| ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 |
| |
| ; RUN: llc -mtriple=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefix=SI %s |
| ; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefix=VI %s |
| ; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefix=GFX9 %s |
| ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-TRUE16 %s |
| ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-FAKE16 %s |
| |
| define <20 x float> @bitcast_v20i32_to_v20f32(<20 x i32> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v20i32_to_v20f32: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB0_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13 |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1 |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; SI-NEXT: .LBB0_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20i32_to_v20f32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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 v19, vcc, 3, v19 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17 |
| ; 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_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_v20i32_to_v20f32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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 v19, 3, v19 |
| ; GFX9-NEXT: v_add_u32_e32 v18, 3, v18 |
| ; GFX9-NEXT: v_add_u32_e32 v17, 3, v17 |
| ; 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_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_v20i32_to_v20f32: |
| ; 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, v20 |
| ; 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 v19, 3, v19 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17 |
| ; 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_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 <20 x i32> %a, splat (i32 3) |
| %a2 = bitcast <20 x i32> %a1 to <20 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x i32> %a to <20 x float> |
| br label %end |
| |
| end: |
| %phi = phi <20 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x float> %phi |
| } |
| |
| define inreg <20 x float> @bitcast_v20i32_to_v20f32_scalar(<20 x i32> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v20i32_to_v20f32_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB1_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB1_3 |
| ; SI-NEXT: .LBB1_2: ; %cmp.true |
| ; SI-NEXT: s_add_i32 s6, s6, 3 |
| ; SI-NEXT: s_add_i32 s7, s7, 3 |
| ; SI-NEXT: s_add_i32 s8, s8, 3 |
| ; SI-NEXT: s_add_i32 s9, s9, 3 |
| ; SI-NEXT: s_add_i32 s10, s10, 3 |
| ; SI-NEXT: s_add_i32 s11, s11, 3 |
| ; SI-NEXT: s_add_i32 s29, s29, 3 |
| ; SI-NEXT: s_add_i32 s28, s28, 3 |
| ; SI-NEXT: s_add_i32 s27, s27, 3 |
| ; SI-NEXT: s_add_i32 s26, s26, 3 |
| ; SI-NEXT: s_add_i32 s25, s25, 3 |
| ; SI-NEXT: s_add_i32 s24, s24, 3 |
| ; SI-NEXT: s_add_i32 s23, s23, 3 |
| ; SI-NEXT: s_add_i32 s22, s22, 3 |
| ; SI-NEXT: s_add_i32 s21, s21, 3 |
| ; SI-NEXT: s_add_i32 s20, s20, 3 |
| ; SI-NEXT: s_add_i32 s19, s19, 3 |
| ; SI-NEXT: s_add_i32 s18, s18, 3 |
| ; SI-NEXT: s_add_i32 s17, s17, 3 |
| ; SI-NEXT: s_add_i32 s16, s16, 3 |
| ; SI-NEXT: .LBB1_3: ; %end |
| ; SI-NEXT: v_mov_b32_e32 v0, s16 |
| ; SI-NEXT: v_mov_b32_e32 v1, s17 |
| ; SI-NEXT: v_mov_b32_e32 v2, s18 |
| ; SI-NEXT: v_mov_b32_e32 v3, s19 |
| ; SI-NEXT: v_mov_b32_e32 v4, s20 |
| ; SI-NEXT: v_mov_b32_e32 v5, s21 |
| ; SI-NEXT: v_mov_b32_e32 v6, s22 |
| ; SI-NEXT: v_mov_b32_e32 v7, s23 |
| ; SI-NEXT: v_mov_b32_e32 v8, s24 |
| ; SI-NEXT: v_mov_b32_e32 v9, s25 |
| ; SI-NEXT: v_mov_b32_e32 v10, s26 |
| ; SI-NEXT: v_mov_b32_e32 v11, s27 |
| ; SI-NEXT: v_mov_b32_e32 v12, s28 |
| ; SI-NEXT: v_mov_b32_e32 v13, s29 |
| ; SI-NEXT: v_mov_b32_e32 v14, s11 |
| ; SI-NEXT: v_mov_b32_e32 v15, s10 |
| ; SI-NEXT: v_mov_b32_e32 v16, s9 |
| ; SI-NEXT: v_mov_b32_e32 v17, s8 |
| ; SI-NEXT: v_mov_b32_e32 v18, s7 |
| ; SI-NEXT: v_mov_b32_e32 v19, s6 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB1_4: |
| ; SI-NEXT: s_branch .LBB1_2 |
| ; |
| ; VI-LABEL: bitcast_v20i32_to_v20f32_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB1_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB1_3 |
| ; VI-NEXT: .LBB1_2: ; %cmp.true |
| ; VI-NEXT: s_add_i32 s6, s6, 3 |
| ; VI-NEXT: s_add_i32 s7, s7, 3 |
| ; VI-NEXT: s_add_i32 s8, s8, 3 |
| ; VI-NEXT: s_add_i32 s9, s9, 3 |
| ; VI-NEXT: s_add_i32 s10, s10, 3 |
| ; VI-NEXT: s_add_i32 s11, s11, 3 |
| ; VI-NEXT: s_add_i32 s29, s29, 3 |
| ; VI-NEXT: s_add_i32 s28, s28, 3 |
| ; VI-NEXT: s_add_i32 s27, s27, 3 |
| ; VI-NEXT: s_add_i32 s26, s26, 3 |
| ; VI-NEXT: s_add_i32 s25, s25, 3 |
| ; VI-NEXT: s_add_i32 s24, s24, 3 |
| ; VI-NEXT: s_add_i32 s23, s23, 3 |
| ; VI-NEXT: s_add_i32 s22, s22, 3 |
| ; VI-NEXT: s_add_i32 s21, s21, 3 |
| ; VI-NEXT: s_add_i32 s20, s20, 3 |
| ; VI-NEXT: s_add_i32 s19, s19, 3 |
| ; VI-NEXT: s_add_i32 s18, s18, 3 |
| ; VI-NEXT: s_add_i32 s17, s17, 3 |
| ; VI-NEXT: s_add_i32 s16, s16, 3 |
| ; VI-NEXT: .LBB1_3: ; %end |
| ; VI-NEXT: v_mov_b32_e32 v0, s16 |
| ; VI-NEXT: v_mov_b32_e32 v1, s17 |
| ; VI-NEXT: v_mov_b32_e32 v2, s18 |
| ; VI-NEXT: v_mov_b32_e32 v3, s19 |
| ; VI-NEXT: v_mov_b32_e32 v4, s20 |
| ; VI-NEXT: v_mov_b32_e32 v5, s21 |
| ; VI-NEXT: v_mov_b32_e32 v6, s22 |
| ; VI-NEXT: v_mov_b32_e32 v7, s23 |
| ; VI-NEXT: v_mov_b32_e32 v8, s24 |
| ; VI-NEXT: v_mov_b32_e32 v9, s25 |
| ; VI-NEXT: v_mov_b32_e32 v10, s26 |
| ; VI-NEXT: v_mov_b32_e32 v11, s27 |
| ; VI-NEXT: v_mov_b32_e32 v12, s28 |
| ; VI-NEXT: v_mov_b32_e32 v13, s29 |
| ; VI-NEXT: v_mov_b32_e32 v14, s11 |
| ; VI-NEXT: v_mov_b32_e32 v15, s10 |
| ; VI-NEXT: v_mov_b32_e32 v16, s9 |
| ; VI-NEXT: v_mov_b32_e32 v17, s8 |
| ; VI-NEXT: v_mov_b32_e32 v18, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB1_4: |
| ; VI-NEXT: s_branch .LBB1_2 |
| ; |
| ; GFX9-LABEL: bitcast_v20i32_to_v20f32_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s10, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s11, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB1_4 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB1_3 |
| ; GFX9-NEXT: .LBB1_2: ; %cmp.true |
| ; GFX9-NEXT: s_add_i32 s6, s6, 3 |
| ; GFX9-NEXT: s_add_i32 s7, s7, 3 |
| ; GFX9-NEXT: s_add_i32 s8, s8, 3 |
| ; GFX9-NEXT: s_add_i32 s9, s9, 3 |
| ; GFX9-NEXT: s_add_i32 s10, s10, 3 |
| ; GFX9-NEXT: s_add_i32 s11, s11, 3 |
| ; GFX9-NEXT: s_add_i32 s29, s29, 3 |
| ; GFX9-NEXT: s_add_i32 s28, s28, 3 |
| ; GFX9-NEXT: s_add_i32 s27, s27, 3 |
| ; GFX9-NEXT: s_add_i32 s26, s26, 3 |
| ; GFX9-NEXT: s_add_i32 s25, s25, 3 |
| ; GFX9-NEXT: s_add_i32 s24, s24, 3 |
| ; GFX9-NEXT: s_add_i32 s23, s23, 3 |
| ; GFX9-NEXT: s_add_i32 s22, s22, 3 |
| ; GFX9-NEXT: s_add_i32 s21, s21, 3 |
| ; GFX9-NEXT: s_add_i32 s20, s20, 3 |
| ; GFX9-NEXT: s_add_i32 s19, s19, 3 |
| ; GFX9-NEXT: s_add_i32 s18, s18, 3 |
| ; GFX9-NEXT: s_add_i32 s17, s17, 3 |
| ; GFX9-NEXT: s_add_i32 s16, s16, 3 |
| ; GFX9-NEXT: .LBB1_3: ; %end |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s28 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s29 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s6 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; GFX9-NEXT: .LBB1_4: |
| ; GFX9-NEXT: s_branch .LBB1_2 |
| ; |
| ; GFX11-LABEL: bitcast_v20i32_to_v20f32_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB1_4 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB1_3 |
| ; GFX11-NEXT: .LBB1_2: ; %cmp.true |
| ; GFX11-NEXT: s_add_i32 s4, s4, 3 |
| ; GFX11-NEXT: s_add_i32 s5, s5, 3 |
| ; GFX11-NEXT: s_add_i32 s29, s29, 3 |
| ; GFX11-NEXT: s_add_i32 s28, s28, 3 |
| ; GFX11-NEXT: s_add_i32 s27, s27, 3 |
| ; GFX11-NEXT: s_add_i32 s26, s26, 3 |
| ; GFX11-NEXT: s_add_i32 s25, s25, 3 |
| ; GFX11-NEXT: s_add_i32 s24, s24, 3 |
| ; GFX11-NEXT: s_add_i32 s23, s23, 3 |
| ; GFX11-NEXT: s_add_i32 s22, s22, 3 |
| ; GFX11-NEXT: s_add_i32 s21, s21, 3 |
| ; GFX11-NEXT: s_add_i32 s20, s20, 3 |
| ; GFX11-NEXT: s_add_i32 s19, s19, 3 |
| ; GFX11-NEXT: s_add_i32 s18, s18, 3 |
| ; GFX11-NEXT: s_add_i32 s17, s17, 3 |
| ; GFX11-NEXT: s_add_i32 s16, s16, 3 |
| ; GFX11-NEXT: s_add_i32 s3, s3, 3 |
| ; GFX11-NEXT: s_add_i32 s2, s2, 3 |
| ; GFX11-NEXT: s_add_i32 s1, s1, 3 |
| ; GFX11-NEXT: s_add_i32 s0, s0, 3 |
| ; GFX11-NEXT: .LBB1_3: ; %end |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB1_4: |
| ; GFX11-NEXT: s_branch .LBB1_2 |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <20 x i32> %a, splat (i32 3) |
| %a2 = bitcast <20 x i32> %a1 to <20 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x i32> %a to <20 x float> |
| br label %end |
| |
| end: |
| %phi = phi <20 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x float> %phi |
| } |
| |
| define <20 x i32> @bitcast_v20f32_to_v20i32(<20 x float> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v20f32_to_v20i32: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB2_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_add_f32_e32 v19, 1.0, v19 |
| ; SI-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; SI-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; SI-NEXT: v_add_f32_e32 v16, 1.0, v16 |
| ; SI-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; SI-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; SI-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; SI-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; SI-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; SI-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; SI-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; SI-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; SI-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; SI-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; SI-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; SI-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; SI-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; SI-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; SI-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; SI-NEXT: .LBB2_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20f32_to_v20i32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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_f32_e32 v19, 1.0, v19 |
| ; VI-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; VI-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; 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_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: .LBB2_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v20f32_to_v20i32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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_f32_e32 v19, 1.0, v19 |
| ; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; 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_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: .LBB2_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v20f32_to_v20i32: |
| ; 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, v20 |
| ; 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_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18 |
| ; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16 |
| ; 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: .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 = fadd <20 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <20 x float> %a1 to <20 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x float> %a to <20 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <20 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x i32> %phi |
| } |
| |
| define inreg <20 x i32> @bitcast_v20f32_to_v20i32_scalar(<20 x float> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v20f32_to_v20i32_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; SI-NEXT: s_mov_b32 s49, s29 |
| ; SI-NEXT: s_mov_b32 s48, s28 |
| ; SI-NEXT: s_mov_b32 s47, s27 |
| ; SI-NEXT: s_mov_b32 s46, s26 |
| ; SI-NEXT: s_mov_b32 s45, s25 |
| ; SI-NEXT: s_mov_b32 s44, s24 |
| ; SI-NEXT: s_mov_b32 s43, s23 |
| ; SI-NEXT: s_mov_b32 s42, s22 |
| ; SI-NEXT: s_mov_b32 s41, s21 |
| ; SI-NEXT: s_mov_b32 s40, s20 |
| ; SI-NEXT: s_mov_b32 s39, s19 |
| ; SI-NEXT: s_mov_b32 s38, s18 |
| ; SI-NEXT: s_mov_b32 s37, s17 |
| ; SI-NEXT: s_mov_b32 s36, s16 |
| ; SI-NEXT: v_readfirstlane_b32 s55, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s54, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s53, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s52, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s51, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s50, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB3_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB3_4 |
| ; SI-NEXT: .LBB3_2: ; %cmp.true |
| ; SI-NEXT: v_add_f32_e64 v19, s55, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v18, s54, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v17, s53, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v16, s52, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v15, s51, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v14, s50, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v13, s49, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v12, s48, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v11, s47, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v10, s46, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v9, s45, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v8, s44, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v7, s43, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v6, s42, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v5, s41, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v4, s40, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v3, s39, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v2, s38, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v1, s37, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v0, s36, 1.0 |
| ; SI-NEXT: s_branch .LBB3_5 |
| ; SI-NEXT: .LBB3_3: |
| ; SI-NEXT: s_branch .LBB3_2 |
| ; SI-NEXT: .LBB3_4: |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_mov_b32_e32 v20, s56 |
| ; SI-NEXT: v_mov_b32_e32 v21, s57 |
| ; SI-NEXT: v_mov_b32_e32 v22, s58 |
| ; SI-NEXT: v_mov_b32_e32 v23, s59 |
| ; SI-NEXT: v_mov_b32_e32 v24, s60 |
| ; SI-NEXT: v_mov_b32_e32 v25, s61 |
| ; SI-NEXT: v_mov_b32_e32 v26, s62 |
| ; SI-NEXT: v_mov_b32_e32 v27, s63 |
| ; SI-NEXT: v_mov_b32_e32 v28, s64 |
| ; SI-NEXT: v_mov_b32_e32 v29, s65 |
| ; SI-NEXT: v_mov_b32_e32 v30, s66 |
| ; SI-NEXT: v_mov_b32_e32 v31, s67 |
| ; SI-NEXT: .LBB3_5: ; %end |
| ; SI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20f32_to_v20i32_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; VI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; VI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; VI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; VI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; VI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; VI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; VI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; VI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; VI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; VI-NEXT: s_mov_b32 s49, s29 |
| ; VI-NEXT: s_mov_b32 s48, s28 |
| ; VI-NEXT: s_mov_b32 s47, s27 |
| ; VI-NEXT: s_mov_b32 s46, s26 |
| ; VI-NEXT: s_mov_b32 s45, s25 |
| ; VI-NEXT: s_mov_b32 s44, s24 |
| ; VI-NEXT: s_mov_b32 s43, s23 |
| ; VI-NEXT: s_mov_b32 s42, s22 |
| ; VI-NEXT: s_mov_b32 s41, s21 |
| ; VI-NEXT: s_mov_b32 s40, s20 |
| ; VI-NEXT: s_mov_b32 s39, s19 |
| ; VI-NEXT: s_mov_b32 s38, s18 |
| ; VI-NEXT: s_mov_b32 s37, s17 |
| ; VI-NEXT: s_mov_b32 s36, s16 |
| ; VI-NEXT: v_readfirstlane_b32 s55, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s54, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s53, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s52, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s51, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s50, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB3_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB3_4 |
| ; VI-NEXT: .LBB3_2: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e64 v19, s55, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v18, s54, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v17, s53, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v16, s52, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v15, s51, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v14, s50, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v13, s49, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v12, s48, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v11, s47, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v10, s46, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v9, s45, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v8, s44, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v7, s43, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v6, s42, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v5, s41, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v4, s40, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v3, s39, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v2, s38, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v1, s37, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v0, s36, 1.0 |
| ; VI-NEXT: s_branch .LBB3_5 |
| ; VI-NEXT: .LBB3_3: |
| ; VI-NEXT: s_branch .LBB3_2 |
| ; VI-NEXT: .LBB3_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_mov_b32_e32 v20, s56 |
| ; VI-NEXT: v_mov_b32_e32 v21, s57 |
| ; VI-NEXT: v_mov_b32_e32 v22, s58 |
| ; VI-NEXT: v_mov_b32_e32 v23, s59 |
| ; VI-NEXT: v_mov_b32_e32 v24, s60 |
| ; VI-NEXT: v_mov_b32_e32 v25, s61 |
| ; VI-NEXT: v_mov_b32_e32 v26, s62 |
| ; VI-NEXT: v_mov_b32_e32 v27, s63 |
| ; VI-NEXT: v_mov_b32_e32 v28, s64 |
| ; VI-NEXT: v_mov_b32_e32 v29, s65 |
| ; VI-NEXT: v_mov_b32_e32 v30, s66 |
| ; VI-NEXT: v_mov_b32_e32 v31, s67 |
| ; VI-NEXT: .LBB3_5: ; %end |
| ; VI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; VI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; VI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; VI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; VI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; VI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; VI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; VI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; VI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; VI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; VI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; VI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v20f32_to_v20i32_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_mov_b32 s49, s29 |
| ; GFX9-NEXT: s_mov_b32 s48, s28 |
| ; GFX9-NEXT: s_mov_b32 s47, s27 |
| ; GFX9-NEXT: s_mov_b32 s46, s26 |
| ; GFX9-NEXT: s_mov_b32 s45, s25 |
| ; GFX9-NEXT: s_mov_b32 s44, s24 |
| ; GFX9-NEXT: s_mov_b32 s43, s23 |
| ; GFX9-NEXT: s_mov_b32 s42, s22 |
| ; GFX9-NEXT: s_mov_b32 s41, s21 |
| ; GFX9-NEXT: s_mov_b32 s40, s20 |
| ; GFX9-NEXT: s_mov_b32 s39, s19 |
| ; GFX9-NEXT: s_mov_b32 s38, s18 |
| ; GFX9-NEXT: s_mov_b32 s37, s17 |
| ; GFX9-NEXT: s_mov_b32 s36, s16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s55, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s54, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s53, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s52, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s51, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s50, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB3_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB3_4 |
| ; GFX9-NEXT: .LBB3_2: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e64 v19, s55, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v18, s54, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v17, s53, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v16, s52, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v15, s51, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v14, s50, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v13, s49, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v12, s48, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v11, s47, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v10, s46, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v9, s45, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v8, s44, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v7, s43, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v6, s42, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v5, s41, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v4, s40, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v3, s39, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v2, s38, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v1, s37, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v0, s36, 1.0 |
| ; GFX9-NEXT: s_branch .LBB3_5 |
| ; GFX9-NEXT: .LBB3_3: |
| ; GFX9-NEXT: s_branch .LBB3_2 |
| ; GFX9-NEXT: .LBB3_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB3_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v20f32_to_v20i32_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_xor_saveexec_b32 s4, -1 |
| ; GFX11-NEXT: scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill |
| ; GFX11-NEXT: s_mov_b32 exec_lo, s4 |
| ; GFX11-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX11-NEXT: s_mov_b32 s36, s0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s0, v2 |
| ; GFX11-NEXT: s_mov_b32 s47, s23 |
| ; GFX11-NEXT: s_mov_b32 s46, s22 |
| ; GFX11-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX11-NEXT: s_mov_b32 s45, s21 |
| ; GFX11-NEXT: s_mov_b32 s44, s20 |
| ; GFX11-NEXT: s_mov_b32 s43, s19 |
| ; GFX11-NEXT: s_mov_b32 s42, s18 |
| ; GFX11-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX11-NEXT: s_mov_b32 s41, s17 |
| ; GFX11-NEXT: s_mov_b32 s40, s16 |
| ; GFX11-NEXT: s_mov_b32 s38, s2 |
| ; GFX11-NEXT: s_mov_b32 s37, s1 |
| ; GFX11-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX11-NEXT: s_mov_b32 s39, s3 |
| ; GFX11-NEXT: s_cmp_lg_u32 s0, 0 |
| ; GFX11-NEXT: s_mov_b32 s0, 0 |
| ; GFX11-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX11-NEXT: s_mov_b32 s48, s24 |
| ; GFX11-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX11-NEXT: s_mov_b32 s49, s25 |
| ; GFX11-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX11-NEXT: s_mov_b32 s50, s26 |
| ; GFX11-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX11-NEXT: s_mov_b32 s51, s27 |
| ; GFX11-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX11-NEXT: s_mov_b32 s52, s28 |
| ; GFX11-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX11-NEXT: s_mov_b32 s53, s29 |
| ; GFX11-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX11-NEXT: v_readfirstlane_b32 s54, v0 |
| ; GFX11-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX11-NEXT: v_readfirstlane_b32 s55, v1 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB3_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB3_4 |
| ; GFX11-NEXT: .LBB3_2: ; %cmp.true |
| ; GFX11-NEXT: v_add_f32_e64 v19, s55, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v18, s54, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v17, s53, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v16, s52, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v15, s51, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v14, s50, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v13, s49, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v12, s48, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v11, s47, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v10, s46, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v9, s45, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v8, s44, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v7, s43, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v6, s42, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v5, s41, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v4, s40, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v3, s39, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v2, s38, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v1, s37, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v0, s36, 1.0 |
