59, 8, v10 |
| ; GFX9-NEXT: s_waitcnt vmcnt(17) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v58, 8, v12 |
| ; GFX9-NEXT: s_waitcnt vmcnt(16) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v61, 8, v14 |
| ; GFX9-NEXT: s_waitcnt vmcnt(15) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v60, 8, v16 |
| ; GFX9-NEXT: s_waitcnt vmcnt(14) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v22, 8, v18 |
| ; GFX9-NEXT: s_waitcnt vmcnt(13) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v36, 8, v20 |
| ; GFX9-NEXT: s_waitcnt vmcnt(12) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v30, 8, v24 |
| ; GFX9-NEXT: buffer_load_ushort v18, off, s[0:3], s32 offset:116 |
| ; GFX9-NEXT: s_waitcnt vmcnt(11) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v24, 8, v28 |
| ; GFX9-NEXT: s_waitcnt vmcnt(10) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v20, 8, v31 |
| ; GFX9-NEXT: s_waitcnt vmcnt(9) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v31, 8, v32 |
| ; GFX9-NEXT: buffer_load_ushort v32, off, s[0:3], s32 offset:108 |
| ; GFX9-NEXT: buffer_load_ushort v16, off, s[0:3], s32 offset:92 |
| ; GFX9-NEXT: buffer_load_ushort v28, off, s[0:3], s32 offset:84 |
| ; GFX9-NEXT: buffer_load_ushort v34, off, s[0:3], s32 offset:76 |
| ; GFX9-NEXT: buffer_load_ushort v35, off, s[0:3], s32 offset:60 |
| ; GFX9-NEXT: buffer_load_ushort v37, off, s[0:3], s32 offset:52 |
| ; GFX9-NEXT: buffer_load_ushort v63, off, s[0:3], s32 offset:44 |
| ; GFX9-NEXT: buffer_load_ushort v62, off, s[0:3], s32 offset:28 |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v26, 8, v26 |
| ; GFX9-NEXT: s_waitcnt vmcnt(16) |
| ; GFX9-NEXT: v_lshlrev_b16_e32 v33, 8, v33 |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB55_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(23) |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v39, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(9) |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v62, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v63, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v35, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v34, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v16, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v32, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_or_b32_sdwa v16, v38, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr35 |
| ; GFX9-NEXT: ; implicit-def: $vgpr34 |
| ; GFX9-NEXT: ; implicit-def: $vgpr32 |
| ; GFX9-NEXT: ; implicit-def: $vgpr38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr33 |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v0, v1, v0, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v3, v2, s6 |
| ; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr49 |
| ; GFX9-NEXT: ; implicit-def: $vgpr48 |
| ; GFX9-NEXT: ; implicit-def: $vgpr53 |
| ; GFX9-NEXT: ; implicit-def: $vgpr50 |
| ; GFX9-NEXT: ; implicit-def: $vgpr40 |
| ; GFX9-NEXT: ; implicit-def: $vgpr17 |
| ; GFX9-NEXT: ; implicit-def: $vgpr19 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v8, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v2, v3, v2, s6 |
| ; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr55 |
| ; GFX9-NEXT: ; implicit-def: $vgpr54 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v3, v4, v3, s6 |
| ; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr41 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v4, v5, v4, s6 |
| ; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr43 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v5, v5, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v5, v6, v5, s6 |
| ; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v6, v6, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v6, v7, v6, s6 |
| ; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v7, v8, v7, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v8, v51, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v8, v9, v8, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v9, v52, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v9, v10, v9, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v10, v42, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v10, v11, v10, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v11, v37, v61 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v11, v12, v11, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v12, v44, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v12, v13, v12, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v13, v28, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v13, v14, v13, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v14, v47, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v14, v15, v14, s6 |
| ; GFX9-NEXT: v_or_b32_sdwa v15, v18, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD |
| ; GFX9-NEXT: v_perm_b32 v15, v16, v15, s6 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; kill: killed $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr51 |
| ; GFX9-NEXT: ; implicit-def: $vgpr52 |
| ; GFX9-NEXT: ; implicit-def: $vgpr42 |
| ; GFX9-NEXT: ; implicit-def: $vgpr37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr16 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr18 |
| ; GFX9-NEXT: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr30 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr31 |
| ; GFX9-NEXT: .LBB55_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB55_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: s_waitcnt vmcnt(8) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v18 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v15, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v28 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v30, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v13, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v16 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v26, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v16, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v44 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v12, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v34 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v22, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v37 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v61, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v11, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v35 |
| ; GFX9-NEXT: v_add_u16_e32 v2, 3, v47 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_or_b32_sdwa v2, v24, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v24, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v42 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v10, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v63 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v26, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v52 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v9, 0x300, v0 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v62 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v28, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v51 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v8, 0x300, v0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v39 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v30, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_add_u16_e32 v3, 3, v32 |
| ; GFX9-NEXT: v_or_b32_sdwa v3, v20, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v20, 0x300, v3 |
| ; GFX9-NEXT: v_add_u16_e32 v14, 0x300, v2 |
| ; GFX9-NEXT: v_add_u16_e32 v1, 3, v38 |
| ; GFX9-NEXT: v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v18, 0x300, v1 |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v8, v30, v8, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v28, v9, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v26, v10, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v24, v11, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v22, v12, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v16, v13, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v20, v14, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v18, v15, s6 |
| ; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v7, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_add_u16_e32 v31, 3, v31 |
| ; GFX9-NEXT: v_or_b32_sdwa v31, v48, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v31, 0x300, v31 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v27, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v7, v27, v7, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v6, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v23, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v23, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v6, v23, v6, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v21, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v5, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v19, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v19, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v5, v19, v5, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v4, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v17, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v17, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v4, v17, v4, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v3, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v21, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v3, v21, v3, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v55, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v2, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v25, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v2, v25, v2, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v1, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v29, 0x300, v0 |
| ; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload |
| ; GFX9-NEXT: v_perm_b32 v1, v29, v1, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; GFX9-NEXT: v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 |
| ; GFX9-NEXT: v_add_u16_e32 v0, 0x300, v0 |
| ; GFX9-NEXT: v_perm_b32 v0, v31, v0, s6 |
| ; GFX9-NEXT: .LBB55_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v64i8_to_v32bf16: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_dual_mov_b32 v32, v14 :: v_dual_mov_b32 v31, v8 |
| ; GFX11-NEXT: v_dual_mov_b32 v34, v12 :: v_dual_mov_b32 v35, v10 |
| ; GFX11-NEXT: v_dual_mov_b32 v38, v6 :: v_dual_mov_b32 v33, v4 |
| ; GFX11-NEXT: v_dual_mov_b32 v37, v2 :: v_dual_mov_b32 v36, v0 |
| ; GFX11-NEXT: s_clause 0x1f |
| ; GFX11-NEXT: scratch_load_u16 v0, off, s32 offset:128 |
| ; GFX11-NEXT: scratch_load_u16 v67, off, s32 offset:124 |
| ; GFX11-NEXT: scratch_load_u16 v2, off, s32 offset:120 |
| ; GFX11-NEXT: scratch_load_u16 v70, off, s32 offset:116 |
| ; GFX11-NEXT: scratch_load_u16 v4, off, s32 offset:112 |
| ; GFX11-NEXT: scratch_load_u16 v66, off, s32 offset:108 |
| ; GFX11-NEXT: scratch_load_u16 v6, off, s32 offset:104 |
| ; GFX11-NEXT: scratch_load_u16 v69, off, s32 offset:100 |
| ; GFX11-NEXT: scratch_load_u16 v8, off, s32 offset:96 |
| ; GFX11-NEXT: scratch_load_u16 v64, off, s32 offset:92 |
| ; GFX11-NEXT: scratch_load_u16 v10, off, s32 offset:88 |
| ; GFX11-NEXT: scratch_load_b32 v12, off, s32 offset:132 |
| ; GFX11-NEXT: scratch_load_u16 v14, off, s32 |
| ; GFX11-NEXT: scratch_load_u16 v96, off, s32 offset:8 |
| ; GFX11-NEXT: scratch_load_u16 v100, off, s32 offset:16 |
| ; GFX11-NEXT: scratch_load_u16 v101, off, s32 offset:24 |
| ; GFX11-NEXT: scratch_load_u16 v102, off, s32 offset:32 |
| ; GFX11-NEXT: scratch_load_u16 v103, off, s32 offset:40 |
| ; GFX11-NEXT: scratch_load_u16 v113, off, s32 offset:48 |
| ; GFX11-NEXT: scratch_load_u16 v114, off, s32 offset:56 |
| ; GFX11-NEXT: scratch_load_u16 v115, off, s32 offset:64 |
| ; GFX11-NEXT: scratch_load_u16 v116, off, s32 offset:72 |
| ; GFX11-NEXT: scratch_load_u16 v117, off, s32 offset:80 |
| ; GFX11-NEXT: scratch_load_u16 v99, off, s32 offset:84 |
| ; GFX11-NEXT: scratch_load_u16 v85, off, s32 offset:76 |
| ; GFX11-NEXT: scratch_load_u16 v97, off, s32 offset:68 |
| ; GFX11-NEXT: scratch_load_u16 v82, off, s32 offset:60 |
| ; GFX11-NEXT: scratch_load_u16 v86, off, s32 offset:52 |
| ; GFX11-NEXT: scratch_load_u16 v84, off, s32 offset:44 |
| ; GFX11-NEXT: scratch_load_u16 v87, off, s32 offset:36 |
| ; GFX11-NEXT: scratch_load_u16 v68, off, s32 offset:28 |
| ; GFX11-NEXT: scratch_load_u16 v80, off, s32 offset:20 |
| ; GFX11-NEXT: s_clause 0x1 |
| ; GFX11-NEXT: scratch_load_u16 v65, off, s32 offset:12 |
| ; GFX11-NEXT: scratch_load_u16 v71, off, s32 offset:4 |
| ; GFX11-NEXT: v_lshlrev_b16 v50, 8, v1 |
| ; GFX11-NEXT: v_lshlrev_b16 v51, 8, v3 |
| ; GFX11-NEXT: v_lshlrev_b16 v49, 8, v5 |
| ; GFX11-NEXT: v_lshlrev_b16 v39, 8, v7 |
| ; GFX11-NEXT: v_lshlrev_b16 v54, 8, v9 |
| ; GFX11-NEXT: v_lshlrev_b16 v48, 8, v11 |
| ; GFX11-NEXT: v_lshlrev_b16 v55, 8, v13 |
| ; GFX11-NEXT: v_lshlrev_b16 v52, 8, v15 |
| ; GFX11-NEXT: v_lshlrev_b16 v53, 8, v17 |
| ; GFX11-NEXT: v_lshlrev_b16 v17, 8, v19 |
| ; GFX11-NEXT: v_lshlrev_b16 v83, 8, v21 |
| ; GFX11-NEXT: v_lshlrev_b16 v81, 8, v23 |
| ; GFX11-NEXT: v_lshlrev_b16 v21, 8, v25 |
| ; GFX11-NEXT: v_lshlrev_b16 v19, 8, v27 |
| ; GFX11-NEXT: v_lshlrev_b16 v27, 8, v29 |
| ; GFX11-NEXT: s_waitcnt vmcnt(33) |
| ; GFX11-NEXT: v_lshlrev_b16 v119, 8, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(31) |
| ; GFX11-NEXT: v_lshlrev_b16 v118, 8, v2 |
| ; GFX11-NEXT: s_waitcnt vmcnt(22) |
| ; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12 |
| ; GFX11-NEXT: s_waitcnt vmcnt(21) |
| ; GFX11-NEXT: v_lshlrev_b16 v23, 8, v14 |
| ; GFX11-NEXT: s_waitcnt vmcnt(20) |
| ; GFX11-NEXT: v_lshlrev_b16 v98, 8, v96 |
| ; GFX11-NEXT: s_waitcnt vmcnt(19) |
| ; GFX11-NEXT: v_lshlrev_b16 v25, 8, v100 |
| ; GFX11-NEXT: s_waitcnt vmcnt(18) |
| ; GFX11-NEXT: v_lshlrev_b16 v96, 8, v101 |
| ; GFX11-NEXT: s_waitcnt vmcnt(17) |
| ; GFX11-NEXT: v_lshlrev_b16 v29, 8, v102 |
| ; GFX11-NEXT: s_waitcnt vmcnt(16) |
| ; GFX11-NEXT: v_lshlrev_b16 v112, 8, v103 |
| ; GFX11-NEXT: s_waitcnt vmcnt(15) |
| ; GFX11-NEXT: v_lshlrev_b16 v103, 8, v113 |
| ; GFX11-NEXT: s_waitcnt vmcnt(14) |
| ; GFX11-NEXT: v_lshlrev_b16 v101, 8, v114 |
| ; GFX11-NEXT: s_waitcnt vmcnt(13) |
| ; GFX11-NEXT: v_lshlrev_b16 v100, 8, v115 |
| ; GFX11-NEXT: s_waitcnt vmcnt(12) |
| ; GFX11-NEXT: v_lshlrev_b16 v114, 8, v116 |
| ; GFX11-NEXT: s_waitcnt vmcnt(11) |
| ; GFX11-NEXT: v_lshlrev_b16 v102, 8, v117 |
| ; GFX11-NEXT: v_lshlrev_b16 v117, 8, v10 |
| ; GFX11-NEXT: v_lshlrev_b16 v113, 8, v8 |
| ; GFX11-NEXT: v_lshlrev_b16 v116, 8, v6 |