| ; GFX11-NEXT: s_branch .LBB3_5 |
| ; GFX11-NEXT: .LBB3_3: |
| ; GFX11-NEXT: s_branch .LBB3_2 |
| ; GFX11-NEXT: .LBB3_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67 |
| ; GFX11-NEXT: .LBB3_5: ; %end |
| ; GFX11-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX11-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX11-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX11-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX11-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX11-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX11-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX11-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX11-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX11-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX11-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX11-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX11-NEXT: s_xor_saveexec_b32 s0, -1 |
| ; GFX11-NEXT: scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload |
| ; GFX11-NEXT: s_mov_b32 exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; 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 <20 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <20 x float> %a1 to <20 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x float> %a to <20 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <20 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x i32> %phi |
| } |
| |
| define <10 x i64> @bitcast_v20i32_to_v10i64(<20 x i32> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v20i32_to_v10i64: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB4_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13 |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1 |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; SI-NEXT: .LBB4_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20i32_to_v10i64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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 v19, vcc, 3, v19 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17 |
| ; 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_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_v20i32_to_v10i64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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 v19, 3, v19 |
| ; GFX9-NEXT: v_add_u32_e32 v18, 3, v18 |
| ; GFX9-NEXT: v_add_u32_e32 v17, 3, v17 |
| ; 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_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_v20i32_to_v10i64: |
| ; 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, v20 |
| ; 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 v19, 3, v19 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17 |
| ; 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_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 <20 x i32> %a, splat (i32 3) |
| %a2 = bitcast <20 x i32> %a1 to <10 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x i32> %a to <10 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <10 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x i64> %phi |
| } |
| |
| define inreg <10 x i64> @bitcast_v20i32_to_v10i64_scalar(<20 x i32> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v20i32_to_v10i64_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB5_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB5_3 |
| ; SI-NEXT: .LBB5_2: ; %cmp.true |
| ; SI-NEXT: s_add_i32 s6, s6, 3 |
| ; SI-NEXT: s_add_i32 s7, s7, 3 |
| ; SI-NEXT: s_add_i32 s8, s8, 3 |
| ; SI-NEXT: s_add_i32 s9, s9, 3 |
| ; SI-NEXT: s_add_i32 s10, s10, 3 |
| ; SI-NEXT: s_add_i32 s11, s11, 3 |
| ; SI-NEXT: s_add_i32 s29, s29, 3 |
| ; SI-NEXT: s_add_i32 s28, s28, 3 |
| ; SI-NEXT: s_add_i32 s27, s27, 3 |
| ; SI-NEXT: s_add_i32 s26, s26, 3 |
| ; SI-NEXT: s_add_i32 s25, s25, 3 |
| ; SI-NEXT: s_add_i32 s24, s24, 3 |
| ; SI-NEXT: s_add_i32 s23, s23, 3 |
| ; SI-NEXT: s_add_i32 s22, s22, 3 |
| ; SI-NEXT: s_add_i32 s21, s21, 3 |
| ; SI-NEXT: s_add_i32 s20, s20, 3 |
| ; SI-NEXT: s_add_i32 s19, s19, 3 |
| ; SI-NEXT: s_add_i32 s18, s18, 3 |
| ; SI-NEXT: s_add_i32 s17, s17, 3 |
| ; SI-NEXT: s_add_i32 s16, s16, 3 |
| ; SI-NEXT: .LBB5_3: ; %end |
| ; SI-NEXT: v_mov_b32_e32 v0, s16 |
| ; SI-NEXT: v_mov_b32_e32 v1, s17 |
| ; SI-NEXT: v_mov_b32_e32 v2, s18 |
| ; SI-NEXT: v_mov_b32_e32 v3, s19 |
| ; SI-NEXT: v_mov_b32_e32 v4, s20 |
| ; SI-NEXT: v_mov_b32_e32 v5, s21 |
| ; SI-NEXT: v_mov_b32_e32 v6, s22 |
| ; SI-NEXT: v_mov_b32_e32 v7, s23 |
| ; SI-NEXT: v_mov_b32_e32 v8, s24 |
| ; SI-NEXT: v_mov_b32_e32 v9, s25 |
| ; SI-NEXT: v_mov_b32_e32 v10, s26 |
| ; SI-NEXT: v_mov_b32_e32 v11, s27 |
| ; SI-NEXT: v_mov_b32_e32 v12, s28 |
| ; SI-NEXT: v_mov_b32_e32 v13, s29 |
| ; SI-NEXT: v_mov_b32_e32 v14, s11 |
| ; SI-NEXT: v_mov_b32_e32 v15, s10 |
| ; SI-NEXT: v_mov_b32_e32 v16, s9 |
| ; SI-NEXT: v_mov_b32_e32 v17, s8 |
| ; SI-NEXT: v_mov_b32_e32 v18, s7 |
| ; SI-NEXT: v_mov_b32_e32 v19, s6 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB5_4: |
| ; SI-NEXT: s_branch .LBB5_2 |
| ; |
| ; VI-LABEL: bitcast_v20i32_to_v10i64_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB5_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB5_3 |
| ; VI-NEXT: .LBB5_2: ; %cmp.true |
| ; VI-NEXT: s_add_i32 s6, s6, 3 |
| ; VI-NEXT: s_add_i32 s7, s7, 3 |
| ; VI-NEXT: s_add_i32 s8, s8, 3 |
| ; VI-NEXT: s_add_i32 s9, s9, 3 |
| ; VI-NEXT: s_add_i32 s10, s10, 3 |
| ; VI-NEXT: s_add_i32 s11, s11, 3 |
| ; VI-NEXT: s_add_i32 s29, s29, 3 |
| ; VI-NEXT: s_add_i32 s28, s28, 3 |
| ; VI-NEXT: s_add_i32 s27, s27, 3 |
| ; VI-NEXT: s_add_i32 s26, s26, 3 |
| ; VI-NEXT: s_add_i32 s25, s25, 3 |
| ; VI-NEXT: s_add_i32 s24, s24, 3 |
| ; VI-NEXT: s_add_i32 s23, s23, 3 |
| ; VI-NEXT: s_add_i32 s22, s22, 3 |
| ; VI-NEXT: s_add_i32 s21, s21, 3 |
| ; VI-NEXT: s_add_i32 s20, s20, 3 |
| ; VI-NEXT: s_add_i32 s19, s19, 3 |
| ; VI-NEXT: s_add_i32 s18, s18, 3 |
| ; VI-NEXT: s_add_i32 s17, s17, 3 |
| ; VI-NEXT: s_add_i32 s16, s16, 3 |
| ; VI-NEXT: .LBB5_3: ; %end |
| ; VI-NEXT: v_mov_b32_e32 v0, s16 |
| ; VI-NEXT: v_mov_b32_e32 v1, s17 |
| ; VI-NEXT: v_mov_b32_e32 v2, s18 |
| ; VI-NEXT: v_mov_b32_e32 v3, s19 |
| ; VI-NEXT: v_mov_b32_e32 v4, s20 |
| ; VI-NEXT: v_mov_b32_e32 v5, s21 |
| ; VI-NEXT: v_mov_b32_e32 v6, s22 |
| ; VI-NEXT: v_mov_b32_e32 v7, s23 |
| ; VI-NEXT: v_mov_b32_e32 v8, s24 |
| ; VI-NEXT: v_mov_b32_e32 v9, s25 |
| ; VI-NEXT: v_mov_b32_e32 v10, s26 |
| ; VI-NEXT: v_mov_b32_e32 v11, s27 |
| ; VI-NEXT: v_mov_b32_e32 v12, s28 |
| ; VI-NEXT: v_mov_b32_e32 v13, s29 |
| ; VI-NEXT: v_mov_b32_e32 v14, s11 |
| ; VI-NEXT: v_mov_b32_e32 v15, s10 |
| ; VI-NEXT: v_mov_b32_e32 v16, s9 |
| ; VI-NEXT: v_mov_b32_e32 v17, s8 |
| ; VI-NEXT: v_mov_b32_e32 v18, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB5_4: |
| ; VI-NEXT: s_branch .LBB5_2 |
| ; |
| ; GFX9-LABEL: bitcast_v20i32_to_v10i64_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s10, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s11, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB5_4 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB5_3 |
| ; GFX9-NEXT: .LBB5_2: ; %cmp.true |
| ; GFX9-NEXT: s_add_i32 s6, s6, 3 |
| ; GFX9-NEXT: s_add_i32 s7, s7, 3 |
| ; GFX9-NEXT: s_add_i32 s8, s8, 3 |
| ; GFX9-NEXT: s_add_i32 s9, s9, 3 |
| ; GFX9-NEXT: s_add_i32 s10, s10, 3 |
| ; GFX9-NEXT: s_add_i32 s11, s11, 3 |
| ; GFX9-NEXT: s_add_i32 s29, s29, 3 |
| ; GFX9-NEXT: s_add_i32 s28, s28, 3 |
| ; GFX9-NEXT: s_add_i32 s27, s27, 3 |
| ; GFX9-NEXT: s_add_i32 s26, s26, 3 |
| ; GFX9-NEXT: s_add_i32 s25, s25, 3 |
| ; GFX9-NEXT: s_add_i32 s24, s24, 3 |
| ; GFX9-NEXT: s_add_i32 s23, s23, 3 |
| ; GFX9-NEXT: s_add_i32 s22, s22, 3 |
| ; GFX9-NEXT: s_add_i32 s21, s21, 3 |
| ; GFX9-NEXT: s_add_i32 s20, s20, 3 |
| ; GFX9-NEXT: s_add_i32 s19, s19, 3 |
| ; GFX9-NEXT: s_add_i32 s18, s18, 3 |
| ; GFX9-NEXT: s_add_i32 s17, s17, 3 |
| ; GFX9-NEXT: s_add_i32 s16, s16, 3 |
| ; GFX9-NEXT: .LBB5_3: ; %end |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s28 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s29 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s6 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; GFX9-NEXT: .LBB5_4: |
| ; GFX9-NEXT: s_branch .LBB5_2 |
| ; |
| ; GFX11-LABEL: bitcast_v20i32_to_v10i64_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB5_4 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB5_3 |
| ; GFX11-NEXT: .LBB5_2: ; %cmp.true |
| ; GFX11-NEXT: s_add_i32 s4, s4, 3 |
| ; GFX11-NEXT: s_add_i32 s5, s5, 3 |
| ; GFX11-NEXT: s_add_i32 s29, s29, 3 |
| ; GFX11-NEXT: s_add_i32 s28, s28, 3 |
| ; GFX11-NEXT: s_add_i32 s27, s27, 3 |
| ; GFX11-NEXT: s_add_i32 s26, s26, 3 |
| ; GFX11-NEXT: s_add_i32 s25, s25, 3 |
| ; GFX11-NEXT: s_add_i32 s24, s24, 3 |
| ; GFX11-NEXT: s_add_i32 s23, s23, 3 |
| ; GFX11-NEXT: s_add_i32 s22, s22, 3 |
| ; GFX11-NEXT: s_add_i32 s21, s21, 3 |
| ; GFX11-NEXT: s_add_i32 s20, s20, 3 |
| ; GFX11-NEXT: s_add_i32 s19, s19, 3 |
| ; GFX11-NEXT: s_add_i32 s18, s18, 3 |
| ; GFX11-NEXT: s_add_i32 s17, s17, 3 |
| ; GFX11-NEXT: s_add_i32 s16, s16, 3 |
| ; GFX11-NEXT: s_add_i32 s3, s3, 3 |
| ; GFX11-NEXT: s_add_i32 s2, s2, 3 |
| ; GFX11-NEXT: s_add_i32 s1, s1, 3 |
| ; GFX11-NEXT: s_add_i32 s0, s0, 3 |
| ; GFX11-NEXT: .LBB5_3: ; %end |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB5_4: |
| ; GFX11-NEXT: s_branch .LBB5_2 |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <20 x i32> %a, splat (i32 3) |
| %a2 = bitcast <20 x i32> %a1 to <10 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x i32> %a to <10 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <10 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x i64> %phi |
| } |
| |
| define <20 x i32> @bitcast_v10i64_to_v20i32(<10 x i64> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v10i64_to_v20i32: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB6_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; SI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; SI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; SI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; SI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; SI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; SI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; SI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; SI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; SI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; SI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; SI-NEXT: .LBB6_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10i64_to_v20i32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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 v18, vcc, 3, v18 |
| ; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc |
| ; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16 |
| ; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc |
| ; 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: .LBB6_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10i64_to_v20i32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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_co_u32_e32 v18, vcc, 3, v18 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, 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: 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: .LBB6_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v10i64_to_v20i32: |
| ; 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, v20 |
| ; 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_co_u32 v18, vcc_lo, v18, 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 v19, null, 0, v19, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo |
| ; 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: .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 <10 x i64> %a, splat (i64 3) |
| %a2 = bitcast <10 x i64> %a1 to <20 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x i64> %a to <20 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <20 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x i32> %phi |
| } |
| |
| define inreg <20 x i32> @bitcast_v10i64_to_v20i32_scalar(<10 x i64> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v10i64_to_v20i32_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB7_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB7_3 |
| ; SI-NEXT: .LBB7_2: ; %cmp.true |
| ; SI-NEXT: s_add_u32 s7, s7, 3 |
| ; SI-NEXT: s_addc_u32 s6, s6, 0 |
| ; SI-NEXT: s_add_u32 s9, s9, 3 |
| ; SI-NEXT: s_addc_u32 s8, s8, 0 |
| ; SI-NEXT: s_add_u32 s11, s11, 3 |
| ; SI-NEXT: s_addc_u32 s10, s10, 0 |
| ; SI-NEXT: s_add_u32 s28, s28, 3 |
| ; SI-NEXT: s_addc_u32 s29, s29, 0 |
| ; SI-NEXT: s_add_u32 s26, s26, 3 |
| ; SI-NEXT: s_addc_u32 s27, s27, 0 |
| ; SI-NEXT: s_add_u32 s24, s24, 3 |
| ; SI-NEXT: s_addc_u32 s25, s25, 0 |
| ; SI-NEXT: s_add_u32 s22, s22, 3 |
| ; SI-NEXT: s_addc_u32 s23, s23, 0 |
| ; SI-NEXT: s_add_u32 s20, s20, 3 |
| ; SI-NEXT: s_addc_u32 s21, s21, 0 |
| ; SI-NEXT: s_add_u32 s18, s18, 3 |
| ; SI-NEXT: s_addc_u32 s19, s19, 0 |
| ; SI-NEXT: s_add_u32 s16, s16, 3 |
| ; SI-NEXT: s_addc_u32 s17, s17, 0 |
| ; SI-NEXT: .LBB7_3: ; %end |
| ; SI-NEXT: v_mov_b32_e32 v0, s16 |
| ; SI-NEXT: v_mov_b32_e32 v1, s17 |
| ; SI-NEXT: v_mov_b32_e32 v2, s18 |
| ; SI-NEXT: v_mov_b32_e32 v3, s19 |
| ; SI-NEXT: v_mov_b32_e32 v4, s20 |
| ; SI-NEXT: v_mov_b32_e32 v5, s21 |
| ; SI-NEXT: v_mov_b32_e32 v6, s22 |
| ; SI-NEXT: v_mov_b32_e32 v7, s23 |
| ; SI-NEXT: v_mov_b32_e32 v8, s24 |
| ; SI-NEXT: v_mov_b32_e32 v9, s25 |
| ; SI-NEXT: v_mov_b32_e32 v10, s26 |
| ; SI-NEXT: v_mov_b32_e32 v11, s27 |
| ; SI-NEXT: v_mov_b32_e32 v12, s28 |
| ; SI-NEXT: v_mov_b32_e32 v13, s29 |
| ; SI-NEXT: v_mov_b32_e32 v14, s11 |
| ; SI-NEXT: v_mov_b32_e32 v15, s10 |
| ; SI-NEXT: v_mov_b32_e32 v16, s9 |
| ; SI-NEXT: v_mov_b32_e32 v17, s8 |
| ; SI-NEXT: v_mov_b32_e32 v18, s7 |
| ; SI-NEXT: v_mov_b32_e32 v19, s6 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB7_4: |
| ; SI-NEXT: s_branch .LBB7_2 |
| ; |
| ; VI-LABEL: bitcast_v10i64_to_v20i32_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB7_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB7_3 |
| ; VI-NEXT: .LBB7_2: ; %cmp.true |
| ; VI-NEXT: s_add_u32 s7, s7, 3 |
| ; VI-NEXT: s_addc_u32 s6, s6, 0 |
| ; VI-NEXT: s_add_u32 s9, s9, 3 |
| ; VI-NEXT: s_addc_u32 s8, s8, 0 |
| ; VI-NEXT: s_add_u32 s11, s11, 3 |
| ; VI-NEXT: s_addc_u32 s10, s10, 0 |
| ; VI-NEXT: s_add_u32 s28, s28, 3 |
| ; VI-NEXT: s_addc_u32 s29, s29, 0 |
| ; VI-NEXT: s_add_u32 s26, s26, 3 |
| ; VI-NEXT: s_addc_u32 s27, s27, 0 |
| ; VI-NEXT: s_add_u32 s24, s24, 3 |
| ; VI-NEXT: s_addc_u32 s25, s25, 0 |
| ; VI-NEXT: s_add_u32 s22, s22, 3 |
| ; VI-NEXT: s_addc_u32 s23, s23, 0 |
| ; VI-NEXT: s_add_u32 s20, s20, 3 |
| ; VI-NEXT: s_addc_u32 s21, s21, 0 |
| ; VI-NEXT: s_add_u32 s18, s18, 3 |
| ; VI-NEXT: s_addc_u32 s19, s19, 0 |
| ; VI-NEXT: s_add_u32 s16, s16, 3 |
| ; VI-NEXT: s_addc_u32 s17, s17, 0 |
| ; VI-NEXT: .LBB7_3: ; %end |
| ; VI-NEXT: v_mov_b32_e32 v0, s16 |
| ; VI-NEXT: v_mov_b32_e32 v1, s17 |
| ; VI-NEXT: v_mov_b32_e32 v2, s18 |
| ; VI-NEXT: v_mov_b32_e32 v3, s19 |
| ; VI-NEXT: v_mov_b32_e32 v4, s20 |
| ; VI-NEXT: v_mov_b32_e32 v5, s21 |
| ; VI-NEXT: v_mov_b32_e32 v6, s22 |
| ; VI-NEXT: v_mov_b32_e32 v7, s23 |
| ; VI-NEXT: v_mov_b32_e32 v8, s24 |
| ; VI-NEXT: v_mov_b32_e32 v9, s25 |
| ; VI-NEXT: v_mov_b32_e32 v10, s26 |
| ; VI-NEXT: v_mov_b32_e32 v11, s27 |
| ; VI-NEXT: v_mov_b32_e32 v12, s28 |
| ; VI-NEXT: v_mov_b32_e32 v13, s29 |
| ; VI-NEXT: v_mov_b32_e32 v14, s11 |
| ; VI-NEXT: v_mov_b32_e32 v15, s10 |
| ; VI-NEXT: v_mov_b32_e32 v16, s9 |
| ; VI-NEXT: v_mov_b32_e32 v17, s8 |
| ; VI-NEXT: v_mov_b32_e32 v18, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB7_4: |
| ; VI-NEXT: s_branch .LBB7_2 |
| ; |
| ; GFX9-LABEL: bitcast_v10i64_to_v20i32_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s10, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s11, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB7_4 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB7_3 |
| ; GFX9-NEXT: .LBB7_2: ; %cmp.true |
| ; GFX9-NEXT: s_add_u32 s7, s7, 3 |
| ; GFX9-NEXT: s_addc_u32 s6, s6, 0 |
| ; GFX9-NEXT: s_add_u32 s9, s9, 3 |
| ; GFX9-NEXT: s_addc_u32 s8, s8, 0 |
| ; GFX9-NEXT: s_add_u32 s11, s11, 3 |
| ; GFX9-NEXT: s_addc_u32 s10, s10, 0 |
| ; GFX9-NEXT: s_add_u32 s28, s28, 3 |
| ; GFX9-NEXT: s_addc_u32 s29, s29, 0 |
| ; GFX9-NEXT: s_add_u32 s26, s26, 3 |
| ; GFX9-NEXT: s_addc_u32 s27, s27, 0 |
| ; GFX9-NEXT: s_add_u32 s24, s24, 3 |
| ; GFX9-NEXT: s_addc_u32 s25, s25, 0 |
| ; GFX9-NEXT: s_add_u32 s22, s22, 3 |
| ; GFX9-NEXT: s_addc_u32 s23, s23, 0 |
| ; GFX9-NEXT: s_add_u32 s20, s20, 3 |
| ; GFX9-NEXT: s_addc_u32 s21, s21, 0 |
| ; GFX9-NEXT: s_add_u32 s18, s18, 3 |
| ; GFX9-NEXT: s_addc_u32 s19, s19, 0 |
| ; GFX9-NEXT: s_add_u32 s16, s16, 3 |
| ; GFX9-NEXT: s_addc_u32 s17, s17, 0 |
| ; GFX9-NEXT: .LBB7_3: ; %end |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s28 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s29 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s6 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; GFX9-NEXT: .LBB7_4: |
| ; GFX9-NEXT: s_branch .LBB7_2 |
| ; |
| ; GFX11-LABEL: bitcast_v10i64_to_v20i32_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB7_4 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB7_3 |
| ; GFX11-NEXT: .LBB7_2: ; %cmp.true |
| ; GFX11-NEXT: s_add_u32 s5, s5, 3 |
| ; GFX11-NEXT: s_addc_u32 s4, s4, 0 |
| ; GFX11-NEXT: s_add_u32 s28, s28, 3 |
| ; GFX11-NEXT: s_addc_u32 s29, s29, 0 |
| ; GFX11-NEXT: s_add_u32 s26, s26, 3 |
| ; GFX11-NEXT: s_addc_u32 s27, s27, 0 |
| ; GFX11-NEXT: s_add_u32 s24, s24, 3 |
| ; GFX11-NEXT: s_addc_u32 s25, s25, 0 |
| ; GFX11-NEXT: s_add_u32 s22, s22, 3 |
| ; GFX11-NEXT: s_addc_u32 s23, s23, 0 |
| ; GFX11-NEXT: s_add_u32 s20, s20, 3 |
| ; GFX11-NEXT: s_addc_u32 s21, s21, 0 |
| ; GFX11-NEXT: s_add_u32 s18, s18, 3 |
| ; GFX11-NEXT: s_addc_u32 s19, s19, 0 |
| ; GFX11-NEXT: s_add_u32 s16, s16, 3 |
| ; GFX11-NEXT: s_addc_u32 s17, s17, 0 |
| ; GFX11-NEXT: s_add_u32 s2, s2, 3 |
| ; GFX11-NEXT: s_addc_u32 s3, s3, 0 |
| ; GFX11-NEXT: s_add_u32 s0, s0, 3 |
| ; GFX11-NEXT: s_addc_u32 s1, s1, 0 |
| ; GFX11-NEXT: .LBB7_3: ; %end |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB7_4: |
| ; GFX11-NEXT: s_branch .LBB7_2 |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <10 x i64> %a, splat (i64 3) |
| %a2 = bitcast <10 x i64> %a1 to <20 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x i64> %a to <20 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <20 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x i32> %phi |
| } |
| |
| define <10 x double> @bitcast_v20i32_to_v10f64(<20 x i32> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v20i32_to_v10f64: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB8_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13 |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1 |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; SI-NEXT: .LBB8_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20i32_to_v10f64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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 v19, vcc, 3, v19 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17 |
| ; 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_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_v20i32_to_v10f64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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 v19, 3, v19 |
| ; GFX9-NEXT: v_add_u32_e32 v18, 3, v18 |
| ; GFX9-NEXT: v_add_u32_e32 v17, 3, v17 |
| ; 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_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_v20i32_to_v10f64: |
| ; 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, v20 |
| ; 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 v19, 3, v19 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18 |
| ; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17 |
| ; 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_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 <20 x i32> %a, splat (i32 3) |
| %a2 = bitcast <20 x i32> %a1 to <10 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x i32> %a to <10 x double> |
| br label %end |
| |
| end: |
| %phi = phi <10 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x double> %phi |
| } |
| |
| define inreg <10 x double> @bitcast_v20i32_to_v10f64_scalar(<20 x i32> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v20i32_to_v10f64_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB9_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB9_3 |
| ; SI-NEXT: .LBB9_2: ; %cmp.true |
| ; SI-NEXT: s_add_i32 s6, s6, 3 |
| ; SI-NEXT: s_add_i32 s7, s7, 3 |
| ; SI-NEXT: s_add_i32 s8, s8, 3 |
| ; SI-NEXT: s_add_i32 s9, s9, 3 |
| ; SI-NEXT: s_add_i32 s10, s10, 3 |
| ; SI-NEXT: s_add_i32 s11, s11, 3 |
| ; SI-NEXT: s_add_i32 s29, s29, 3 |
| ; SI-NEXT: s_add_i32 s28, s28, 3 |
| ; SI-NEXT: s_add_i32 s27, s27, 3 |
| ; SI-NEXT: s_add_i32 s26, s26, 3 |
| ; SI-NEXT: s_add_i32 s25, s25, 3 |
| ; SI-NEXT: s_add_i32 s24, s24, 3 |
| ; SI-NEXT: s_add_i32 s23, s23, 3 |
| ; SI-NEXT: s_add_i32 s22, s22, 3 |
| ; SI-NEXT: s_add_i32 s21, s21, 3 |
| ; SI-NEXT: s_add_i32 s20, s20, 3 |
| ; SI-NEXT: s_add_i32 s19, s19, 3 |
| ; SI-NEXT: s_add_i32 s18, s18, 3 |
| ; SI-NEXT: s_add_i32 s17, s17, 3 |
| ; SI-NEXT: s_add_i32 s16, s16, 3 |
| ; SI-NEXT: .LBB9_3: ; %end |
| ; SI-NEXT: v_mov_b32_e32 v0, s16 |
| ; SI-NEXT: v_mov_b32_e32 v1, s17 |
| ; SI-NEXT: v_mov_b32_e32 v2, s18 |
| ; SI-NEXT: v_mov_b32_e32 v3, s19 |
| ; SI-NEXT: v_mov_b32_e32 v4, s20 |
| ; SI-NEXT: v_mov_b32_e32 v5, s21 |
| ; SI-NEXT: v_mov_b32_e32 v6, s22 |
| ; SI-NEXT: v_mov_b32_e32 v7, s23 |
| ; SI-NEXT: v_mov_b32_e32 v8, s24 |
| ; SI-NEXT: v_mov_b32_e32 v9, s25 |
| ; SI-NEXT: v_mov_b32_e32 v10, s26 |
| ; SI-NEXT: v_mov_b32_e32 v11, s27 |
| ; SI-NEXT: v_mov_b32_e32 v12, s28 |
| ; SI-NEXT: v_mov_b32_e32 v13, s29 |
| ; SI-NEXT: v_mov_b32_e32 v14, s11 |
| ; SI-NEXT: v_mov_b32_e32 v15, s10 |
| ; SI-NEXT: v_mov_b32_e32 v16, s9 |
| ; SI-NEXT: v_mov_b32_e32 v17, s8 |
| ; SI-NEXT: v_mov_b32_e32 v18, s7 |
| ; SI-NEXT: v_mov_b32_e32 v19, s6 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB9_4: |
| ; SI-NEXT: s_branch .LBB9_2 |
| ; |
| ; VI-LABEL: bitcast_v20i32_to_v10f64_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB9_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB9_3 |
| ; VI-NEXT: .LBB9_2: ; %cmp.true |
| ; VI-NEXT: s_add_i32 s6, s6, 3 |
| ; VI-NEXT: s_add_i32 s7, s7, 3 |
| ; VI-NEXT: s_add_i32 s8, s8, 3 |
| ; VI-NEXT: s_add_i32 s9, s9, 3 |
| ; VI-NEXT: s_add_i32 s10, s10, 3 |
| ; VI-NEXT: s_add_i32 s11, s11, 3 |
| ; VI-NEXT: s_add_i32 s29, s29, 3 |
| ; VI-NEXT: s_add_i32 s28, s28, 3 |
| ; VI-NEXT: s_add_i32 s27, s27, 3 |
| ; VI-NEXT: s_add_i32 s26, s26, 3 |
| ; VI-NEXT: s_add_i32 s25, s25, 3 |
| ; VI-NEXT: s_add_i32 s24, s24, 3 |
| ; VI-NEXT: s_add_i32 s23, s23, 3 |
| ; VI-NEXT: s_add_i32 s22, s22, 3 |
| ; VI-NEXT: s_add_i32 s21, s21, 3 |
| ; VI-NEXT: s_add_i32 s20, s20, 3 |
| ; VI-NEXT: s_add_i32 s19, s19, 3 |
| ; VI-NEXT: s_add_i32 s18, s18, 3 |
| ; VI-NEXT: s_add_i32 s17, s17, 3 |
| ; VI-NEXT: s_add_i32 s16, s16, 3 |
| ; VI-NEXT: .LBB9_3: ; %end |
| ; VI-NEXT: v_mov_b32_e32 v0, s16 |
| ; VI-NEXT: v_mov_b32_e32 v1, s17 |
| ; VI-NEXT: v_mov_b32_e32 v2, s18 |
| ; VI-NEXT: v_mov_b32_e32 v3, s19 |
| ; VI-NEXT: v_mov_b32_e32 v4, s20 |
| ; VI-NEXT: v_mov_b32_e32 v5, s21 |
| ; VI-NEXT: v_mov_b32_e32 v6, s22 |
| ; VI-NEXT: v_mov_b32_e32 v7, s23 |
| ; VI-NEXT: v_mov_b32_e32 v8, s24 |
| ; VI-NEXT: v_mov_b32_e32 v9, s25 |
| ; VI-NEXT: v_mov_b32_e32 v10, s26 |
| ; VI-NEXT: v_mov_b32_e32 v11, s27 |
| ; VI-NEXT: v_mov_b32_e32 v12, s28 |
| ; VI-NEXT: v_mov_b32_e32 v13, s29 |
| ; VI-NEXT: v_mov_b32_e32 v14, s11 |
| ; VI-NEXT: v_mov_b32_e32 v15, s10 |
| ; VI-NEXT: v_mov_b32_e32 v16, s9 |
| ; VI-NEXT: v_mov_b32_e32 v17, s8 |
| ; VI-NEXT: v_mov_b32_e32 v18, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB9_4: |
| ; VI-NEXT: s_branch .LBB9_2 |
| ; |
| ; GFX9-LABEL: bitcast_v20i32_to_v10f64_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s10, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s11, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB9_4 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB9_3 |
| ; GFX9-NEXT: .LBB9_2: ; %cmp.true |
| ; GFX9-NEXT: s_add_i32 s6, s6, 3 |
| ; GFX9-NEXT: s_add_i32 s7, s7, 3 |
| ; GFX9-NEXT: s_add_i32 s8, s8, 3 |
| ; GFX9-NEXT: s_add_i32 s9, s9, 3 |
| ; GFX9-NEXT: s_add_i32 s10, s10, 3 |
| ; GFX9-NEXT: s_add_i32 s11, s11, 3 |
| ; GFX9-NEXT: s_add_i32 s29, s29, 3 |
| ; GFX9-NEXT: s_add_i32 s28, s28, 3 |
| ; GFX9-NEXT: s_add_i32 s27, s27, 3 |
| ; GFX9-NEXT: s_add_i32 s26, s26, 3 |
| ; GFX9-NEXT: s_add_i32 s25, s25, 3 |
| ; GFX9-NEXT: s_add_i32 s24, s24, 3 |
| ; GFX9-NEXT: s_add_i32 s23, s23, 3 |
| ; GFX9-NEXT: s_add_i32 s22, s22, 3 |
| ; GFX9-NEXT: s_add_i32 s21, s21, 3 |
| ; GFX9-NEXT: s_add_i32 s20, s20, 3 |
| ; GFX9-NEXT: s_add_i32 s19, s19, 3 |
| ; GFX9-NEXT: s_add_i32 s18, s18, 3 |
| ; GFX9-NEXT: s_add_i32 s17, s17, 3 |
| ; GFX9-NEXT: s_add_i32 s16, s16, 3 |
| ; GFX9-NEXT: .LBB9_3: ; %end |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s28 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s29 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s6 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; GFX9-NEXT: .LBB9_4: |
| ; GFX9-NEXT: s_branch .LBB9_2 |
| ; |
| ; GFX11-LABEL: bitcast_v20i32_to_v10f64_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB9_4 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB9_3 |
| ; GFX11-NEXT: .LBB9_2: ; %cmp.true |
| ; GFX11-NEXT: s_add_i32 s4, s4, 3 |
| ; GFX11-NEXT: s_add_i32 s5, s5, 3 |
| ; GFX11-NEXT: s_add_i32 s29, s29, 3 |
| ; GFX11-NEXT: s_add_i32 s28, s28, 3 |
| ; GFX11-NEXT: s_add_i32 s27, s27, 3 |
| ; GFX11-NEXT: s_add_i32 s26, s26, 3 |
| ; GFX11-NEXT: s_add_i32 s25, s25, 3 |
| ; GFX11-NEXT: s_add_i32 s24, s24, 3 |
| ; GFX11-NEXT: s_add_i32 s23, s23, 3 |
| ; GFX11-NEXT: s_add_i32 s22, s22, 3 |
| ; GFX11-NEXT: s_add_i32 s21, s21, 3 |
| ; GFX11-NEXT: s_add_i32 s20, s20, 3 |
| ; GFX11-NEXT: s_add_i32 s19, s19, 3 |
| ; GFX11-NEXT: s_add_i32 s18, s18, 3 |
| ; GFX11-NEXT: s_add_i32 s17, s17, 3 |
| ; GFX11-NEXT: s_add_i32 s16, s16, 3 |
| ; GFX11-NEXT: s_add_i32 s3, s3, 3 |
| ; GFX11-NEXT: s_add_i32 s2, s2, 3 |
| ; GFX11-NEXT: s_add_i32 s1, s1, 3 |
| ; GFX11-NEXT: s_add_i32 s0, s0, 3 |
| ; GFX11-NEXT: .LBB9_3: ; %end |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB9_4: |
| ; GFX11-NEXT: s_branch .LBB9_2 |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <20 x i32> %a, splat (i32 3) |
| %a2 = bitcast <20 x i32> %a1 to <10 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x i32> %a to <10 x double> |
| br label %end |
| |
| end: |
| %phi = phi <10 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x double> %phi |
| } |
| |
| define <20 x i32> @bitcast_v10f64_to_v20i32(<10 x double> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v10f64_to_v20i32: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB10_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0 |
| ; SI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; SI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; SI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; SI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; SI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; SI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; SI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; SI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; SI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; SI-NEXT: .LBB10_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10f64_to_v20i32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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_f64 v[18:19], v[18:19], 1.0 |
| ; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; 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: .LBB10_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10f64_to_v20i32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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_f64 v[18:19], v[18:19], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; 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: .LBB10_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v10f64_to_v20i32: |
| ; 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, v20 |
| ; 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_f64 v[18:19], v[18:19], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; 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: .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 = fadd <10 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <10 x double> %a1 to <20 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x double> %a to <20 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <20 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x i32> %phi |
| } |
| |
| define inreg <20 x i32> @bitcast_v10f64_to_v20i32_scalar(<10 x double> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v10f64_to_v20i32_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; SI-NEXT: s_mov_b32 s49, s29 |
| ; SI-NEXT: s_mov_b32 s48, s28 |
| ; SI-NEXT: s_mov_b32 s47, s27 |
| ; SI-NEXT: s_mov_b32 s46, s26 |
| ; SI-NEXT: s_mov_b32 s45, s25 |
| ; SI-NEXT: s_mov_b32 s44, s24 |
| ; SI-NEXT: s_mov_b32 s43, s23 |
| ; SI-NEXT: s_mov_b32 s42, s22 |
| ; SI-NEXT: s_mov_b32 s41, s21 |
| ; SI-NEXT: s_mov_b32 s40, s20 |
| ; SI-NEXT: s_mov_b32 s39, s19 |
| ; SI-NEXT: s_mov_b32 s38, s18 |
| ; SI-NEXT: s_mov_b32 s37, s17 |
| ; SI-NEXT: s_mov_b32 s36, s16 |
| ; SI-NEXT: v_readfirstlane_b32 s55, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s54, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s53, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s52, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s51, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s50, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB11_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB11_4 |
| ; SI-NEXT: .LBB11_2: ; %cmp.true |
| ; SI-NEXT: v_add_f64 v[18:19], s[54:55], 1.0 |
| ; SI-NEXT: v_add_f64 v[16:17], s[52:53], 1.0 |
| ; SI-NEXT: v_add_f64 v[14:15], s[50:51], 1.0 |
| ; SI-NEXT: v_add_f64 v[12:13], s[48:49], 1.0 |
| ; SI-NEXT: v_add_f64 v[10:11], s[46:47], 1.0 |
| ; SI-NEXT: v_add_f64 v[8:9], s[44:45], 1.0 |
| ; SI-NEXT: v_add_f64 v[6:7], s[42:43], 1.0 |
| ; SI-NEXT: v_add_f64 v[4:5], s[40:41], 1.0 |
| ; SI-NEXT: v_add_f64 v[2:3], s[38:39], 1.0 |
| ; SI-NEXT: v_add_f64 v[0:1], s[36:37], 1.0 |
| ; SI-NEXT: s_branch .LBB11_5 |
| ; SI-NEXT: .LBB11_3: |
| ; SI-NEXT: s_branch .LBB11_2 |
| ; SI-NEXT: .LBB11_4: |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_mov_b32_e32 v20, s56 |
| ; SI-NEXT: v_mov_b32_e32 v21, s57 |
| ; SI-NEXT: v_mov_b32_e32 v22, s58 |
| ; SI-NEXT: v_mov_b32_e32 v23, s59 |
| ; SI-NEXT: v_mov_b32_e32 v24, s60 |
| ; SI-NEXT: v_mov_b32_e32 v25, s61 |
| ; SI-NEXT: v_mov_b32_e32 v26, s62 |
| ; SI-NEXT: v_mov_b32_e32 v27, s63 |
| ; SI-NEXT: v_mov_b32_e32 v28, s64 |
| ; SI-NEXT: v_mov_b32_e32 v29, s65 |
| ; SI-NEXT: v_mov_b32_e32 v30, s66 |
| ; SI-NEXT: v_mov_b32_e32 v31, s67 |
| ; SI-NEXT: .LBB11_5: ; %end |
| ; SI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10f64_to_v20i32_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; VI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; VI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; VI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; VI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; VI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; VI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; VI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; VI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; VI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; VI-NEXT: s_mov_b32 s49, s29 |
| ; VI-NEXT: s_mov_b32 s48, s28 |
| ; VI-NEXT: s_mov_b32 s47, s27 |
| ; VI-NEXT: s_mov_b32 s46, s26 |
| ; VI-NEXT: s_mov_b32 s45, s25 |
| ; VI-NEXT: s_mov_b32 s44, s24 |
| ; VI-NEXT: s_mov_b32 s43, s23 |
| ; VI-NEXT: s_mov_b32 s42, s22 |
| ; VI-NEXT: s_mov_b32 s41, s21 |
| ; VI-NEXT: s_mov_b32 s40, s20 |
| ; VI-NEXT: s_mov_b32 s39, s19 |
| ; VI-NEXT: s_mov_b32 s38, s18 |
| ; VI-NEXT: s_mov_b32 s37, s17 |
| ; VI-NEXT: s_mov_b32 s36, s16 |
| ; VI-NEXT: v_readfirstlane_b32 s55, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s54, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s53, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s52, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s51, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s50, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB11_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB11_4 |
| ; VI-NEXT: .LBB11_2: ; %cmp.true |
| ; VI-NEXT: v_add_f64 v[18:19], s[54:55], 1.0 |
| ; VI-NEXT: v_add_f64 v[16:17], s[52:53], 1.0 |
| ; VI-NEXT: v_add_f64 v[14:15], s[50:51], 1.0 |
| ; VI-NEXT: v_add_f64 v[12:13], s[48:49], 1.0 |
| ; VI-NEXT: v_add_f64 v[10:11], s[46:47], 1.0 |
| ; VI-NEXT: v_add_f64 v[8:9], s[44:45], 1.0 |
| ; VI-NEXT: v_add_f64 v[6:7], s[42:43], 1.0 |
| ; VI-NEXT: v_add_f64 v[4:5], s[40:41], 1.0 |
| ; VI-NEXT: v_add_f64 v[2:3], s[38:39], 1.0 |
| ; VI-NEXT: v_add_f64 v[0:1], s[36:37], 1.0 |
| ; VI-NEXT: s_branch .LBB11_5 |
| ; VI-NEXT: .LBB11_3: |
| ; VI-NEXT: s_branch .LBB11_2 |
| ; VI-NEXT: .LBB11_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_mov_b32_e32 v20, s56 |
| ; VI-NEXT: v_mov_b32_e32 v21, s57 |
| ; VI-NEXT: v_mov_b32_e32 v22, s58 |
| ; VI-NEXT: v_mov_b32_e32 v23, s59 |
| ; VI-NEXT: v_mov_b32_e32 v24, s60 |
| ; VI-NEXT: v_mov_b32_e32 v25, s61 |
| ; VI-NEXT: v_mov_b32_e32 v26, s62 |
| ; VI-NEXT: v_mov_b32_e32 v27, s63 |
| ; VI-NEXT: v_mov_b32_e32 v28, s64 |
| ; VI-NEXT: v_mov_b32_e32 v29, s65 |
| ; VI-NEXT: v_mov_b32_e32 v30, s66 |
| ; VI-NEXT: v_mov_b32_e32 v31, s67 |
| ; VI-NEXT: .LBB11_5: ; %end |
| ; VI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; VI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; VI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; VI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; VI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; VI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; VI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; VI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; VI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; VI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; VI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; VI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10f64_to_v20i32_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_mov_b32 s49, s29 |
| ; GFX9-NEXT: s_mov_b32 s48, s28 |
| ; GFX9-NEXT: s_mov_b32 s47, s27 |
| ; GFX9-NEXT: s_mov_b32 s46, s26 |
| ; GFX9-NEXT: s_mov_b32 s45, s25 |
| ; GFX9-NEXT: s_mov_b32 s44, s24 |
| ; GFX9-NEXT: s_mov_b32 s43, s23 |
| ; GFX9-NEXT: s_mov_b32 s42, s22 |
| ; GFX9-NEXT: s_mov_b32 s41, s21 |
| ; GFX9-NEXT: s_mov_b32 s40, s20 |
| ; GFX9-NEXT: s_mov_b32 s39, s19 |
| ; GFX9-NEXT: s_mov_b32 s38, s18 |
| ; GFX9-NEXT: s_mov_b32 s37, s17 |
| ; GFX9-NEXT: s_mov_b32 s36, s16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s55, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s54, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s53, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s52, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s51, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s50, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB11_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB11_4 |
| ; GFX9-NEXT: .LBB11_2: ; %cmp.true |
| ; GFX9-NEXT: v_add_f64 v[18:19], s[54:55], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[16:17], s[52:53], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[14:15], s[50:51], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[12:13], s[48:49], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[10:11], s[46:47], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[8:9], s[44:45], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[6:7], s[42:43], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[4:5], s[40:41], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[2:3], s[38:39], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[0:1], s[36:37], 1.0 |
| ; GFX9-NEXT: s_branch .LBB11_5 |
| ; GFX9-NEXT: .LBB11_3: |
| ; GFX9-NEXT: s_branch .LBB11_2 |
| ; GFX9-NEXT: .LBB11_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB11_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v10f64_to_v20i32_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_xor_saveexec_b32 s4, -1 |
| ; GFX11-NEXT: scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill |
| ; GFX11-NEXT: s_mov_b32 exec_lo, s4 |
| ; GFX11-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX11-NEXT: s_mov_b32 s36, s0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s0, v2 |
| ; GFX11-NEXT: s_mov_b32 s47, s23 |
| ; GFX11-NEXT: s_mov_b32 s46, s22 |
| ; GFX11-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX11-NEXT: s_mov_b32 s45, s21 |
| ; GFX11-NEXT: s_mov_b32 s44, s20 |
| ; GFX11-NEXT: s_mov_b32 s43, s19 |
| ; GFX11-NEXT: s_mov_b32 s42, s18 |
| ; GFX11-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX11-NEXT: s_mov_b32 s41, s17 |
| ; GFX11-NEXT: s_mov_b32 s40, s16 |
| ; GFX11-NEXT: s_mov_b32 s38, s2 |
| ; GFX11-NEXT: s_mov_b32 s37, s1 |
| ; GFX11-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX11-NEXT: s_mov_b32 s39, s3 |
| ; GFX11-NEXT: s_cmp_lg_u32 s0, 0 |
| ; GFX11-NEXT: s_mov_b32 s0, 0 |
| ; GFX11-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX11-NEXT: s_mov_b32 s48, s24 |
| ; GFX11-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX11-NEXT: s_mov_b32 s49, s25 |
| ; GFX11-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX11-NEXT: s_mov_b32 s50, s26 |
| ; GFX11-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX11-NEXT: s_mov_b32 s51, s27 |
| ; GFX11-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX11-NEXT: s_mov_b32 s52, s28 |
| ; GFX11-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX11-NEXT: s_mov_b32 s53, s29 |
| ; GFX11-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX11-NEXT: v_readfirstlane_b32 s54, v0 |
| ; GFX11-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX11-NEXT: v_readfirstlane_b32 s55, v1 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB11_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB11_4 |
| ; GFX11-NEXT: .LBB11_2: ; %cmp.true |
| ; GFX11-NEXT: v_add_f64 v[18:19], s[54:55], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[16:17], s[52:53], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[14:15], s[50:51], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[12:13], s[48:49], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[10:11], s[46:47], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[8:9], s[44:45], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[6:7], s[42:43], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[4:5], s[40:41], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[2:3], s[38:39], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[0:1], s[36:37], 1.0 |
| ; GFX11-NEXT: s_branch .LBB11_5 |
| ; GFX11-NEXT: .LBB11_3: |
| ; GFX11-NEXT: s_branch .LBB11_2 |
| ; GFX11-NEXT: .LBB11_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67 |
| ; GFX11-NEXT: .LBB11_5: ; %end |
| ; GFX11-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX11-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX11-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX11-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX11-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX11-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX11-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX11-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX11-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX11-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX11-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX11-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX11-NEXT: s_xor_saveexec_b32 s0, -1 |
| ; GFX11-NEXT: scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload |
| ; GFX11-NEXT: s_mov_b32 exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; 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 <10 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <10 x double> %a1 to <20 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x double> %a to <20 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <20 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x i32> %phi |
| } |
| |
| define <40 x i16> @bitcast_v20i32_to_v40i16(<20 x i32> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v20i32_to_v40i16: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr31 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr29 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr26 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr25 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr24 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr23 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; implicit-def: $vgpr22 |
| ; SI-NEXT: ; implicit-def: $vgpr30 |
| ; SI-NEXT: ; implicit-def: $vgpr21 |
| ; SI-NEXT: ; implicit-def: $vgpr28 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr27 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB12_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v31, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v34, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB12_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB12_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1 |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13 |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v31, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v34, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB12_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v34 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v31 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v29 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v25 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v35 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v23 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v32 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v22 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v30 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v28 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v27 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v34 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v31 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v29 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v26 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v25 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v24 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20i32_to_v40i16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; 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: 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 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; 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 v19, vcc, 3, v19 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17 |
| ; 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_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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB12_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v20i32_to_v40i16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; 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 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; 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 v19, 3, v19 |
| ; GFX9-NEXT: v_add_u32_e32 v18, 3, v18 |
| ; GFX9-NEXT: v_add_u32_e32 v17, 3, v17 |
| ; 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_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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB12_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_mov_b32 s4, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v39, v0, s4 |
| ; GFX9-NEXT: v_perm_b32 v1, v38, v1, s4 |
| ; GFX9-NEXT: v_perm_b32 v2, v37, v2, s4 |
| ; GFX9-NEXT: v_perm_b32 v3, v36, v3, s4 |
| ; GFX9-NEXT: v_perm_b32 v4, v35, v4, s4 |
| ; GFX9-NEXT: v_perm_b32 v5, v34, v5, s4 |
| ; GFX9-NEXT: v_perm_b32 v6, v33, v6, s4 |
| ; GFX9-NEXT: v_perm_b32 v7, v32, v7, s4 |
| ; GFX9-NEXT: v_perm_b32 v8, v31, v8, s4 |
| ; GFX9-NEXT: v_perm_b32 v9, v30, v9, s4 |
| ; GFX9-NEXT: v_perm_b32 v10, v29, v10, s4 |
| ; GFX9-NEXT: v_perm_b32 v11, v28, v11, s4 |
| ; GFX9-NEXT: v_perm_b32 v12, v27, v12, s4 |
| ; GFX9-NEXT: v_perm_b32 v13, v26, v13, s4 |
| ; GFX9-NEXT: v_perm_b32 v14, v25, v14, s4 |
| ; GFX9-NEXT: v_perm_b32 v15, v24, v15, s4 |
| ; GFX9-NEXT: v_perm_b32 v16, v23, v16, s4 |
| ; GFX9-NEXT: v_perm_b32 v17, v22, v17, s4 |
| ; GFX9-NEXT: v_perm_b32 v18, v21, v18, s4 |
| ; GFX9-NEXT: v_perm_b32 v19, v20, v19, s4 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v20i32_to_v40i16: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB12_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v19, 3, v19 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v18, 3, v18 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v17, 3, v17 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v16, 3, v16 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v15, 3, v15 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v14, 3, v14 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v13, 3, v13 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v12, 3, v12 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v11, 3, v11 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v10, 3, v10 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, 3, v9 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 3, v8 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v7, 3, v7 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v6 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v5 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v4 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, 3, v3 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v2, 3, v2 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v1 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v0 |
| ; GFX11-TRUE16-NEXT: .LBB12_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v20i32_to_v40i16: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB12_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB12_2: ; %Flow |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB12_4 |
| ; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v19, 3, v19 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v18, 3, v18 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v17, 3, v17 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v16, 3, v16 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v15, 3, v15 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v14, 3, v14 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v13, 3, v13 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v12, 3, v12 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v11, 3, v11 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v10, 3, v10 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, 3, v9 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v8 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v7 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 3, v6 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 3, v5 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v4 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v3 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v2 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 3, v1 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB12_4: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v39, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v37, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v36, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v35, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v34, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v33, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v32, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v31, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v30, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v29, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v28, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v27, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v26, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v25, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v24, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v23, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v22, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v21, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v20, v19, 0x5040100 |