| ; GFX11-NEXT: v_lshlrev_b16 v115, 8, v4 |
| ; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 |
| ; GFX11-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB55_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execnz .LBB55_4 |
| ; GFX11-NEXT: .LBB55_2: ; %end |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB55_3: ; %cmp.false |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v36 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v37 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v38 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v31 |
| ; GFX11-NEXT: v_and_b32_e32 v5, 0xff, v35 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v0, v50 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v1, v51 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v34 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v33 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v16 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v18 |
| ; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x5040100 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v3, v39 |
| ; GFX11-NEXT: v_or_b32_e32 v3, v4, v54 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v5, v48 |
| ; GFX11-NEXT: v_or_b32_e32 v5, v6, v55 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v32 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v20 |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v2, v49 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v53 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v52 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v83 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v81 |
| ; GFX11-NEXT: v_perm_b32 v1, v1, v2, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v2, v4, v3, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v3, v6, v5, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v4, v8, v7, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v5, v10, v9, 0x5040100 |
| ; GFX11-NEXT: v_and_b32_e32 v6, 0xff, v24 |
| ; GFX11-NEXT: v_and_b32_e32 v7, 0xff, v26 |
| ; GFX11-NEXT: v_and_b32_e32 v8, 0xff, v28 |
| ; GFX11-NEXT: v_and_b32_e32 v9, 0xff, v30 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_and_b32_e32 v10, 0xff, v71 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v65 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v80 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v68 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v87 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v84 |
| ; GFX11-NEXT: v_or_b32_e32 v6, v6, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v7, v7, v19 |
| ; GFX11-NEXT: v_or_b32_e32 v8, v8, v27 |
| ; GFX11-NEXT: v_or_b32_e32 v9, v9, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v10, v10, v98 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v25 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v96 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v29 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v112 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v103 |
| ; GFX11-NEXT: v_perm_b32 v6, v7, v6, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v7, v9, v8, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v8, v11, v10, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v9, v13, v12, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v10, v15, v14, 0x5040100 |
| ; GFX11-NEXT: v_and_b32_e32 v11, 0xff, v86 |
| ; GFX11-NEXT: v_and_b32_e32 v12, 0xff, v82 |
| ; GFX11-NEXT: v_and_b32_e32 v13, 0xff, v97 |
| ; GFX11-NEXT: v_and_b32_e32 v14, 0xff, v85 |
| ; GFX11-NEXT: v_and_b32_e32 v15, 0xff, v99 |
| ; GFX11-NEXT: v_and_b32_e32 v16, 0xff, v64 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v69 |
| ; GFX11-NEXT: v_and_b32_e32 v18, 0xff, v66 |
| ; GFX11-NEXT: v_and_b32_e32 v19, 0xff, v70 |
| ; GFX11-NEXT: v_and_b32_e32 v20, 0xff, v67 |
| ; GFX11-NEXT: v_or_b32_e32 v11, v11, v101 |
| ; GFX11-NEXT: v_or_b32_e32 v12, v12, v100 |
| ; GFX11-NEXT: v_or_b32_e32 v13, v13, v114 |
| ; GFX11-NEXT: v_or_b32_e32 v14, v14, v102 |
| ; GFX11-NEXT: v_or_b32_e32 v15, v15, v117 |
| ; GFX11-NEXT: v_or_b32_e32 v16, v16, v113 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v17, v116 |
| ; GFX11-NEXT: v_or_b32_e32 v18, v18, v115 |
| ; GFX11-NEXT: v_or_b32_e32 v19, v19, v118 |
| ; GFX11-NEXT: v_or_b32_e32 v20, v20, v119 |
| ; GFX11-NEXT: v_perm_b32 v11, v12, v11, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v12, v14, v13, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v13, v16, v15, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v14, v18, v17, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v15, v20, v19, 0x5040100 |
| ; GFX11-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-NEXT: ; implicit-def: $vgpr16 |
| ; GFX11-NEXT: ; implicit-def: $vgpr18 |
| ; GFX11-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-NEXT: ; implicit-def: $vgpr71 |
| ; GFX11-NEXT: ; implicit-def: $vgpr65 |
| ; GFX11-NEXT: ; implicit-def: $vgpr80 |
| ; GFX11-NEXT: ; implicit-def: $vgpr68 |
| ; GFX11-NEXT: ; implicit-def: $vgpr87 |
| ; GFX11-NEXT: ; implicit-def: $vgpr84 |
| ; GFX11-NEXT: ; implicit-def: $vgpr86 |
| ; GFX11-NEXT: ; implicit-def: $vgpr82 |
| ; GFX11-NEXT: ; implicit-def: $vgpr97 |
| ; GFX11-NEXT: ; implicit-def: $vgpr85 |
| ; GFX11-NEXT: ; implicit-def: $vgpr99 |
| ; GFX11-NEXT: ; implicit-def: $vgpr64 |
| ; GFX11-NEXT: ; implicit-def: $vgpr69 |
| ; GFX11-NEXT: ; implicit-def: $vgpr66 |
| ; GFX11-NEXT: ; implicit-def: $vgpr70 |
| ; GFX11-NEXT: ; implicit-def: $vgpr67 |
| ; GFX11-NEXT: ; implicit-def: $vgpr50 |
| ; GFX11-NEXT: ; implicit-def: $vgpr51 |
| ; GFX11-NEXT: ; implicit-def: $vgpr49 |
| ; GFX11-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-NEXT: ; implicit-def: $vgpr54 |
| ; GFX11-NEXT: ; implicit-def: $vgpr48 |
| ; GFX11-NEXT: ; implicit-def: $vgpr55 |
| ; GFX11-NEXT: ; implicit-def: $vgpr52 |
| ; GFX11-NEXT: ; implicit-def: $vgpr53 |
| ; GFX11-NEXT: ; implicit-def: $vgpr17 |
| ; GFX11-NEXT: ; implicit-def: $vgpr83 |
| ; GFX11-NEXT: ; implicit-def: $vgpr81 |
| ; GFX11-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-NEXT: ; implicit-def: $vgpr19 |
| ; GFX11-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-NEXT: ; implicit-def: $vgpr98 |
| ; GFX11-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-NEXT: ; implicit-def: $vgpr96 |
| ; GFX11-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-NEXT: ; implicit-def: $vgpr112 |
| ; GFX11-NEXT: ; implicit-def: $vgpr103 |
| ; GFX11-NEXT: ; implicit-def: $vgpr101 |
| ; GFX11-NEXT: ; implicit-def: $vgpr100 |
| ; GFX11-NEXT: ; implicit-def: $vgpr114 |
| ; GFX11-NEXT: ; implicit-def: $vgpr102 |
| ; GFX11-NEXT: ; implicit-def: $vgpr117 |
| ; GFX11-NEXT: ; implicit-def: $vgpr113 |
| ; GFX11-NEXT: ; implicit-def: $vgpr116 |
| ; GFX11-NEXT: ; implicit-def: $vgpr115 |
| ; GFX11-NEXT: ; implicit-def: $vgpr118 |
| ; GFX11-NEXT: ; implicit-def: $vgpr119 |
| ; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-NEXT: s_cbranch_execz .LBB55_2 |
| ; GFX11-NEXT: .LBB55_4: ; %cmp.true |
| ; GFX11-NEXT: v_add_nc_u16 v0, v70, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v67, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v69, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v66, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v5, v64, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: s_waitcnt vmcnt(8) |
| ; GFX11-NEXT: v_add_nc_u16 v6, v97, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v118, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v119, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v99, 3 |
| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-NEXT: v_add_nc_u16 v15, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v64, 0x300, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v116, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v5 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v115, v3 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v6 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v14, 0x300, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v113, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v66, 0x300, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v114, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v4, v117, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v85, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v67, 0x300, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(6) |
| ; GFX11-NEXT: v_add_nc_u16 v0, v86, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v12, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v82, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v13, 0x300, v4 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: s_waitcnt vmcnt(4) |
| ; GFX11-NEXT: v_add_nc_u16 v3, v87, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v84, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v102, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v101, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v100, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v69, 0x300, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v112, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v11, 0x300, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v103, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v70, 0x300, v1 |
| ; GFX11-NEXT: s_waitcnt vmcnt(2) |