| ; GFX11-FAKE16-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 <20 x i32> %a, splat (i32 3) |
| %a2 = bitcast <20 x i32> %a1 to <40 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x i32> %a to <40 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <40 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x i16> %phi |
| } |
| |
| define inreg <40 x i16> @bitcast_v20i32_to_v40i16_scalar(<20 x i32> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v20i32_to_v40i16_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s8, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s5, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s8, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB13_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_lshr_b32 s72, s5, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s17, 16 |
| ; SI-NEXT: s_lshr_b64 s[10:11], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b64 s[12:13], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[14:15], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[40:41], s[28:29], 16 |
| ; SI-NEXT: s_lshr_b64 s[42:43], s[26:27], 16 |
| ; SI-NEXT: s_lshr_b64 s[44:45], s[24:25], 16 |
| ; SI-NEXT: s_lshr_b64 s[46:47], s[22:23], 16 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[20:21], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[18:19], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[16:17], 16 |
| ; SI-NEXT: s_cbranch_execnz .LBB13_3 |
| ; SI-NEXT: .LBB13_2: ; %cmp.true |
| ; SI-NEXT: s_add_i32 s17, s17, 3 |
| ; SI-NEXT: s_add_i32 s16, s16, 3 |
| ; SI-NEXT: s_add_i32 s19, s19, 3 |
| ; SI-NEXT: s_add_i32 s18, s18, 3 |
| ; SI-NEXT: s_add_i32 s21, s21, 3 |
| ; SI-NEXT: s_add_i32 s20, s20, 3 |
| ; SI-NEXT: s_add_i32 s23, s23, 3 |
| ; SI-NEXT: s_add_i32 s22, s22, 3 |
| ; SI-NEXT: s_add_i32 s25, s25, 3 |
| ; SI-NEXT: s_add_i32 s24, s24, 3 |
| ; SI-NEXT: s_add_i32 s27, s27, 3 |
| ; SI-NEXT: s_add_i32 s26, s26, 3 |
| ; SI-NEXT: s_add_i32 s29, s29, 3 |
| ; SI-NEXT: s_add_i32 s28, s28, 3 |
| ; SI-NEXT: s_add_i32 s9, s9, 3 |
| ; SI-NEXT: s_add_i32 s8, s8, 3 |
| ; SI-NEXT: s_add_i32 s7, s7, 3 |
| ; SI-NEXT: s_add_i32 s6, s6, 3 |
| ; SI-NEXT: s_add_i32 s5, s5, 3 |
| ; SI-NEXT: s_add_i32 s4, s4, 3 |
| ; SI-NEXT: s_lshr_b64 s[10:11], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b64 s[12:13], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[14:15], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[40:41], s[28:29], 16 |
| ; SI-NEXT: s_lshr_b64 s[42:43], s[26:27], 16 |
| ; SI-NEXT: s_lshr_b64 s[44:45], s[24:25], 16 |
| ; SI-NEXT: s_lshr_b64 s[46:47], s[22:23], 16 |
| ; SI-NEXT: s_lshr_b32 s72, s5, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s17, 16 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[20:21], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[18:19], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[16:17], 16 |
| ; SI-NEXT: .LBB13_3: ; %end |
| ; SI-NEXT: s_lshl_b32 s11, s60, 16 |
| ; SI-NEXT: s_and_b32 s13, s16, 0xffff |
| ; SI-NEXT: s_or_b32 s11, s13, s11 |
| ; SI-NEXT: s_and_b32 s13, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s15, s89, 16 |
| ; SI-NEXT: s_or_b32 s13, s13, s15 |
| ; SI-NEXT: s_lshl_b32 s15, s58, 16 |
| ; SI-NEXT: s_and_b32 s16, s18, 0xffff |
| ; SI-NEXT: s_or_b32 s15, s16, s15 |
| ; SI-NEXT: s_and_b32 s16, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s17, s88, 16 |
| ; SI-NEXT: s_or_b32 s16, s16, s17 |
| ; SI-NEXT: s_lshl_b32 s17, s56, 16 |
| ; SI-NEXT: s_and_b32 s18, s20, 0xffff |
| ; SI-NEXT: s_or_b32 s17, s18, s17 |
| ; SI-NEXT: s_and_b32 s18, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s19, s79, 16 |
| ; SI-NEXT: s_or_b32 s18, s18, s19 |
| ; SI-NEXT: s_and_b32 s19, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s20, s46, 16 |
| ; SI-NEXT: s_or_b32 s19, s19, s20 |
| ; SI-NEXT: s_and_b32 s20, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s21, s78, 16 |
| ; SI-NEXT: s_or_b32 s20, s20, s21 |
| ; SI-NEXT: s_and_b32 s21, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s22, s44, 16 |
| ; SI-NEXT: s_or_b32 s21, s21, s22 |
| ; SI-NEXT: s_and_b32 s22, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s23, s77, 16 |
| ; SI-NEXT: s_or_b32 s22, s22, s23 |
| ; SI-NEXT: s_and_b32 s23, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s42, 16 |
| ; SI-NEXT: s_or_b32 s23, s23, s24 |
| ; SI-NEXT: s_and_b32 s24, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s25, s76, 16 |
| ; SI-NEXT: s_or_b32 s24, s24, s25 |
| ; SI-NEXT: s_and_b32 s25, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s26, s40, 16 |
| ; SI-NEXT: s_and_b32 s8, s8, 0xffff |
| ; SI-NEXT: s_lshl_b32 s14, s14, 16 |
| ; SI-NEXT: s_and_b32 s6, s6, 0xffff |
| ; SI-NEXT: s_lshl_b32 s12, s12, 16 |
| ; SI-NEXT: s_and_b32 s4, s4, 0xffff |
| ; SI-NEXT: s_lshl_b32 s10, s10, 16 |
| ; SI-NEXT: s_or_b32 s25, s25, s26 |
| ; SI-NEXT: s_and_b32 s26, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s27, s75, 16 |
| ; SI-NEXT: s_or_b32 s8, s8, s14 |
| ; SI-NEXT: s_and_b32 s9, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s14, s74, 16 |
| ; SI-NEXT: s_or_b32 s6, s6, s12 |
| ; SI-NEXT: s_and_b32 s7, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s12, s73, 16 |
| ; SI-NEXT: s_or_b32 s4, s4, s10 |
| ; SI-NEXT: s_and_b32 s5, s5, 0xffff |
| ; SI-NEXT: s_lshl_b32 s10, s72, 16 |
| ; SI-NEXT: s_or_b32 s26, s26, s27 |
| ; SI-NEXT: s_or_b32 s9, s9, s14 |
| ; SI-NEXT: s_or_b32 s7, s7, s12 |
| ; SI-NEXT: s_or_b32 s5, s5, s10 |
| ; SI-NEXT: v_mov_b32_e32 v0, s11 |
| ; SI-NEXT: v_mov_b32_e32 v1, s13 |
| ; SI-NEXT: v_mov_b32_e32 v2, s15 |
| ; SI-NEXT: v_mov_b32_e32 v3, s16 |
| ; SI-NEXT: v_mov_b32_e32 v4, s17 |
| ; SI-NEXT: v_mov_b32_e32 v5, s18 |
| ; SI-NEXT: v_mov_b32_e32 v6, s19 |
| ; SI-NEXT: v_mov_b32_e32 v7, s20 |
| ; SI-NEXT: v_mov_b32_e32 v8, s21 |
| ; SI-NEXT: v_mov_b32_e32 v9, s22 |
| ; SI-NEXT: v_mov_b32_e32 v10, s23 |
| ; SI-NEXT: v_mov_b32_e32 v11, s24 |
| ; SI-NEXT: v_mov_b32_e32 v12, s25 |
| ; SI-NEXT: v_mov_b32_e32 v13, s26 |
| ; SI-NEXT: v_mov_b32_e32 v14, s8 |
| ; SI-NEXT: v_mov_b32_e32 v15, s9 |
| ; SI-NEXT: v_mov_b32_e32 v16, s6 |
| ; SI-NEXT: v_mov_b32_e32 v17, s7 |
| ; SI-NEXT: v_mov_b32_e32 v18, s4 |
| ; SI-NEXT: v_mov_b32_e32 v19, s5 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB13_4: |
| ; SI-NEXT: ; implicit-def: $sgpr60 |
| ; SI-NEXT: ; implicit-def: $sgpr89 |
| ; SI-NEXT: ; implicit-def: $sgpr58 |
| ; SI-NEXT: ; implicit-def: $sgpr88 |
| ; SI-NEXT: ; implicit-def: $sgpr56 |
| ; SI-NEXT: ; implicit-def: $sgpr79 |
| ; SI-NEXT: ; implicit-def: $sgpr46 |
| ; SI-NEXT: ; implicit-def: $sgpr78 |
| ; SI-NEXT: ; implicit-def: $sgpr44 |
| ; SI-NEXT: ; implicit-def: $sgpr77 |
| ; SI-NEXT: ; implicit-def: $sgpr42 |
| ; SI-NEXT: ; implicit-def: $sgpr76 |
| ; SI-NEXT: ; implicit-def: $sgpr40 |
| ; SI-NEXT: ; implicit-def: $sgpr75 |
| ; SI-NEXT: ; implicit-def: $sgpr14 |
| ; SI-NEXT: ; implicit-def: $sgpr74 |
| ; SI-NEXT: ; implicit-def: $sgpr12 |
| ; SI-NEXT: ; implicit-def: $sgpr73 |
| ; SI-NEXT: ; implicit-def: $sgpr10 |
| ; SI-NEXT: ; implicit-def: $sgpr72 |
| ; SI-NEXT: s_branch .LBB13_2 |
| ; |
| ; VI-LABEL: bitcast_v20i32_to_v40i16_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB13_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_lshr_b32 s12, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s10, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s46, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: s_cbranch_execnz .LBB13_3 |
| ; VI-NEXT: .LBB13_2: ; %cmp.true |
| ; VI-NEXT: s_add_i32 s6, s6, 3 |
| ; VI-NEXT: s_add_i32 s7, s7, 3 |
| ; VI-NEXT: s_add_i32 s8, s8, 3 |
| ; VI-NEXT: s_add_i32 s9, s9, 3 |
| ; VI-NEXT: s_add_i32 s10, s10, 3 |
| ; VI-NEXT: s_add_i32 s11, s11, 3 |
| ; VI-NEXT: s_add_i32 s29, s29, 3 |
| ; VI-NEXT: s_add_i32 s28, s28, 3 |
| ; VI-NEXT: s_add_i32 s27, s27, 3 |
| ; VI-NEXT: s_add_i32 s26, s26, 3 |
| ; VI-NEXT: s_add_i32 s25, s25, 3 |
| ; VI-NEXT: s_add_i32 s24, s24, 3 |
| ; VI-NEXT: s_add_i32 s23, s23, 3 |
| ; VI-NEXT: s_add_i32 s22, s22, 3 |
| ; VI-NEXT: s_add_i32 s21, s21, 3 |
| ; VI-NEXT: s_add_i32 s20, s20, 3 |
| ; VI-NEXT: s_add_i32 s19, s19, 3 |
| ; VI-NEXT: s_add_i32 s18, s18, 3 |
| ; VI-NEXT: s_add_i32 s17, s17, 3 |
| ; VI-NEXT: s_add_i32 s16, s16, 3 |
| ; VI-NEXT: s_lshr_b32 s12, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s10, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s46, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: .LBB13_3: ; %end |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s63, 16 |
| ; VI-NEXT: s_or_b32 s4, s4, s5 |
| ; VI-NEXT: s_and_b32 s5, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s16, s62, 16 |
| ; VI-NEXT: s_or_b32 s5, s5, s16 |
| ; VI-NEXT: s_and_b32 s16, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s17, s61, 16 |
| ; VI-NEXT: s_or_b32 s16, s16, s17 |
| ; VI-NEXT: s_and_b32 s17, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s18, s60, 16 |
| ; VI-NEXT: s_or_b32 s17, s17, s18 |
| ; VI-NEXT: s_and_b32 s18, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s19, s59, 16 |
| ; VI-NEXT: s_or_b32 s18, s18, s19 |
| ; VI-NEXT: s_and_b32 s19, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s20, s58, 16 |
| ; VI-NEXT: s_or_b32 s19, s19, s20 |
| ; VI-NEXT: s_and_b32 s20, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s21, s57, 16 |
| ; VI-NEXT: s_or_b32 s20, s20, s21 |
| ; VI-NEXT: s_and_b32 s21, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s22, s56, 16 |
| ; VI-NEXT: s_or_b32 s21, s21, s22 |
| ; VI-NEXT: s_and_b32 s22, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s23, s47, 16 |
| ; VI-NEXT: s_or_b32 s22, s22, s23 |
| ; VI-NEXT: s_and_b32 s23, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s24, s46, 16 |
| ; VI-NEXT: s_or_b32 s23, s23, s24 |
| ; VI-NEXT: s_and_b32 s24, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s25, s45, 16 |
| ; VI-NEXT: s_or_b32 s24, s24, s25 |
| ; VI-NEXT: s_and_b32 s25, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s26, s44, 16 |
| ; VI-NEXT: s_or_b32 s25, s25, s26 |
| ; VI-NEXT: s_and_b32 s26, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s27, s43, 16 |
| ; VI-NEXT: s_or_b32 s26, s26, s27 |
| ; VI-NEXT: s_and_b32 s27, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s28, s42, 16 |
| ; VI-NEXT: s_or_b32 s27, s27, s28 |
| ; VI-NEXT: s_and_b32 s11, 0xffff, s11 |
| ; VI-NEXT: s_lshl_b32 s28, s41, 16 |
| ; VI-NEXT: s_or_b32 s11, s11, s28 |
| ; VI-NEXT: s_and_b32 s10, 0xffff, s10 |
| ; VI-NEXT: s_lshl_b32 s28, s40, 16 |
| ; VI-NEXT: s_and_b32 s9, 0xffff, s9 |
| ; VI-NEXT: s_lshl_b32 s15, s15, 16 |
| ; VI-NEXT: s_and_b32 s8, 0xffff, s8 |
| ; VI-NEXT: s_lshl_b32 s14, s14, 16 |
| ; VI-NEXT: s_and_b32 s7, 0xffff, s7 |
| ; VI-NEXT: s_lshl_b32 s13, s13, 16 |
| ; VI-NEXT: s_and_b32 s6, 0xffff, s6 |
| ; VI-NEXT: s_lshl_b32 s12, s12, 16 |
| ; VI-NEXT: s_or_b32 s10, s10, s28 |
| ; VI-NEXT: s_or_b32 s9, s9, s15 |
| ; VI-NEXT: s_or_b32 s8, s8, s14 |
| ; VI-NEXT: s_or_b32 s7, s7, s13 |
| ; VI-NEXT: s_or_b32 s6, s6, s12 |
| ; VI-NEXT: v_mov_b32_e32 v0, s4 |
| ; VI-NEXT: v_mov_b32_e32 v1, s5 |
| ; VI-NEXT: v_mov_b32_e32 v2, s16 |
| ; VI-NEXT: v_mov_b32_e32 v3, s17 |
| ; VI-NEXT: v_mov_b32_e32 v4, s18 |
| ; VI-NEXT: v_mov_b32_e32 v5, s19 |
| ; VI-NEXT: v_mov_b32_e32 v6, s20 |
| ; VI-NEXT: v_mov_b32_e32 v7, s21 |
| ; VI-NEXT: v_mov_b32_e32 v8, s22 |
| ; VI-NEXT: v_mov_b32_e32 v9, s23 |
| ; VI-NEXT: v_mov_b32_e32 v10, s24 |
| ; VI-NEXT: v_mov_b32_e32 v11, s25 |
| ; VI-NEXT: v_mov_b32_e32 v12, s26 |
| ; VI-NEXT: v_mov_b32_e32 v13, s27 |
| ; VI-NEXT: v_mov_b32_e32 v14, s11 |
| ; VI-NEXT: v_mov_b32_e32 v15, s10 |
| ; VI-NEXT: v_mov_b32_e32 v16, s9 |
| ; VI-NEXT: v_mov_b32_e32 v17, s8 |
| ; VI-NEXT: v_mov_b32_e32 v18, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB13_4: |
| ; VI-NEXT: ; implicit-def: $sgpr63 |
| ; VI-NEXT: ; implicit-def: $sgpr62 |
| ; VI-NEXT: ; implicit-def: $sgpr61 |
| ; VI-NEXT: ; implicit-def: $sgpr60 |
| ; VI-NEXT: ; implicit-def: $sgpr59 |
| ; VI-NEXT: ; implicit-def: $sgpr58 |
| ; VI-NEXT: ; implicit-def: $sgpr57 |
| ; VI-NEXT: ; implicit-def: $sgpr56 |
| ; VI-NEXT: ; implicit-def: $sgpr47 |
| ; VI-NEXT: ; implicit-def: $sgpr46 |
| ; VI-NEXT: ; implicit-def: $sgpr45 |
| ; VI-NEXT: ; implicit-def: $sgpr44 |
| ; VI-NEXT: ; implicit-def: $sgpr43 |
| ; VI-NEXT: ; implicit-def: $sgpr42 |
| ; VI-NEXT: ; implicit-def: $sgpr41 |
| ; VI-NEXT: ; implicit-def: $sgpr40 |
| ; VI-NEXT: ; implicit-def: $sgpr15 |
| ; VI-NEXT: ; implicit-def: $sgpr14 |
| ; VI-NEXT: ; implicit-def: $sgpr13 |
| ; VI-NEXT: ; implicit-def: $sgpr12 |
| ; VI-NEXT: s_branch .LBB13_2 |
| ; |
| ; GFX9-LABEL: bitcast_v20i32_to_v40i16_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s10, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s11, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB13_4 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_lshr_b32 s12, s6, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s7, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s8, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s9, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s10, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s11, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s43, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s58, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s60, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s62, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s16, 16 |
| ; GFX9-NEXT: s_cbranch_execnz .LBB13_3 |
| ; GFX9-NEXT: .LBB13_2: ; %cmp.true |
| ; GFX9-NEXT: s_add_i32 s6, s6, 3 |
| ; GFX9-NEXT: s_add_i32 s7, s7, 3 |
| ; GFX9-NEXT: s_add_i32 s8, s8, 3 |
| ; GFX9-NEXT: s_add_i32 s9, s9, 3 |
| ; GFX9-NEXT: s_add_i32 s10, s10, 3 |
| ; GFX9-NEXT: s_add_i32 s11, s11, 3 |
| ; GFX9-NEXT: s_add_i32 s29, s29, 3 |
| ; GFX9-NEXT: s_add_i32 s28, s28, 3 |
| ; GFX9-NEXT: s_add_i32 s27, s27, 3 |
| ; GFX9-NEXT: s_add_i32 s26, s26, 3 |
| ; GFX9-NEXT: s_add_i32 s25, s25, 3 |
| ; GFX9-NEXT: s_add_i32 s24, s24, 3 |
| ; GFX9-NEXT: s_add_i32 s23, s23, 3 |
| ; GFX9-NEXT: s_add_i32 s22, s22, 3 |
| ; GFX9-NEXT: s_add_i32 s21, s21, 3 |
| ; GFX9-NEXT: s_add_i32 s20, s20, 3 |
| ; GFX9-NEXT: s_add_i32 s19, s19, 3 |
| ; GFX9-NEXT: s_add_i32 s18, s18, 3 |
| ; GFX9-NEXT: s_add_i32 s17, s17, 3 |
| ; GFX9-NEXT: s_add_i32 s16, s16, 3 |
| ; GFX9-NEXT: s_lshr_b32 s12, s6, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s7, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s8, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s9, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s10, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s11, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s43, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s58, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s60, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s62, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s16, 16 |
| ; GFX9-NEXT: .LBB13_3: ; %end |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s63 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s5, s17, s62 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s61 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s60 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s59 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s58 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s57 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s56 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s47 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s46 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s45 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s44 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s43 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s42 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s11, s11, s41 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s10, s10, s40 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s9, s9, s15 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s14 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s13 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s12 |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s5 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s6 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; GFX9-NEXT: .LBB13_4: |
| ; GFX9-NEXT: ; implicit-def: $sgpr63 |
| ; GFX9-NEXT: ; implicit-def: $sgpr62 |
| ; GFX9-NEXT: ; implicit-def: $sgpr61 |
| ; GFX9-NEXT: ; implicit-def: $sgpr60 |
| ; GFX9-NEXT: ; implicit-def: $sgpr59 |
| ; GFX9-NEXT: ; implicit-def: $sgpr58 |
| ; GFX9-NEXT: ; implicit-def: $sgpr57 |
| ; GFX9-NEXT: ; implicit-def: $sgpr56 |
| ; GFX9-NEXT: ; implicit-def: $sgpr47 |
| ; GFX9-NEXT: ; implicit-def: $sgpr46 |
| ; GFX9-NEXT: ; implicit-def: $sgpr45 |
| ; GFX9-NEXT: ; implicit-def: $sgpr44 |
| ; GFX9-NEXT: ; implicit-def: $sgpr43 |
| ; GFX9-NEXT: ; implicit-def: $sgpr42 |
| ; GFX9-NEXT: ; implicit-def: $sgpr41 |
| ; GFX9-NEXT: ; implicit-def: $sgpr40 |
| ; GFX9-NEXT: ; implicit-def: $sgpr15 |
| ; GFX9-NEXT: ; implicit-def: $sgpr14 |
| ; GFX9-NEXT: ; implicit-def: $sgpr13 |
| ; GFX9-NEXT: ; implicit-def: $sgpr12 |
| ; GFX9-NEXT: s_branch .LBB13_2 |
| ; |
| ; GFX11-LABEL: bitcast_v20i32_to_v40i16_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-NEXT: s_mov_b32 s58, 0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB13_4 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-NEXT: s_lshr_b32 s6, s4, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s5, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s40, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s41, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s45, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s46, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s56, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s58 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB13_3 |
| ; GFX11-NEXT: .LBB13_2: ; %cmp.true |
| ; GFX11-NEXT: s_add_i32 s4, s4, 3 |
| ; GFX11-NEXT: s_add_i32 s5, s5, 3 |
| ; GFX11-NEXT: s_add_i32 s29, s29, 3 |
| ; GFX11-NEXT: s_add_i32 s28, s28, 3 |
| ; GFX11-NEXT: s_add_i32 s27, s27, 3 |
| ; GFX11-NEXT: s_add_i32 s26, s26, 3 |
| ; GFX11-NEXT: s_add_i32 s25, s25, 3 |
| ; GFX11-NEXT: s_add_i32 s24, s24, 3 |
| ; GFX11-NEXT: s_add_i32 s23, s23, 3 |
| ; GFX11-NEXT: s_add_i32 s22, s22, 3 |
| ; GFX11-NEXT: s_add_i32 s21, s21, 3 |
| ; GFX11-NEXT: s_add_i32 s20, s20, 3 |
| ; GFX11-NEXT: s_add_i32 s19, s19, 3 |
| ; GFX11-NEXT: s_add_i32 s18, s18, 3 |
| ; GFX11-NEXT: s_add_i32 s17, s17, 3 |
| ; GFX11-NEXT: s_add_i32 s16, s16, 3 |
| ; GFX11-NEXT: s_add_i32 s3, s3, 3 |
| ; GFX11-NEXT: s_add_i32 s2, s2, 3 |
| ; GFX11-NEXT: s_add_i32 s1, s1, 3 |
| ; GFX11-NEXT: s_add_i32 s0, s0, 3 |
| ; GFX11-NEXT: s_lshr_b32 s6, s4, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s5, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s40, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s41, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s45, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s46, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s56, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: .LBB13_3: ; %end |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s57 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s56 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s47 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s46 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s45 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s44 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s43 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s42 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s41 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s40 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s15, s22, s15 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s14, s23, s14 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s13, s24, s13 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s12, s25, s12 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s11, s26, s11 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s10, s27, s10 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s9, s28, s9 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s8, s29, s8 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s7 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s6 |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB13_4: |
| ; GFX11-NEXT: ; implicit-def: $sgpr57 |
| ; GFX11-NEXT: ; implicit-def: $sgpr56 |
| ; GFX11-NEXT: ; implicit-def: $sgpr47 |
| ; GFX11-NEXT: ; implicit-def: $sgpr46 |
| ; GFX11-NEXT: ; implicit-def: $sgpr45 |
| ; GFX11-NEXT: ; implicit-def: $sgpr44 |
| ; GFX11-NEXT: ; implicit-def: $sgpr43 |
| ; GFX11-NEXT: ; implicit-def: $sgpr42 |
| ; GFX11-NEXT: ; implicit-def: $sgpr41 |
| ; GFX11-NEXT: ; implicit-def: $sgpr40 |
| ; GFX11-NEXT: ; implicit-def: $sgpr15 |
| ; GFX11-NEXT: ; implicit-def: $sgpr14 |
| ; GFX11-NEXT: ; implicit-def: $sgpr13 |
| ; GFX11-NEXT: ; implicit-def: $sgpr12 |
| ; GFX11-NEXT: ; implicit-def: $sgpr11 |
| ; GFX11-NEXT: ; implicit-def: $sgpr10 |
| ; GFX11-NEXT: ; implicit-def: $sgpr9 |
| ; GFX11-NEXT: ; implicit-def: $sgpr8 |
| ; GFX11-NEXT: ; implicit-def: $sgpr7 |
| ; GFX11-NEXT: ; implicit-def: $sgpr6 |
| ; GFX11-NEXT: s_branch .LBB13_2 |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <20 x i32> %a, splat (i32 3) |
| %a2 = bitcast <20 x i32> %a1 to <40 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x i32> %a to <40 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <40 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x i16> %phi |
| } |
| |
| define <20 x i32> @bitcast_v40i16_to_v20i32(<40 x i16> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v40i16_to_v20i32: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: v_mov_b32_e32 v32, v19 |
| ; SI-NEXT: v_mov_b32_e32 v33, v18 |
| ; SI-NEXT: v_mov_b32_e32 v34, v17 |
| ; SI-NEXT: v_mov_b32_e32 v35, v16 |
| ; SI-NEXT: v_mov_b32_e32 v36, v15 |
| ; SI-NEXT: v_mov_b32_e32 v37, v14 |
| ; SI-NEXT: v_mov_b32_e32 v38, v13 |
| ; SI-NEXT: v_mov_b32_e32 v39, v12 |
| ; SI-NEXT: v_mov_b32_e32 v48, v11 |
| ; SI-NEXT: v_mov_b32_e32 v49, v10 |
| ; SI-NEXT: v_mov_b32_e32 v50, v9 |
| ; SI-NEXT: v_mov_b32_e32 v51, v8 |
| ; SI-NEXT: v_mov_b32_e32 v52, v7 |
| ; SI-NEXT: v_mov_b32_e32 v53, v6 |
| ; SI-NEXT: v_mov_b32_e32 v54, v5 |
| ; SI-NEXT: v_mov_b32_e32 v55, v4 |
| ; SI-NEXT: v_mov_b32_e32 v40, v3 |
| ; SI-NEXT: v_mov_b32_e32 v41, v2 |
| ; SI-NEXT: v_mov_b32_e32 v42, v1 |
| ; SI-NEXT: v_mov_b32_e32 v43, v0 |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v33 |
| ; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v34 |
| ; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v35 |
| ; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v36 |
| ; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v37 |
| ; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v38 |
| ; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v39 |
| ; SI-NEXT: v_lshrrev_b32_e32 v8, 16, v48 |
| ; SI-NEXT: v_lshrrev_b32_e32 v9, 16, v49 |
| ; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v50 |
| ; SI-NEXT: v_lshrrev_b32_e32 v11, 16, v51 |