| ; GFX11-NEXT: v_add_nc_u16 v1, v80, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v10, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v68, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v68, 0x300, v0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; GFX11-NEXT: v_add_nc_u16 v0, v71, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v65, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v28, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v96, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v29, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v98, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v9, 0x300, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v25, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v25, 0x300, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v27, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v8, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v0, v30, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v27, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v24, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v7, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v26, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v20, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v22, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v23, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v21, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v19, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v19, 0x300, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v83, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v6, 0x300, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v81, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v20, 0x300, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v2, v16, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v5, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v0, v34, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v16, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v31, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v2, 0xff, v2 |
| ; GFX11-NEXT: v_add_nc_u16 v3, v18, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v0, 0xff, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v4, v32, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v53, v2 |
| ; GFX11-NEXT: v_and_b32_e32 v3, 0xff, v3 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v55, v0 |
| ; GFX11-NEXT: v_and_b32_e32 v4, 0xff, v4 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v54, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v18, 0x300, v2 |
| ; GFX11-NEXT: v_or_b32_e32 v2, v17, v3 |
| ; GFX11-NEXT: v_add_nc_u16 v3, 0x300, v0 |
| ; GFX11-NEXT: v_or_b32_e32 v0, v52, v4 |
| ; GFX11-NEXT: v_add_nc_u16 v4, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v1, v33, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v17, v36, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v21, v37, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v22, v38, 3 |
| ; GFX11-NEXT: v_add_nc_u16 v23, v35, 3 |
| ; GFX11-NEXT: v_and_b32_e32 v1, 0xff, v1 |
| ; GFX11-NEXT: v_and_b32_e32 v17, 0xff, v17 |
| ; GFX11-NEXT: v_and_b32_e32 v21, 0xff, v21 |
| ; GFX11-NEXT: v_and_b32_e32 v22, 0xff, v22 |
| ; GFX11-NEXT: v_and_b32_e32 v23, 0xff, v23 |
| ; GFX11-NEXT: v_or_b32_e32 v1, v49, v1 |
| ; GFX11-NEXT: v_or_b32_e32 v17, v50, v17 |
| ; GFX11-NEXT: v_or_b32_e32 v21, v51, v21 |
| ; GFX11-NEXT: v_or_b32_e32 v22, v39, v22 |
| ; GFX11-NEXT: v_or_b32_e32 v23, v48, v23 |
| ; GFX11-NEXT: v_add_nc_u16 v1, 0x300, v1 |
| ; GFX11-NEXT: v_add_nc_u16 v17, 0x300, v17 |
| ; GFX11-NEXT: v_add_nc_u16 v21, 0x300, v21 |
| ; GFX11-NEXT: v_add_nc_u16 v22, 0x300, v22 |
| ; GFX11-NEXT: v_add_nc_u16 v23, 0x300, v23 |
| ; GFX11-NEXT: v_add_nc_u16 v24, 0x300, v0 |
| ; GFX11-NEXT: v_add_nc_u16 v26, 0x300, v2 |
| ; GFX11-NEXT: v_perm_b32 v0, v21, v17, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v1, v22, v1, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v2, v23, v4, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v3, v24, v3, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v4, v26, v18, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v5, v16, v5, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v6, v20, v6, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v7, v19, v7, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v8, v27, v8, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v9, v25, v9, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v10, v68, v10, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v11, v70, v11, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v12, v69, v12, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v13, v67, v13, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v14, v66, v14, 0x5040100 |
| ; GFX11-NEXT: v_perm_b32 v15, v64, v15, 0x5040100 |
| ; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <64 x i8> %a, splat (i8 3) |
| %a2 = bitcast <64 x i8> %a1 to <32 x bfloat> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <64 x i8> %a to <32 x bfloat> |
| br label %end |
| |
| end: |
| %phi = phi <32 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <32 x bfloat> %phi |
| } |