| ; SI-NEXT: v_lshrrev_b32_e32 v12, 16, v52 |
| ; SI-NEXT: v_lshrrev_b32_e32 v13, 16, v53 |
| ; SI-NEXT: v_lshrrev_b32_e32 v14, 16, v54 |
| ; SI-NEXT: v_lshrrev_b32_e32 v15, 16, v55 |
| ; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v40 |
| ; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v41 |
| ; SI-NEXT: v_lshrrev_b32_e32 v18, 16, v42 |
| ; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v43 |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_waitcnt expcnt(4) |
| ; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v56, 16, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v47, 16, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v12 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB14_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v39 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v43 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v42 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v41 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v40 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v55 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v54 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v53 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v52 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v51 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v50 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v49 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v48 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v59 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v58 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v57 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v56 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v47 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v46 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v45 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v44 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v63 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v62 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v61 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v60 |
| ; SI-NEXT: ; implicit-def: $vgpr43 |
| ; SI-NEXT: ; implicit-def: $vgpr42 |
| ; SI-NEXT: ; implicit-def: $vgpr41 |
| ; SI-NEXT: ; implicit-def: $vgpr40 |
| ; SI-NEXT: ; implicit-def: $vgpr55 |
| ; SI-NEXT: ; implicit-def: $vgpr54 |
| ; SI-NEXT: ; implicit-def: $vgpr53 |
| ; SI-NEXT: ; implicit-def: $vgpr52 |
| ; SI-NEXT: ; implicit-def: $vgpr51 |
| ; SI-NEXT: ; implicit-def: $vgpr50 |
| ; SI-NEXT: ; implicit-def: $vgpr49 |
| ; SI-NEXT: ; implicit-def: $vgpr48 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr59 |
| ; SI-NEXT: ; implicit-def: $vgpr58 |
| ; SI-NEXT: ; implicit-def: $vgpr57 |
| ; SI-NEXT: ; implicit-def: $vgpr56 |
| ; SI-NEXT: ; implicit-def: $vgpr47 |
| ; SI-NEXT: ; implicit-def: $vgpr46 |
| ; SI-NEXT: ; implicit-def: $vgpr45 |
| ; SI-NEXT: ; implicit-def: $vgpr44 |
| ; SI-NEXT: ; implicit-def: $vgpr63 |
| ; SI-NEXT: ; implicit-def: $vgpr62 |
| ; SI-NEXT: ; implicit-def: $vgpr61 |
| ; SI-NEXT: ; implicit-def: $vgpr60 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v38 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v37 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v36 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v35 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v34 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v33 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v19 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v32 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: .LBB14_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB14_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v39 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v43 |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v42 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v41 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v40 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v55 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v54 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v53 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v52 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v51 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v50 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v49 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v48 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_or_b32_e32 v0, v59, v0 |
| ; SI-NEXT: s_mov_b32 s6, 0x30000 |
| ; SI-NEXT: v_or_b32_e32 v1, v58, v1 |
| ; SI-NEXT: v_or_b32_e32 v2, v57, v2 |
| ; SI-NEXT: v_or_b32_e32 v3, v56, v3 |
| ; SI-NEXT: v_or_b32_e32 v4, v47, v4 |
| ; SI-NEXT: v_or_b32_e32 v5, v46, v5 |
| ; SI-NEXT: v_or_b32_e32 v6, v45, v6 |
| ; SI-NEXT: v_or_b32_e32 v7, v44, v7 |
| ; SI-NEXT: v_or_b32_e32 v8, v63, v8 |
| ; SI-NEXT: v_or_b32_e32 v9, v62, v9 |
| ; SI-NEXT: v_or_b32_e32 v10, v61, v10 |
| ; SI-NEXT: v_or_b32_e32 v11, v60, v11 |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0 |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, s6, v3 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, s6, v5 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, s6, v7 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v38 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_or_b32_e32 v13, v14, v13 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v37 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_or_b32_e32 v14, v15, v14 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v36 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v35 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v34 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_or_b32_e32 v17, v18, v17 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v33 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_or_b32_e32 v18, v19, v18 |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v32 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_or_b32_e32 v19, v20, v19 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, s6, v11 |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, s6, v13 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v14 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, s6, v15 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, s6, v16 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, s6, v17 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v18 |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 0x30000, v19 |
| ; SI-NEXT: .LBB14_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40i16_to_v20i32: |
| ; 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:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: v_mov_b32_e32 v32, v19 |
| ; VI-NEXT: v_mov_b32_e32 v33, v18 |
| ; VI-NEXT: v_mov_b32_e32 v34, v17 |
| ; VI-NEXT: v_mov_b32_e32 v35, v16 |
| ; VI-NEXT: v_mov_b32_e32 v36, v15 |
| ; VI-NEXT: v_mov_b32_e32 v37, v14 |
| ; VI-NEXT: v_mov_b32_e32 v38, v13 |
| ; VI-NEXT: v_mov_b32_e32 v39, v12 |
| ; VI-NEXT: v_mov_b32_e32 v48, v11 |
| ; VI-NEXT: v_mov_b32_e32 v49, v10 |
| ; VI-NEXT: v_mov_b32_e32 v50, v9 |
| ; VI-NEXT: v_mov_b32_e32 v51, v8 |
| ; VI-NEXT: v_mov_b32_e32 v52, v7 |
| ; VI-NEXT: v_mov_b32_e32 v53, v6 |
| ; VI-NEXT: v_mov_b32_e32 v54, v5 |
| ; VI-NEXT: v_mov_b32_e32 v55, v4 |
| ; VI-NEXT: v_mov_b32_e32 v40, v3 |
| ; VI-NEXT: v_mov_b32_e32 v41, v2 |
| ; VI-NEXT: v_mov_b32_e32 v42, v1 |
| ; VI-NEXT: v_mov_b32_e32 v43, v0 |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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 .LBB14_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_mov_b32_e32 v19, 16 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v0, v19, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v1, v19, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v2, v19, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v3, v19, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v4, v19, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v5, v19, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v6, v19, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v7, v19, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v8, v19, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v9, v19, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v10, v19, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v11, v19, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v12, v19, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v13, v19, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v14, v19, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v15, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v16, v19, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v17, v19, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v18, v19, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v19, v19, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v40, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v55, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v54, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v53, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v52, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v51, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v50, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v49, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v48, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v39, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v38, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v37, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v36, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v35, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v34, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v33, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v32, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; 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: .LBB14_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB14_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v19, 3 |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v43 |
| ; VI-NEXT: v_add_u16_sdwa v1, v43, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v42 |
| ; VI-NEXT: v_add_u16_sdwa v2, v42, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v1, v1, v2 |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v41 |
| ; VI-NEXT: v_add_u16_sdwa v3, v41, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v40 |
| ; VI-NEXT: v_add_u16_sdwa v4, v40, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v55 |
| ; VI-NEXT: v_add_u16_sdwa v5, v55, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v54 |
| ; VI-NEXT: v_add_u16_sdwa v6, v54, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v53 |
| ; VI-NEXT: v_add_u16_sdwa v7, v53, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v52 |
| ; VI-NEXT: v_add_u16_sdwa v8, v52, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v51 |
| ; VI-NEXT: v_add_u16_sdwa v9, v51, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v50 |
| ; VI-NEXT: v_add_u16_sdwa v10, v50, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v49 |
| ; VI-NEXT: v_add_u16_sdwa v11, v49, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v48 |
| ; VI-NEXT: v_add_u16_sdwa v12, v48, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v39 |
| ; VI-NEXT: v_add_u16_sdwa v13, v39, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v38 |
| ; VI-NEXT: v_add_u16_sdwa v14, v38, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v37 |
| ; VI-NEXT: v_add_u16_sdwa v15, v37, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; VI-NEXT: v_add_u16_e32 v15, 3, v36 |
| ; VI-NEXT: v_add_u16_sdwa v16, v36, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v35 |
| ; VI-NEXT: v_add_u16_sdwa v17, v35, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; VI-NEXT: v_add_u16_e32 v17, 3, v34 |
| ; VI-NEXT: v_add_u16_sdwa v18, v34, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; VI-NEXT: v_add_u16_e32 v18, 3, v33 |
| ; VI-NEXT: v_add_u16_sdwa v20, v33, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; VI-NEXT: v_add_u16_e32 v20, 3, v32 |
| ; VI-NEXT: v_add_u16_sdwa v19, v32, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v19, v20, v19 |
| ; VI-NEXT: .LBB14_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40i16_to_v20i32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_mov_b32_e32 v32, v19 |
| ; 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: v_mov_b32_e32 v33, v18 |
| ; GFX9-NEXT: v_mov_b32_e32 v43, v0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, v17 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, v16 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, v15 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, v14 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, v13 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, v12 |
| ; GFX9-NEXT: v_mov_b32_e32 v48, v11 |
| ; GFX9-NEXT: v_mov_b32_e32 v49, v10 |
| ; GFX9-NEXT: v_mov_b32_e32 v50, v9 |
| ; GFX9-NEXT: v_mov_b32_e32 v51, v8 |
| ; GFX9-NEXT: v_mov_b32_e32 v52, v7 |
| ; GFX9-NEXT: v_mov_b32_e32 v53, v6 |
| ; GFX9-NEXT: v_mov_b32_e32 v54, v5 |
| ; GFX9-NEXT: v_mov_b32_e32 v55, v4 |
| ; GFX9-NEXT: v_mov_b32_e32 v40, v3 |
| ; GFX9-NEXT: v_mov_b32_e32 v41, v2 |
| ; GFX9-NEXT: v_mov_b32_e32 v42, v1 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v48 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v49 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v50 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v51 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v52 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v53 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v54 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v55 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43 |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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 .LBB14_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v35 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v34 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v33 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v32 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; 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: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: .LBB14_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB14_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_pk_add_u16 v0, v0, 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 v2, v2, 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 v4, v4, 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 v6, v6, 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 v8, v8, 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 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; GFX9-NEXT: v_pk_add_u16 v12, v12, 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 v14, v14, 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 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB14_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; 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-TRUE16-LABEL: bitcast_v40i16_to_v20i32: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB14_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: .LBB14_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v40i16_to_v20i32: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v37, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v39, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v48, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v36, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v35, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v34, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v33, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v32, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v31, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v30, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v29, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v28, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v27, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v26, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v25, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v24, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v23, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v22, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v21, v19, 0x5040100 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB14_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: .LBB14_2: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-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 <40 x i16> %a, splat (i16 3) |
| %a2 = bitcast <40 x i16> %a1 to <20 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x i16> %a to <20 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <20 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x i32> %phi |
| } |
| |
| define inreg <20 x i32> @bitcast_v40i16_to_v20i32_scalar(<40 x i16> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v40i16_to_v20i32_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v20, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v20, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v20, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v20, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v20, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v20, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v20, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v20, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v20, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v20, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v20, s54, 10 |
| ; SI-NEXT: v_writelane_b32 v20, s55, 11 |
| ; SI-NEXT: v_writelane_b32 v20, s64, 12 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s12, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s15, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s74, v1 |
| ; SI-NEXT: v_readfirstlane_b32 s77, v0 |
| ; SI-NEXT: v_writelane_b32 v20, s65, 13 |
| ; SI-NEXT: s_lshr_b32 s11, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s13, s28, 16 |
| ; SI-NEXT: s_lshr_b32 s72, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s26, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; SI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; SI-NEXT: s_lshr_b32 s92, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s93, s18, 16 |
| ; SI-NEXT: s_lshr_b32 s94, s17, 16 |
| ; SI-NEXT: s_lshr_b32 s95, s16, 16 |
| ; SI-NEXT: s_lshr_b32 s6, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s8, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s10, s12, 16 |
| ; SI-NEXT: s_lshr_b32 s14, s15, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s74, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s77, 16 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v20, s66, 14 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_writelane_b32 v20, s67, 15 |
| ; SI-NEXT: s_cbranch_scc0 .LBB15_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s95, 16 |
| ; SI-NEXT: s_or_b32 s36, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s94, 16 |
| ; SI-NEXT: s_or_b32 s37, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s93, 16 |
| ; SI-NEXT: s_or_b32 s38, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s92, 16 |
| ; SI-NEXT: s_or_b32 s39, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; SI-NEXT: s_or_b32 s40, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; SI-NEXT: s_or_b32 s41, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; SI-NEXT: s_or_b32 s42, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; SI-NEXT: s_or_b32 s43, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; SI-NEXT: s_or_b32 s44, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; SI-NEXT: s_or_b32 s45, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; SI-NEXT: s_or_b32 s46, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; SI-NEXT: s_or_b32 s47, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; SI-NEXT: s_or_b32 s48, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; SI-NEXT: s_or_b32 s49, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s77, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; SI-NEXT: s_or_b32 s50, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s74, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; SI-NEXT: s_or_b32 s51, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s15, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; SI-NEXT: s_or_b32 s52, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; SI-NEXT: s_or_b32 s53, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; SI-NEXT: s_or_b32 s54, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; SI-NEXT: s_or_b32 s55, s4, s5 |
| ; SI-NEXT: s_cbranch_execnz .LBB15_3 |
| ; SI-NEXT: .LBB15_2: ; %cmp.true |
| ; SI-NEXT: s_add_i32 s16, s16, 3 |
| ; SI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s95, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s17, s17, 3 |
| ; SI-NEXT: s_add_i32 s36, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s94, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s18, s18, 3 |
| ; SI-NEXT: s_add_i32 s37, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s93, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s19, s19, 3 |
| ; SI-NEXT: s_add_i32 s38, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s92, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s20, s20, 3 |
| ; SI-NEXT: s_add_i32 s39, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s21, s21, 3 |
| ; SI-NEXT: s_add_i32 s40, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s22, s22, 3 |
| ; SI-NEXT: s_add_i32 s41, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s23, s23, 3 |
| ; SI-NEXT: s_add_i32 s42, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s24, s24, 3 |
| ; SI-NEXT: s_add_i32 s43, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s25, s25, 3 |
| ; SI-NEXT: s_add_i32 s44, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s26, s26, 3 |
| ; SI-NEXT: s_add_i32 s45, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s27, s27, 3 |
| ; SI-NEXT: s_add_i32 s46, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s28, s28, 3 |
| ; SI-NEXT: s_add_i32 s47, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s29, s29, 3 |
| ; SI-NEXT: s_add_i32 s48, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s77, s77, 3 |
| ; SI-NEXT: s_add_i32 s49, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s77, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s74, s74, 3 |
| ; SI-NEXT: s_add_i32 s50, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s74, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s15, s15, 3 |
| ; SI-NEXT: s_add_i32 s51, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s15, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s12, s12, 3 |
| ; SI-NEXT: s_add_i32 s52, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s9, s9, 3 |
| ; SI-NEXT: s_add_i32 s53, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s7, s7, 3 |
| ; SI-NEXT: s_add_i32 s54, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s55, s4, 0x30000 |
| ; SI-NEXT: .LBB15_3: ; %end |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_readlane_b32 s67, v20, 15 |
| ; SI-NEXT: v_readlane_b32 s66, v20, 14 |
| ; SI-NEXT: v_readlane_b32 s65, v20, 13 |
| ; SI-NEXT: v_readlane_b32 s64, v20, 12 |
| ; SI-NEXT: v_readlane_b32 s55, v20, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v20, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v20, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v20, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v20, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v20, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v20, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v20, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v20, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v20, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v20, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v20, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB15_4: |
| ; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; SI-NEXT: s_branch .LBB15_2 |
| ; |
| ; VI-LABEL: bitcast_v40i16_to_v20i32_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v20, s30, 0 |
| ; VI-NEXT: v_writelane_b32 v20, s31, 1 |
| ; VI-NEXT: v_writelane_b32 v20, s34, 2 |
| ; VI-NEXT: v_writelane_b32 v20, s35, 3 |
| ; VI-NEXT: v_writelane_b32 v20, s36, 4 |
| ; VI-NEXT: v_writelane_b32 v20, s37, 5 |
| ; VI-NEXT: v_writelane_b32 v20, s38, 6 |
| ; VI-NEXT: v_writelane_b32 v20, s39, 7 |
| ; VI-NEXT: v_writelane_b32 v20, s48, 8 |
| ; VI-NEXT: v_writelane_b32 v20, s49, 9 |
| ; VI-NEXT: v_writelane_b32 v20, s50, 10 |
| ; VI-NEXT: v_writelane_b32 v20, s51, 11 |
| ; VI-NEXT: v_writelane_b32 v20, s52, 12 |
| ; VI-NEXT: v_writelane_b32 v20, s53, 13 |
| ; VI-NEXT: v_writelane_b32 v20, s54, 14 |
| ; VI-NEXT: v_writelane_b32 v20, s55, 15 |
| ; VI-NEXT: v_writelane_b32 v20, s64, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s12, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s15, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s74, v1 |
| ; VI-NEXT: v_readfirstlane_b32 s77, v0 |
| ; VI-NEXT: v_writelane_b32 v20, s65, 17 |
| ; VI-NEXT: s_lshr_b32 s11, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s72, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s75, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s78, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s30, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s31, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s34, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s35, s16, 16 |
| ; VI-NEXT: s_lshr_b32 s6, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s8, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s10, s12, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s15, 16 |
| ; VI-NEXT: s_lshr_b32 s73, s74, 16 |
| ; VI-NEXT: s_lshr_b32 s76, s77, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v20, s66, 18 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_writelane_b32 v20, s67, 19 |
| ; VI-NEXT: s_cbranch_scc0 .LBB15_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s35, 16 |
| ; VI-NEXT: s_or_b32 s36, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s5, s34, 16 |
| ; VI-NEXT: s_or_b32 s37, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s5, s31, 16 |
| ; VI-NEXT: s_or_b32 s38, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s5, s30, 16 |
| ; VI-NEXT: s_or_b32 s39, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; VI-NEXT: s_or_b32 s40, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; VI-NEXT: s_or_b32 s41, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; VI-NEXT: s_or_b32 s42, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; VI-NEXT: s_or_b32 s43, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; VI-NEXT: s_or_b32 s44, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; VI-NEXT: s_or_b32 s45, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; VI-NEXT: s_or_b32 s46, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; VI-NEXT: s_or_b32 s47, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; VI-NEXT: s_or_b32 s48, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; VI-NEXT: s_or_b32 s49, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s77 |
| ; VI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; VI-NEXT: s_or_b32 s50, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s74 |
| ; VI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; VI-NEXT: s_or_b32 s51, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s15 |
| ; VI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; VI-NEXT: s_or_b32 s52, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s12 |
| ; VI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; VI-NEXT: s_or_b32 s53, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s9 |
| ; VI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; VI-NEXT: s_or_b32 s54, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s7 |
| ; VI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; VI-NEXT: s_or_b32 s55, s4, s5 |
| ; VI-NEXT: s_cbranch_execnz .LBB15_3 |
| ; VI-NEXT: .LBB15_2: ; %cmp.true |
| ; VI-NEXT: s_add_i32 s16, s16, 3 |
| ; VI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s35, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s17, s17, 3 |
| ; VI-NEXT: s_add_i32 s36, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s34, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s18, s18, 3 |
| ; VI-NEXT: s_add_i32 s37, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s31, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s19, s19, 3 |
| ; VI-NEXT: s_add_i32 s38, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s30, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s20, s20, 3 |
| ; VI-NEXT: s_add_i32 s39, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s21, s21, 3 |
| ; VI-NEXT: s_add_i32 s40, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s22, s22, 3 |
| ; VI-NEXT: s_add_i32 s41, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s23, s23, 3 |
| ; VI-NEXT: s_add_i32 s42, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s24, s24, 3 |
| ; VI-NEXT: s_add_i32 s43, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s25, s25, 3 |
| ; VI-NEXT: s_add_i32 s44, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s26, s26, 3 |
| ; VI-NEXT: s_add_i32 s45, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s27, s27, 3 |
| ; VI-NEXT: s_add_i32 s46, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s28, s28, 3 |
| ; VI-NEXT: s_add_i32 s47, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s29, s29, 3 |
| ; VI-NEXT: s_add_i32 s48, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s77, s77, 3 |
| ; VI-NEXT: s_add_i32 s49, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s77, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s74, s74, 3 |
| ; VI-NEXT: s_add_i32 s50, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s74, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s15, s15, 3 |
| ; VI-NEXT: s_add_i32 s51, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s15, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s12, s12, 3 |
| ; VI-NEXT: s_add_i32 s52, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s9, s9, 3 |
| ; VI-NEXT: s_add_i32 s53, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s9, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s7, s7, 3 |
| ; VI-NEXT: s_add_i32 s54, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s7, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s55, s4, 0x30000 |
| ; VI-NEXT: .LBB15_3: ; %end |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_readlane_b32 s67, v20, 19 |
| ; VI-NEXT: v_readlane_b32 s66, v20, 18 |
| ; VI-NEXT: v_readlane_b32 s65, v20, 17 |
| ; VI-NEXT: v_readlane_b32 s64, v20, 16 |
| ; VI-NEXT: v_readlane_b32 s55, v20, 15 |
| ; VI-NEXT: v_readlane_b32 s54, v20, 14 |
| ; VI-NEXT: v_readlane_b32 s53, v20, 13 |
| ; VI-NEXT: v_readlane_b32 s52, v20, 12 |
| ; VI-NEXT: v_readlane_b32 s51, v20, 11 |
| ; VI-NEXT: v_readlane_b32 s50, v20, 10 |
| ; VI-NEXT: v_readlane_b32 s49, v20, 9 |
| ; VI-NEXT: v_readlane_b32 s48, v20, 8 |
| ; VI-NEXT: v_readlane_b32 s39, v20, 7 |
| ; VI-NEXT: v_readlane_b32 s38, v20, 6 |
| ; VI-NEXT: v_readlane_b32 s37, v20, 5 |
| ; VI-NEXT: v_readlane_b32 s36, v20, 4 |
| ; VI-NEXT: v_readlane_b32 s35, v20, 3 |
| ; VI-NEXT: v_readlane_b32 s34, v20, 2 |
| ; VI-NEXT: v_readlane_b32 s31, v20, 1 |
| ; VI-NEXT: v_readlane_b32 s30, v20, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB15_4: |
| ; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; VI-NEXT: s_branch .LBB15_2 |
| ; |
| ; GFX9-LABEL: bitcast_v40i16_to_v20i32_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_readfirstlane_b32 s56, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s58, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s60, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s62, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s72, v1 |
| ; GFX9-NEXT: v_readfirstlane_b32 s74, v0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: s_lshr_b32 s4, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s5, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s6, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s7, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s8, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s9, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s10, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s11, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s12, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s16, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s56, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s58, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s60, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s62, 16 |
| ; GFX9-NEXT: s_lshr_b32 s73, s72, 16 |
| ; GFX9-NEXT: s_lshr_b32 s75, s74, 16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s42, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_cmp_lg_u32 s42, 0 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s50, s74, s75 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s51, s72, s73 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s52, s62, s63 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s53, s60, s61 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s54, s58, s59 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s55, s56, s57 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB15_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB15_4 |
| ; GFX9-NEXT: .LBB15_2: ; %cmp.true |
| ; GFX9-NEXT: v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_branch .LBB15_5 |
| ; GFX9-NEXT: .LBB15_3: |
| ; GFX9-NEXT: s_branch .LBB15_2 |
| ; GFX9-NEXT: .LBB15_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB15_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v40i16_to_v20i32_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s45, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s46, v0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s56, v2 |
| ; GFX11-NEXT: s_lshr_b32 s41, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s6, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s5, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s4, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: s_lshr_b32 s58, s45, 16 |
| ; GFX11-NEXT: s_lshr_b32 s59, s46, 16 |
| ; GFX11-NEXT: s_mov_b32 s40, 0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s56, 0 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s57 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s47 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s44 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s18, s46, s59 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s19, s45, s58 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB15_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB15_4 |
| ; GFX11-NEXT: .LBB15_2: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v18, s18, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v19, s19, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB15_3: |
| ; GFX11-NEXT: s_branch .LBB15_2 |
| ; GFX11-NEXT: .LBB15_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31 |
| ; 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 <40 x i16> %a, splat (i16 3) |
| %a2 = bitcast <40 x i16> %a1 to <20 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x i16> %a to <20 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <20 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x i32> %phi |
| } |
| |
| define <40 x half> @bitcast_v20i32_to_v40f16(<20 x i32> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v20i32_to_v40f16: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr31 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr29 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr26 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr25 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr24 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr23 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; implicit-def: $vgpr22 |
| ; SI-NEXT: ; implicit-def: $vgpr30 |
| ; SI-NEXT: ; implicit-def: $vgpr21 |
| ; SI-NEXT: ; implicit-def: $vgpr28 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr27 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB16_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v31, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v34, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB16_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB16_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1 |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13 |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v31, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v34, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB16_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v34 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v31 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v29 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v25 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v35 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v23 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v32 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v22 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v30 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v28 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v27 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v34 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v31 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v29 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v26 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v25 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v24 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20i32_to_v40f16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; 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: 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 .LBB16_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB16_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB16_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19 |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18 |
| ; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17 |
| ; 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_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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB16_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v20i32_to_v40f16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; 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 .LBB16_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB16_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB16_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: v_add_u32_e32 v19, 3, v19 |
| ; GFX9-NEXT: v_add_u32_e32 v18, 3, v18 |
| ; GFX9-NEXT: v_add_u32_e32 v17, 3, v17 |
| ; 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_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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB16_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_mov_b32 s4, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v39, v0, s4 |
| ; GFX9-NEXT: v_perm_b32 v1, v38, v1, s4 |
| ; GFX9-NEXT: v_perm_b32 v2, v37, v2, s4 |
| ; GFX9-NEXT: v_perm_b32 v3, v36, v3, s4 |
| ; GFX9-NEXT: v_perm_b32 v4, v35, v4, s4 |
| ; GFX9-NEXT: v_perm_b32 v5, v34, v5, s4 |
| ; GFX9-NEXT: v_perm_b32 v6, v33, v6, s4 |
| ; GFX9-NEXT: v_perm_b32 v7, v32, v7, s4 |
| ; GFX9-NEXT: v_perm_b32 v8, v31, v8, s4 |
| ; GFX9-NEXT: v_perm_b32 v9, v30, v9, s4 |
| ; GFX9-NEXT: v_perm_b32 v10, v29, v10, s4 |
| ; GFX9-NEXT: v_perm_b32 v11, v28, v11, s4 |
| ; GFX9-NEXT: v_perm_b32 v12, v27, v12, s4 |
| ; GFX9-NEXT: v_perm_b32 v13, v26, v13, s4 |
| ; GFX9-NEXT: v_perm_b32 v14, v25, v14, s4 |
| ; GFX9-NEXT: v_perm_b32 v15, v24, v15, s4 |
| ; GFX9-NEXT: v_perm_b32 v16, v23, v16, s4 |
| ; GFX9-NEXT: v_perm_b32 v17, v22, v17, s4 |
| ; GFX9-NEXT: v_perm_b32 v18, v21, v18, s4 |
| ; GFX9-NEXT: v_perm_b32 v19, v20, v19, s4 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v20i32_to_v40f16: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB16_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v19, 3, v19 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v18, 3, v18 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v17, 3, v17 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v16, 3, v16 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v15, 3, v15 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v14, 3, v14 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v13, 3, v13 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v12, 3, v12 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v11, 3, v11 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v10, 3, v10 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, 3, v9 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 3, v8 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v7, 3, v7 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v6 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v5 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v4 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, 3, v3 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v2, 3, v2 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v1 |
| ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v0 |
| ; GFX11-TRUE16-NEXT: .LBB16_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v20i32_to_v40f16: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB16_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB16_2: ; %Flow |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB16_4 |
| ; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v19, 3, v19 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v18, 3, v18 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v17, 3, v17 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v16, 3, v16 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v15, 3, v15 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v14, 3, v14 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v13, 3, v13 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v12, 3, v12 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v11, 3, v11 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v10, 3, v10 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, 3, v9 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v8 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v7 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 3, v6 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 3, v5 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v4 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v3 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v2 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 3, v1 |
| ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB16_4: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v39, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v37, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v36, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v35, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v34, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v33, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v32, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v31, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v30, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v29, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v28, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v27, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v26, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v25, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v24, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v23, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v22, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v21, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v20, v19, 0x5040100 |
| ; GFX11-FAKE16-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 <20 x i32> %a, splat (i32 3) |
| %a2 = bitcast <20 x i32> %a1 to <40 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x i32> %a to <40 x half> |
| br label %end |
| |
| end: |
| %phi = phi <40 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x half> %phi |
| } |
| |
| define inreg <40 x half> @bitcast_v20i32_to_v40f16_scalar(<20 x i32> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v20i32_to_v40f16_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s8, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s5, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s8, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB17_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_lshr_b32 s72, s5, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s17, 16 |
| ; SI-NEXT: s_lshr_b64 s[10:11], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b64 s[12:13], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[14:15], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[40:41], s[28:29], 16 |
| ; SI-NEXT: s_lshr_b64 s[42:43], s[26:27], 16 |
| ; SI-NEXT: s_lshr_b64 s[44:45], s[24:25], 16 |
| ; SI-NEXT: s_lshr_b64 s[46:47], s[22:23], 16 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[20:21], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[18:19], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[16:17], 16 |
| ; SI-NEXT: s_cbranch_execnz .LBB17_3 |
| ; SI-NEXT: .LBB17_2: ; %cmp.true |
| ; SI-NEXT: s_add_i32 s17, s17, 3 |
| ; SI-NEXT: s_add_i32 s16, s16, 3 |
| ; SI-NEXT: s_add_i32 s19, s19, 3 |
| ; SI-NEXT: s_add_i32 s18, s18, 3 |
| ; SI-NEXT: s_add_i32 s21, s21, 3 |
| ; SI-NEXT: s_add_i32 s20, s20, 3 |
| ; SI-NEXT: s_add_i32 s23, s23, 3 |
| ; SI-NEXT: s_add_i32 s22, s22, 3 |
| ; SI-NEXT: s_add_i32 s25, s25, 3 |
| ; SI-NEXT: s_add_i32 s24, s24, 3 |
| ; SI-NEXT: s_add_i32 s27, s27, 3 |
| ; SI-NEXT: s_add_i32 s26, s26, 3 |
| ; SI-NEXT: s_add_i32 s29, s29, 3 |
| ; SI-NEXT: s_add_i32 s28, s28, 3 |
| ; SI-NEXT: s_add_i32 s9, s9, 3 |
| ; SI-NEXT: s_add_i32 s8, s8, 3 |
| ; SI-NEXT: s_add_i32 s7, s7, 3 |
| ; SI-NEXT: s_add_i32 s6, s6, 3 |
| ; SI-NEXT: s_add_i32 s5, s5, 3 |
| ; SI-NEXT: s_add_i32 s4, s4, 3 |
| ; SI-NEXT: s_lshr_b64 s[10:11], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b64 s[12:13], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[14:15], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[40:41], s[28:29], 16 |
| ; SI-NEXT: s_lshr_b64 s[42:43], s[26:27], 16 |
| ; SI-NEXT: s_lshr_b64 s[44:45], s[24:25], 16 |
| ; SI-NEXT: s_lshr_b64 s[46:47], s[22:23], 16 |
| ; SI-NEXT: s_lshr_b32 s72, s5, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s17, 16 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[20:21], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[18:19], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[16:17], 16 |
| ; SI-NEXT: .LBB17_3: ; %end |
| ; SI-NEXT: s_lshl_b32 s11, s60, 16 |
| ; SI-NEXT: s_and_b32 s13, s16, 0xffff |
| ; SI-NEXT: s_or_b32 s11, s13, s11 |
| ; SI-NEXT: s_and_b32 s13, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s15, s89, 16 |
| ; SI-NEXT: s_or_b32 s13, s13, s15 |
| ; SI-NEXT: s_lshl_b32 s15, s58, 16 |
| ; SI-NEXT: s_and_b32 s16, s18, 0xffff |
| ; SI-NEXT: s_or_b32 s15, s16, s15 |
| ; SI-NEXT: s_and_b32 s16, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s17, s88, 16 |
| ; SI-NEXT: s_or_b32 s16, s16, s17 |
| ; SI-NEXT: s_lshl_b32 s17, s56, 16 |
| ; SI-NEXT: s_and_b32 s18, s20, 0xffff |
| ; SI-NEXT: s_or_b32 s17, s18, s17 |
| ; SI-NEXT: s_and_b32 s18, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s19, s79, 16 |
| ; SI-NEXT: s_or_b32 s18, s18, s19 |
| ; SI-NEXT: s_and_b32 s19, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s20, s46, 16 |
| ; SI-NEXT: s_or_b32 s19, s19, s20 |
| ; SI-NEXT: s_and_b32 s20, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s21, s78, 16 |
| ; SI-NEXT: s_or_b32 s20, s20, s21 |
| ; SI-NEXT: s_and_b32 s21, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s22, s44, 16 |
| ; SI-NEXT: s_or_b32 s21, s21, s22 |
| ; SI-NEXT: s_and_b32 s22, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s23, s77, 16 |
| ; SI-NEXT: s_or_b32 s22, s22, s23 |
| ; SI-NEXT: s_and_b32 s23, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s42, 16 |
| ; SI-NEXT: s_or_b32 s23, s23, s24 |
| ; SI-NEXT: s_and_b32 s24, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s25, s76, 16 |
| ; SI-NEXT: s_or_b32 s24, s24, s25 |
| ; SI-NEXT: s_and_b32 s25, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s26, s40, 16 |
| ; SI-NEXT: s_and_b32 s8, s8, 0xffff |
| ; SI-NEXT: s_lshl_b32 s14, s14, 16 |
| ; SI-NEXT: s_and_b32 s6, s6, 0xffff |
| ; SI-NEXT: s_lshl_b32 s12, s12, 16 |
| ; SI-NEXT: s_and_b32 s4, s4, 0xffff |
| ; SI-NEXT: s_lshl_b32 s10, s10, 16 |
| ; SI-NEXT: s_or_b32 s25, s25, s26 |
| ; SI-NEXT: s_and_b32 s26, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s27, s75, 16 |
| ; SI-NEXT: s_or_b32 s8, s8, s14 |
| ; SI-NEXT: s_and_b32 s9, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s14, s74, 16 |
| ; SI-NEXT: s_or_b32 s6, s6, s12 |
| ; SI-NEXT: s_and_b32 s7, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s12, s73, 16 |
| ; SI-NEXT: s_or_b32 s4, s4, s10 |
| ; SI-NEXT: s_and_b32 s5, s5, 0xffff |
| ; SI-NEXT: s_lshl_b32 s10, s72, 16 |
| ; SI-NEXT: s_or_b32 s26, s26, s27 |
| ; SI-NEXT: s_or_b32 s9, s9, s14 |
| ; SI-NEXT: s_or_b32 s7, s7, s12 |
| ; SI-NEXT: s_or_b32 s5, s5, s10 |
| ; SI-NEXT: v_mov_b32_e32 v0, s11 |
| ; SI-NEXT: v_mov_b32_e32 v1, s13 |
| ; SI-NEXT: v_mov_b32_e32 v2, s15 |
| ; SI-NEXT: v_mov_b32_e32 v3, s16 |
| ; SI-NEXT: v_mov_b32_e32 v4, s17 |
| ; SI-NEXT: v_mov_b32_e32 v5, s18 |
| ; SI-NEXT: v_mov_b32_e32 v6, s19 |
| ; SI-NEXT: v_mov_b32_e32 v7, s20 |
| ; SI-NEXT: v_mov_b32_e32 v8, s21 |
| ; SI-NEXT: v_mov_b32_e32 v9, s22 |
| ; SI-NEXT: v_mov_b32_e32 v10, s23 |
| ; SI-NEXT: v_mov_b32_e32 v11, s24 |
| ; SI-NEXT: v_mov_b32_e32 v12, s25 |
| ; SI-NEXT: v_mov_b32_e32 v13, s26 |
| ; SI-NEXT: v_mov_b32_e32 v14, s8 |
| ; SI-NEXT: v_mov_b32_e32 v15, s9 |
| ; SI-NEXT: v_mov_b32_e32 v16, s6 |
| ; SI-NEXT: v_mov_b32_e32 v17, s7 |
| ; SI-NEXT: v_mov_b32_e32 v18, s4 |
| ; SI-NEXT: v_mov_b32_e32 v19, s5 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB17_4: |
| ; SI-NEXT: ; implicit-def: $sgpr60 |
| ; SI-NEXT: ; implicit-def: $sgpr89 |
| ; SI-NEXT: ; implicit-def: $sgpr58 |
| ; SI-NEXT: ; implicit-def: $sgpr88 |
| ; SI-NEXT: ; implicit-def: $sgpr56 |
| ; SI-NEXT: ; implicit-def: $sgpr79 |
| ; SI-NEXT: ; implicit-def: $sgpr46 |
| ; SI-NEXT: ; implicit-def: $sgpr78 |
| ; SI-NEXT: ; implicit-def: $sgpr44 |
| ; SI-NEXT: ; implicit-def: $sgpr77 |
| ; SI-NEXT: ; implicit-def: $sgpr42 |
| ; SI-NEXT: ; implicit-def: $sgpr76 |
| ; SI-NEXT: ; implicit-def: $sgpr40 |
| ; SI-NEXT: ; implicit-def: $sgpr75 |
| ; SI-NEXT: ; implicit-def: $sgpr14 |
| ; SI-NEXT: ; implicit-def: $sgpr74 |
| ; SI-NEXT: ; implicit-def: $sgpr12 |
| ; SI-NEXT: ; implicit-def: $sgpr73 |
| ; SI-NEXT: ; implicit-def: $sgpr10 |
| ; SI-NEXT: ; implicit-def: $sgpr72 |
| ; SI-NEXT: s_branch .LBB17_2 |
| ; |
| ; VI-LABEL: bitcast_v20i32_to_v40f16_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB17_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_lshr_b32 s12, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s10, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s46, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: s_cbranch_execnz .LBB17_3 |
| ; VI-NEXT: .LBB17_2: ; %cmp.true |
| ; VI-NEXT: s_add_i32 s6, s6, 3 |
| ; VI-NEXT: s_add_i32 s7, s7, 3 |
| ; VI-NEXT: s_add_i32 s8, s8, 3 |
| ; VI-NEXT: s_add_i32 s9, s9, 3 |
| ; VI-NEXT: s_add_i32 s10, s10, 3 |
| ; VI-NEXT: s_add_i32 s11, s11, 3 |
| ; VI-NEXT: s_add_i32 s29, s29, 3 |
| ; VI-NEXT: s_add_i32 s28, s28, 3 |
| ; VI-NEXT: s_add_i32 s27, s27, 3 |
| ; VI-NEXT: s_add_i32 s26, s26, 3 |
| ; VI-NEXT: s_add_i32 s25, s25, 3 |
| ; VI-NEXT: s_add_i32 s24, s24, 3 |
| ; VI-NEXT: s_add_i32 s23, s23, 3 |
| ; VI-NEXT: s_add_i32 s22, s22, 3 |
| ; VI-NEXT: s_add_i32 s21, s21, 3 |
| ; VI-NEXT: s_add_i32 s20, s20, 3 |
| ; VI-NEXT: s_add_i32 s19, s19, 3 |
| ; VI-NEXT: s_add_i32 s18, s18, 3 |
| ; VI-NEXT: s_add_i32 s17, s17, 3 |
| ; VI-NEXT: s_add_i32 s16, s16, 3 |
| ; VI-NEXT: s_lshr_b32 s12, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s10, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s46, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: .LBB17_3: ; %end |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s63, 16 |
| ; VI-NEXT: s_or_b32 s4, s4, s5 |
| ; VI-NEXT: s_and_b32 s5, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s16, s62, 16 |
| ; VI-NEXT: s_or_b32 s5, s5, s16 |
| ; VI-NEXT: s_and_b32 s16, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s17, s61, 16 |
| ; VI-NEXT: s_or_b32 s16, s16, s17 |
| ; VI-NEXT: s_and_b32 s17, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s18, s60, 16 |
| ; VI-NEXT: s_or_b32 s17, s17, s18 |
| ; VI-NEXT: s_and_b32 s18, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s19, s59, 16 |
| ; VI-NEXT: s_or_b32 s18, s18, s19 |
| ; VI-NEXT: s_and_b32 s19, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s20, s58, 16 |
| ; VI-NEXT: s_or_b32 s19, s19, s20 |
| ; VI-NEXT: s_and_b32 s20, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s21, s57, 16 |
| ; VI-NEXT: s_or_b32 s20, s20, s21 |
| ; VI-NEXT: s_and_b32 s21, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s22, s56, 16 |
| ; VI-NEXT: s_or_b32 s21, s21, s22 |
| ; VI-NEXT: s_and_b32 s22, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s23, s47, 16 |
| ; VI-NEXT: s_or_b32 s22, s22, s23 |
| ; VI-NEXT: s_and_b32 s23, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s24, s46, 16 |
| ; VI-NEXT: s_or_b32 s23, s23, s24 |
| ; VI-NEXT: s_and_b32 s24, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s25, s45, 16 |
| ; VI-NEXT: s_or_b32 s24, s24, s25 |
| ; VI-NEXT: s_and_b32 s25, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s26, s44, 16 |
| ; VI-NEXT: s_or_b32 s25, s25, s26 |
| ; VI-NEXT: s_and_b32 s26, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s27, s43, 16 |
| ; VI-NEXT: s_or_b32 s26, s26, s27 |
| ; VI-NEXT: s_and_b32 s27, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s28, s42, 16 |
| ; VI-NEXT: s_or_b32 s27, s27, s28 |
| ; VI-NEXT: s_and_b32 s11, 0xffff, s11 |
| ; VI-NEXT: s_lshl_b32 s28, s41, 16 |
| ; VI-NEXT: s_or_b32 s11, s11, s28 |
| ; VI-NEXT: s_and_b32 s10, 0xffff, s10 |
| ; VI-NEXT: s_lshl_b32 s28, s40, 16 |
| ; VI-NEXT: s_and_b32 s9, 0xffff, s9 |
| ; VI-NEXT: s_lshl_b32 s15, s15, 16 |
| ; VI-NEXT: s_and_b32 s8, 0xffff, s8 |
| ; VI-NEXT: s_lshl_b32 s14, s14, 16 |
| ; VI-NEXT: s_and_b32 s7, 0xffff, s7 |
| ; VI-NEXT: s_lshl_b32 s13, s13, 16 |
| ; VI-NEXT: s_and_b32 s6, 0xffff, s6 |
| ; VI-NEXT: s_lshl_b32 s12, s12, 16 |
| ; VI-NEXT: s_or_b32 s10, s10, s28 |
| ; VI-NEXT: s_or_b32 s9, s9, s15 |
| ; VI-NEXT: s_or_b32 s8, s8, s14 |
| ; VI-NEXT: s_or_b32 s7, s7, s13 |
| ; VI-NEXT: s_or_b32 s6, s6, s12 |
| ; VI-NEXT: v_mov_b32_e32 v0, s4 |
| ; VI-NEXT: v_mov_b32_e32 v1, s5 |
| ; VI-NEXT: v_mov_b32_e32 v2, s16 |
| ; VI-NEXT: v_mov_b32_e32 v3, s17 |
| ; VI-NEXT: v_mov_b32_e32 v4, s18 |
| ; VI-NEXT: v_mov_b32_e32 v5, s19 |
| ; VI-NEXT: v_mov_b32_e32 v6, s20 |
| ; VI-NEXT: v_mov_b32_e32 v7, s21 |
| ; VI-NEXT: v_mov_b32_e32 v8, s22 |
| ; VI-NEXT: v_mov_b32_e32 v9, s23 |
| ; VI-NEXT: v_mov_b32_e32 v10, s24 |
| ; VI-NEXT: v_mov_b32_e32 v11, s25 |
| ; VI-NEXT: v_mov_b32_e32 v12, s26 |
| ; VI-NEXT: v_mov_b32_e32 v13, s27 |
| ; VI-NEXT: v_mov_b32_e32 v14, s11 |
| ; VI-NEXT: v_mov_b32_e32 v15, s10 |
| ; VI-NEXT: v_mov_b32_e32 v16, s9 |
| ; VI-NEXT: v_mov_b32_e32 v17, s8 |
| ; VI-NEXT: v_mov_b32_e32 v18, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB17_4: |
| ; VI-NEXT: ; implicit-def: $sgpr63 |
| ; VI-NEXT: ; implicit-def: $sgpr62 |
| ; VI-NEXT: ; implicit-def: $sgpr61 |
| ; VI-NEXT: ; implicit-def: $sgpr60 |
| ; VI-NEXT: ; implicit-def: $sgpr59 |
| ; VI-NEXT: ; implicit-def: $sgpr58 |
| ; VI-NEXT: ; implicit-def: $sgpr57 |
| ; VI-NEXT: ; implicit-def: $sgpr56 |
| ; VI-NEXT: ; implicit-def: $sgpr47 |
| ; VI-NEXT: ; implicit-def: $sgpr46 |
| ; VI-NEXT: ; implicit-def: $sgpr45 |
| ; VI-NEXT: ; implicit-def: $sgpr44 |
| ; VI-NEXT: ; implicit-def: $sgpr43 |
| ; VI-NEXT: ; implicit-def: $sgpr42 |
| ; VI-NEXT: ; implicit-def: $sgpr41 |
| ; VI-NEXT: ; implicit-def: $sgpr40 |
| ; VI-NEXT: ; implicit-def: $sgpr15 |
| ; VI-NEXT: ; implicit-def: $sgpr14 |
| ; VI-NEXT: ; implicit-def: $sgpr13 |
| ; VI-NEXT: ; implicit-def: $sgpr12 |
| ; VI-NEXT: s_branch .LBB17_2 |
| ; |
| ; GFX9-LABEL: bitcast_v20i32_to_v40f16_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s10, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s11, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB17_4 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_lshr_b32 s12, s6, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s7, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s8, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s9, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s10, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s11, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s43, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s58, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s60, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s62, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s16, 16 |
| ; GFX9-NEXT: s_cbranch_execnz .LBB17_3 |
| ; GFX9-NEXT: .LBB17_2: ; %cmp.true |
| ; GFX9-NEXT: s_add_i32 s6, s6, 3 |
| ; GFX9-NEXT: s_add_i32 s7, s7, 3 |
| ; GFX9-NEXT: s_add_i32 s8, s8, 3 |
| ; GFX9-NEXT: s_add_i32 s9, s9, 3 |
| ; GFX9-NEXT: s_add_i32 s10, s10, 3 |
| ; GFX9-NEXT: s_add_i32 s11, s11, 3 |
| ; GFX9-NEXT: s_add_i32 s29, s29, 3 |
| ; GFX9-NEXT: s_add_i32 s28, s28, 3 |
| ; GFX9-NEXT: s_add_i32 s27, s27, 3 |
| ; GFX9-NEXT: s_add_i32 s26, s26, 3 |
| ; GFX9-NEXT: s_add_i32 s25, s25, 3 |
| ; GFX9-NEXT: s_add_i32 s24, s24, 3 |
| ; GFX9-NEXT: s_add_i32 s23, s23, 3 |
| ; GFX9-NEXT: s_add_i32 s22, s22, 3 |
| ; GFX9-NEXT: s_add_i32 s21, s21, 3 |
| ; GFX9-NEXT: s_add_i32 s20, s20, 3 |
| ; GFX9-NEXT: s_add_i32 s19, s19, 3 |
| ; GFX9-NEXT: s_add_i32 s18, s18, 3 |
| ; GFX9-NEXT: s_add_i32 s17, s17, 3 |
| ; GFX9-NEXT: s_add_i32 s16, s16, 3 |
| ; GFX9-NEXT: s_lshr_b32 s12, s6, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s7, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s8, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s9, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s10, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s11, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s43, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s58, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s60, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s62, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s16, 16 |
| ; GFX9-NEXT: .LBB17_3: ; %end |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s63 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s5, s17, s62 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s61 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s60 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s59 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s58 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s57 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s56 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s47 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s46 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s45 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s44 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s43 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s42 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s11, s11, s41 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s10, s10, s40 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s9, s9, s15 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s14 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s13 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s12 |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s5 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s6 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; GFX9-NEXT: .LBB17_4: |
| ; GFX9-NEXT: ; implicit-def: $sgpr63 |
| ; GFX9-NEXT: ; implicit-def: $sgpr62 |
| ; GFX9-NEXT: ; implicit-def: $sgpr61 |
| ; GFX9-NEXT: ; implicit-def: $sgpr60 |
| ; GFX9-NEXT: ; implicit-def: $sgpr59 |
| ; GFX9-NEXT: ; implicit-def: $sgpr58 |
| ; GFX9-NEXT: ; implicit-def: $sgpr57 |
| ; GFX9-NEXT: ; implicit-def: $sgpr56 |
| ; GFX9-NEXT: ; implicit-def: $sgpr47 |
| ; GFX9-NEXT: ; implicit-def: $sgpr46 |
| ; GFX9-NEXT: ; implicit-def: $sgpr45 |
| ; GFX9-NEXT: ; implicit-def: $sgpr44 |
| ; GFX9-NEXT: ; implicit-def: $sgpr43 |
| ; GFX9-NEXT: ; implicit-def: $sgpr42 |
| ; GFX9-NEXT: ; implicit-def: $sgpr41 |
| ; GFX9-NEXT: ; implicit-def: $sgpr40 |
| ; GFX9-NEXT: ; implicit-def: $sgpr15 |
| ; GFX9-NEXT: ; implicit-def: $sgpr14 |
| ; GFX9-NEXT: ; implicit-def: $sgpr13 |
| ; GFX9-NEXT: ; implicit-def: $sgpr12 |
| ; GFX9-NEXT: s_branch .LBB17_2 |
| ; |
| ; GFX11-LABEL: bitcast_v20i32_to_v40f16_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-NEXT: s_mov_b32 s58, 0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB17_4 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-NEXT: s_lshr_b32 s6, s4, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s5, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s40, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s41, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s45, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s46, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s56, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s58 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB17_3 |
| ; GFX11-NEXT: .LBB17_2: ; %cmp.true |
| ; GFX11-NEXT: s_add_i32 s4, s4, 3 |
| ; GFX11-NEXT: s_add_i32 s5, s5, 3 |
| ; GFX11-NEXT: s_add_i32 s29, s29, 3 |
| ; GFX11-NEXT: s_add_i32 s28, s28, 3 |
| ; GFX11-NEXT: s_add_i32 s27, s27, 3 |
| ; GFX11-NEXT: s_add_i32 s26, s26, 3 |
| ; GFX11-NEXT: s_add_i32 s25, s25, 3 |
| ; GFX11-NEXT: s_add_i32 s24, s24, 3 |
| ; GFX11-NEXT: s_add_i32 s23, s23, 3 |
| ; GFX11-NEXT: s_add_i32 s22, s22, 3 |
| ; GFX11-NEXT: s_add_i32 s21, s21, 3 |
| ; GFX11-NEXT: s_add_i32 s20, s20, 3 |
| ; GFX11-NEXT: s_add_i32 s19, s19, 3 |
| ; GFX11-NEXT: s_add_i32 s18, s18, 3 |
| ; GFX11-NEXT: s_add_i32 s17, s17, 3 |
| ; GFX11-NEXT: s_add_i32 s16, s16, 3 |
| ; GFX11-NEXT: s_add_i32 s3, s3, 3 |
| ; GFX11-NEXT: s_add_i32 s2, s2, 3 |
| ; GFX11-NEXT: s_add_i32 s1, s1, 3 |
| ; GFX11-NEXT: s_add_i32 s0, s0, 3 |
| ; GFX11-NEXT: s_lshr_b32 s6, s4, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s5, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s40, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s41, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s45, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s46, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s56, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: .LBB17_3: ; %end |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s57 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s56 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s47 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s46 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s45 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s44 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s43 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s42 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s41 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s40 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s15, s22, s15 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s14, s23, s14 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s13, s24, s13 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s12, s25, s12 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s11, s26, s11 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s10, s27, s10 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s9, s28, s9 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s8, s29, s8 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s7 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s6 |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB17_4: |
| ; GFX11-NEXT: ; implicit-def: $sgpr57 |
| ; GFX11-NEXT: ; implicit-def: $sgpr56 |
| ; GFX11-NEXT: ; implicit-def: $sgpr47 |
| ; GFX11-NEXT: ; implicit-def: $sgpr46 |
| ; GFX11-NEXT: ; implicit-def: $sgpr45 |
| ; GFX11-NEXT: ; implicit-def: $sgpr44 |
| ; GFX11-NEXT: ; implicit-def: $sgpr43 |
| ; GFX11-NEXT: ; implicit-def: $sgpr42 |
| ; GFX11-NEXT: ; implicit-def: $sgpr41 |
| ; GFX11-NEXT: ; implicit-def: $sgpr40 |
| ; GFX11-NEXT: ; implicit-def: $sgpr15 |
| ; GFX11-NEXT: ; implicit-def: $sgpr14 |
| ; GFX11-NEXT: ; implicit-def: $sgpr13 |
| ; GFX11-NEXT: ; implicit-def: $sgpr12 |
| ; GFX11-NEXT: ; implicit-def: $sgpr11 |
| ; GFX11-NEXT: ; implicit-def: $sgpr10 |
| ; GFX11-NEXT: ; implicit-def: $sgpr9 |
| ; GFX11-NEXT: ; implicit-def: $sgpr8 |
| ; GFX11-NEXT: ; implicit-def: $sgpr7 |
| ; GFX11-NEXT: ; implicit-def: $sgpr6 |
| ; GFX11-NEXT: s_branch .LBB17_2 |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <20 x i32> %a, splat (i32 3) |
| %a2 = bitcast <20 x i32> %a1 to <40 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x i32> %a to <40 x half> |
| br label %end |
| |
| end: |
| %phi = phi <40 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x half> %phi |
| } |
| |
| define <20 x i32> @bitcast_v40f16_to_v20i32(<40 x half> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v40f16_to_v20i32: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_mov_b32_e32 v32, v19 |
| ; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: v_mov_b32_e32 v33, v18 |
| ; SI-NEXT: v_mov_b32_e32 v43, v0 |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; SI-NEXT: v_mov_b32_e32 v34, v17 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; SI-NEXT: v_mov_b32_e32 v35, v16 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; SI-NEXT: v_mov_b32_e32 v36, v15 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v35 |
| ; SI-NEXT: v_mov_b32_e32 v37, v14 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v36 |
| ; SI-NEXT: v_mov_b32_e32 v38, v13 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v37 |
| ; SI-NEXT: v_mov_b32_e32 v39, v12 |
| ; SI-NEXT: v_mov_b32_e32 v48, v11 |
| ; SI-NEXT: v_mov_b32_e32 v49, v10 |
| ; SI-NEXT: v_mov_b32_e32 v50, v9 |
| ; SI-NEXT: v_mov_b32_e32 v51, v8 |
| ; SI-NEXT: v_mov_b32_e32 v52, v7 |
| ; SI-NEXT: v_mov_b32_e32 v53, v6 |
| ; SI-NEXT: v_mov_b32_e32 v54, v5 |
| ; SI-NEXT: v_mov_b32_e32 v55, v4 |
| ; SI-NEXT: v_mov_b32_e32 v40, v3 |
| ; SI-NEXT: v_mov_b32_e32 v41, v2 |
| ; SI-NEXT: v_mov_b32_e32 v42, v1 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v38 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v39 |
| ; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v48 |
| ; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v49 |
| ; SI-NEXT: v_lshrrev_b32_e32 v62, 16, v50 |
| ; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v51 |
| ; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v52 |
| ; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v53 |
| ; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v54 |
| ; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v55 |
| ; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v40 |
| ; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v41 |
| ; SI-NEXT: v_lshrrev_b32_e32 v58, 16, v42 |
| ; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v43 |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB18_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v43 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v59 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v42 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v58 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; SI-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v41 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v57 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v40 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v56 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v55 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v47 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v54 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v46 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v53 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v45 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v52 |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v44 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v51 |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v63 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v50 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v62 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v49 |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v61 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v60 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v39 |
| ; SI-NEXT: ; implicit-def: $vgpr43 |
| ; SI-NEXT: ; implicit-def: $vgpr42 |
| ; SI-NEXT: ; implicit-def: $vgpr41 |
| ; SI-NEXT: ; implicit-def: $vgpr40 |
| ; SI-NEXT: ; implicit-def: $vgpr55 |
| ; SI-NEXT: ; implicit-def: $vgpr54 |
| ; SI-NEXT: ; implicit-def: $vgpr53 |
| ; SI-NEXT: ; implicit-def: $vgpr52 |
| ; SI-NEXT: ; implicit-def: $vgpr51 |
| ; SI-NEXT: ; implicit-def: $vgpr50 |
| ; SI-NEXT: ; implicit-def: $vgpr49 |
| ; SI-NEXT: ; implicit-def: $vgpr48 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr59 |
| ; SI-NEXT: ; implicit-def: $vgpr58 |
| ; SI-NEXT: ; implicit-def: $vgpr57 |
| ; SI-NEXT: ; implicit-def: $vgpr56 |
| ; SI-NEXT: ; implicit-def: $vgpr47 |
| ; SI-NEXT: ; implicit-def: $vgpr46 |
| ; SI-NEXT: ; implicit-def: $vgpr45 |
| ; SI-NEXT: ; implicit-def: $vgpr44 |
| ; SI-NEXT: ; implicit-def: $vgpr63 |
| ; SI-NEXT: ; implicit-def: $vgpr62 |
| ; SI-NEXT: ; implicit-def: $vgpr61 |
| ; SI-NEXT: ; implicit-def: $vgpr60 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v38 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v37 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v36 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v35 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v34 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v33 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v19 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v32 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: .LBB18_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB18_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v0, v59 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, v58 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, v43 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v42 |
| ; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; SI-NEXT: v_or_b32_e32 v1, v3, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, v57 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v41 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, v40 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, v55 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v56 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, v54 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, v53 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, v47 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, v44 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, v52 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; SI-NEXT: v_or_b32_e32 v4, v5, v4 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, v46 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, v51 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, v61 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, v45 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, v49 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; SI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, v63 |
| ; SI-NEXT: v_or_b32_e32 v7, v9, v7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, v62 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; SI-NEXT: v_or_b32_e32 v8, v10, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, v50 |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, v38 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, v37 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v11 |
| ; SI-NEXT: v_or_b32_e32 v10, v12, v10 |
| ; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, v60 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, v35 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, v34 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; SI-NEXT: v_or_b32_e32 v11, v13, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, v39 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v21, v32 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, v12 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; SI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v14 |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: v_or_b32_e32 v13, v15, v13 |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v20, v20 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; SI-NEXT: v_or_b32_e32 v14, v16, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, v36 |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v17 |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: v_or_b32_e32 v16, v18, v16 |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_or_b32_e32 v17, v19, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, v33 |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_or_b32_e32 v18, v19, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v19, v21, v19 |
| ; SI-NEXT: .LBB18_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40f16_to_v20i32: |
| ; 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:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: v_mov_b32_e32 v32, v19 |
| ; VI-NEXT: v_mov_b32_e32 v33, v18 |
| ; VI-NEXT: v_mov_b32_e32 v34, v17 |
| ; VI-NEXT: v_mov_b32_e32 v35, v16 |
| ; VI-NEXT: v_mov_b32_e32 v36, v15 |
| ; VI-NEXT: v_mov_b32_e32 v37, v14 |
| ; VI-NEXT: v_mov_b32_e32 v38, v13 |
| ; VI-NEXT: v_mov_b32_e32 v39, v12 |
| ; VI-NEXT: v_mov_b32_e32 v48, v11 |
| ; VI-NEXT: v_mov_b32_e32 v49, v10 |
| ; VI-NEXT: v_mov_b32_e32 v50, v9 |
| ; VI-NEXT: v_mov_b32_e32 v51, v8 |
| ; VI-NEXT: v_mov_b32_e32 v52, v7 |
| ; VI-NEXT: v_mov_b32_e32 v53, v6 |
| ; VI-NEXT: v_mov_b32_e32 v54, v5 |
| ; VI-NEXT: v_mov_b32_e32 v55, v4 |
| ; VI-NEXT: v_mov_b32_e32 v40, v3 |
| ; VI-NEXT: v_mov_b32_e32 v41, v2 |
| ; VI-NEXT: v_mov_b32_e32 v42, v1 |
| ; VI-NEXT: v_mov_b32_e32 v43, v0 |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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 .LBB18_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_mov_b32_e32 v19, 16 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v0, v19, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v1, v19, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v2, v19, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v3, v19, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v4, v19, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v5, v19, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v6, v19, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v7, v19, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v8, v19, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v9, v19, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v10, v19, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v11, v19, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v12, v19, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v13, v19, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v14, v19, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v15, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v16, v19, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v17, v19, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v18, v19, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v19, v19, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v40, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v55, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v54, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v53, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v52, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v51, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v50, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v49, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v48, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v39, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v38, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v37, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v36, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v35, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v34, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v33, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v32, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; 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: .LBB18_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB18_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v19, 0x200 |
| ; VI-NEXT: v_add_f16_sdwa v0, v43, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v1, 0x200, v43 |
| ; VI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; VI-NEXT: v_add_f16_sdwa v1, v42, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v2, 0x200, v42 |
| ; VI-NEXT: v_or_b32_e32 v1, v2, v1 |
| ; VI-NEXT: v_add_f16_sdwa v2, v41, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, 0x200, v41 |
| ; VI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; VI-NEXT: v_add_f16_sdwa v3, v40, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v4, 0x200, v40 |
| ; VI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; VI-NEXT: v_add_f16_sdwa v4, v55, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v5, 0x200, v55 |
| ; VI-NEXT: v_or_b32_e32 v4, v5, v4 |
| ; VI-NEXT: v_add_f16_sdwa v5, v54, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v6, 0x200, v54 |
| ; VI-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; VI-NEXT: v_add_f16_sdwa v6, v53, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v7, 0x200, v53 |
| ; VI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; VI-NEXT: v_add_f16_sdwa v7, v52, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v8, 0x200, v52 |
| ; VI-NEXT: v_or_b32_e32 v7, v8, v7 |
| ; VI-NEXT: v_add_f16_sdwa v8, v51, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v9, 0x200, v51 |
| ; VI-NEXT: v_or_b32_e32 v8, v9, v8 |
| ; VI-NEXT: v_add_f16_sdwa v9, v50, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v10, 0x200, v50 |
| ; VI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; VI-NEXT: v_add_f16_sdwa v10, v49, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v11, 0x200, v49 |
| ; VI-NEXT: v_or_b32_e32 v10, v11, v10 |
| ; VI-NEXT: v_add_f16_sdwa v11, v48, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v12, 0x200, v48 |
| ; VI-NEXT: v_or_b32_e32 v11, v12, v11 |
| ; VI-NEXT: v_add_f16_sdwa v12, v39, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v13, 0x200, v39 |
| ; VI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; VI-NEXT: v_add_f16_sdwa v13, v38, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v14, 0x200, v38 |
| ; VI-NEXT: v_or_b32_e32 v13, v14, v13 |
| ; VI-NEXT: v_add_f16_sdwa v14, v37, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v15, 0x200, v37 |
| ; VI-NEXT: v_or_b32_e32 v14, v15, v14 |
| ; VI-NEXT: v_add_f16_sdwa v15, v36, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v16, 0x200, v36 |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: v_add_f16_sdwa v16, v35, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v17, 0x200, v35 |
| ; VI-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; VI-NEXT: v_add_f16_sdwa v17, v34, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v18, 0x200, v34 |
| ; VI-NEXT: v_or_b32_e32 v17, v18, v17 |
| ; VI-NEXT: v_add_f16_sdwa v18, v33, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v20, 0x200, v33 |
| ; VI-NEXT: v_or_b32_e32 v18, v20, v18 |
| ; VI-NEXT: v_add_f16_sdwa v19, v32, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v20, 0x200, v32 |
| ; VI-NEXT: v_or_b32_e32 v19, v20, v19 |
| ; VI-NEXT: .LBB18_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40f16_to_v20i32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_mov_b32_e32 v32, v19 |
| ; 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: v_mov_b32_e32 v33, v18 |
| ; GFX9-NEXT: v_mov_b32_e32 v43, v0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, v17 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, v16 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, v15 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, v14 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, v13 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, v12 |
| ; GFX9-NEXT: v_mov_b32_e32 v48, v11 |
| ; GFX9-NEXT: v_mov_b32_e32 v49, v10 |
| ; GFX9-NEXT: v_mov_b32_e32 v50, v9 |
| ; GFX9-NEXT: v_mov_b32_e32 v51, v8 |
| ; GFX9-NEXT: v_mov_b32_e32 v52, v7 |
| ; GFX9-NEXT: v_mov_b32_e32 v53, v6 |
| ; GFX9-NEXT: v_mov_b32_e32 v54, v5 |
| ; GFX9-NEXT: v_mov_b32_e32 v55, v4 |
| ; GFX9-NEXT: v_mov_b32_e32 v40, v3 |
| ; GFX9-NEXT: v_mov_b32_e32 v41, v2 |
| ; GFX9-NEXT: v_mov_b32_e32 v42, v1 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v48 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v49 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v50 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v51 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v52 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v53 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v54 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v55 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43 |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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 .LBB18_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v35 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v34 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v33 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v32 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; 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: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: .LBB18_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB18_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: s_movk_i32 s7, 0x200 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_pk_add_f16 v0, v0, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, v1, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, v2, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, v3, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, v4, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, v5, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; GFX9-NEXT: v_pk_add_f16 v12, v12, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, v13, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, v14, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v15, v15, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v16, v16, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v17, v17, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v18, v18, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v19, v19, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB18_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; 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-TRUE16-LABEL: bitcast_v40f16_to_v20i32: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB18_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: .LBB18_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v40f16_to_v20i32: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v37, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v39, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v48, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v36, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v35, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v34, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v33, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v32, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v31, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v30, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v29, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v28, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v27, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v26, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v25, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v24, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v23, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v22, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v21, v19, 0x5040100 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB18_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: .LBB18_2: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-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 <40 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <40 x half> %a1 to <20 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x half> %a to <20 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <20 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x i32> %phi |
| } |
| |
| define inreg <20 x i32> @bitcast_v40f16_to_v20i32_scalar(<40 x half> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v40f16_to_v20i32_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; SI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; SI-NEXT: v_writelane_b32 v32, s64, 12 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s10, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s12, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s14, v1 |
| ; SI-NEXT: v_readfirstlane_b32 s73, v0 |
| ; SI-NEXT: v_writelane_b32 v32, s65, 13 |
| ; SI-NEXT: s_lshr_b32 s72, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s28, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s26, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; SI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; SI-NEXT: s_lshr_b32 s92, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s93, s18, 16 |
| ; SI-NEXT: s_lshr_b32 s94, s17, 16 |
| ; SI-NEXT: s_lshr_b32 s95, s16, 16 |
| ; SI-NEXT: s_lshr_b32 s7, s6, 16 |
| ; SI-NEXT: s_lshr_b32 s9, s8, 16 |
| ; SI-NEXT: s_lshr_b32 s11, s10, 16 |
| ; SI-NEXT: s_lshr_b32 s13, s12, 16 |
| ; SI-NEXT: s_lshr_b32 s15, s14, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s73, 16 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v32, s66, 14 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s67, 15 |
| ; SI-NEXT: s_cbranch_scc0 .LBB19_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s95, 16 |
| ; SI-NEXT: s_or_b32 s36, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s94, 16 |
| ; SI-NEXT: s_or_b32 s37, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s93, 16 |
| ; SI-NEXT: s_or_b32 s38, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s92, 16 |
| ; SI-NEXT: s_or_b32 s39, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; SI-NEXT: s_or_b32 s40, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; SI-NEXT: s_or_b32 s41, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; SI-NEXT: s_or_b32 s42, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; SI-NEXT: s_or_b32 s43, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; SI-NEXT: s_or_b32 s44, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; SI-NEXT: s_or_b32 s45, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s77, 16 |
| ; SI-NEXT: s_or_b32 s46, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; SI-NEXT: s_or_b32 s47, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; SI-NEXT: s_or_b32 s48, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; SI-NEXT: s_or_b32 s49, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s73, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s74, 16 |
| ; SI-NEXT: s_or_b32 s50, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s14, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s15, 16 |
| ; SI-NEXT: s_or_b32 s51, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; SI-NEXT: s_or_b32 s52, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s10, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; SI-NEXT: s_or_b32 s53, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s8, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s9, 16 |
| ; SI-NEXT: s_or_b32 s54, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s6, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s7, 16 |
| ; SI-NEXT: s_or_b32 s55, s4, s5 |
| ; SI-NEXT: s_cbranch_execnz .LBB19_4 |
| ; SI-NEXT: .LBB19_2: ; %cmp.true |
| ; SI-NEXT: v_cvt_f32_f16_e32 v0, s95 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, s16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, s94 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s17 |
| ; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, s18 |
| ; SI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, s93 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_or_b32_e32 v1, v3, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s92 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_or_b32_e32 v2, v4, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, s19 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s91 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, s20 |
| ; SI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s90 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, s21 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_or_b32_e32 v4, v6, v4 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, s89 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_or_b32_e32 v5, v7, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, s22 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, s88 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, s23 |
| ; SI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, s79 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, s24 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_or_b32_e32 v7, v9, v7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, s78 |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; SI-NEXT: v_or_b32_e32 v8, v10, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, s25 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, s77 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, s26 |
| ; SI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, s76 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, s27 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_or_b32_e32 v10, v12, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, s75 |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; SI-NEXT: v_or_b32_e32 v11, v13, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, s28 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, s72 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, s29 |
| ; SI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, s74 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, s73 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_or_b32_e32 v13, v15, v13 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, s15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; SI-NEXT: v_or_b32_e32 v14, v16, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, s14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, s13 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, s12 |
| ; SI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, s11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, s10 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_or_b32_e32 v16, v18, v16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, s9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_or_b32_e32 v17, v19, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, s8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v20, s7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v21, s6 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_or_b32_e32 v18, v19, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v19, v21, v19 |
| ; SI-NEXT: s_branch .LBB19_5 |
| ; SI-NEXT: .LBB19_3: |
| ; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; SI-NEXT: s_branch .LBB19_2 |
| ; SI-NEXT: .LBB19_4: |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_mov_b32_e32 v20, s56 |
| ; SI-NEXT: v_mov_b32_e32 v21, s57 |
| ; SI-NEXT: v_mov_b32_e32 v22, s58 |
| ; SI-NEXT: v_mov_b32_e32 v23, s59 |
| ; SI-NEXT: v_mov_b32_e32 v24, s60 |
| ; SI-NEXT: v_mov_b32_e32 v25, s61 |
| ; SI-NEXT: v_mov_b32_e32 v26, s62 |
| ; SI-NEXT: v_mov_b32_e32 v27, s63 |
| ; SI-NEXT: v_mov_b32_e32 v28, s64 |
| ; SI-NEXT: v_mov_b32_e32 v29, s65 |
| ; SI-NEXT: v_mov_b32_e32 v30, s66 |
| ; SI-NEXT: v_mov_b32_e32 v31, s67 |
| ; SI-NEXT: .LBB19_5: ; %end |
| ; SI-NEXT: v_readlane_b32 s67, v32, 15 |
| ; SI-NEXT: v_readlane_b32 s66, v32, 14 |
| ; SI-NEXT: v_readlane_b32 s65, v32, 13 |
| ; SI-NEXT: v_readlane_b32 s64, v32, 12 |
| ; SI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40f16_to_v20i32_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v32, s30, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s31, 1 |
| ; VI-NEXT: v_writelane_b32 v32, s34, 2 |
| ; VI-NEXT: v_writelane_b32 v32, s35, 3 |
| ; VI-NEXT: v_writelane_b32 v32, s36, 4 |
| ; VI-NEXT: v_writelane_b32 v32, s37, 5 |
| ; VI-NEXT: v_writelane_b32 v32, s38, 6 |
| ; VI-NEXT: v_writelane_b32 v32, s39, 7 |
| ; VI-NEXT: v_writelane_b32 v32, s48, 8 |
| ; VI-NEXT: v_writelane_b32 v32, s49, 9 |
| ; VI-NEXT: v_writelane_b32 v32, s50, 10 |
| ; VI-NEXT: v_writelane_b32 v32, s51, 11 |
| ; VI-NEXT: v_writelane_b32 v32, s52, 12 |
| ; VI-NEXT: v_writelane_b32 v32, s53, 13 |
| ; VI-NEXT: v_writelane_b32 v32, s54, 14 |
| ; VI-NEXT: v_writelane_b32 v32, s55, 15 |
| ; VI-NEXT: v_writelane_b32 v32, s64, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s14, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s73, v1 |
| ; VI-NEXT: v_readfirstlane_b32 s76, v0 |
| ; VI-NEXT: v_writelane_b32 v32, s65, 17 |
| ; VI-NEXT: s_lshr_b32 s10, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s72, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s74, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s77, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s30, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s31, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s34, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s35, s16, 16 |
| ; VI-NEXT: s_lshr_b32 s7, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s9, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s12, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s14, 16 |
| ; VI-NEXT: s_lshr_b32 s75, s73, 16 |
| ; VI-NEXT: s_lshr_b32 s78, s76, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v32, s66, 18 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s67, 19 |
| ; VI-NEXT: s_cbranch_scc0 .LBB19_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s35, 16 |
| ; VI-NEXT: s_or_b32 s36, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s5, s34, 16 |
| ; VI-NEXT: s_or_b32 s37, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s5, s31, 16 |
| ; VI-NEXT: s_or_b32 s38, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s5, s30, 16 |
| ; VI-NEXT: s_or_b32 s39, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; VI-NEXT: s_or_b32 s40, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; VI-NEXT: s_or_b32 s41, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; VI-NEXT: s_or_b32 s42, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; VI-NEXT: s_or_b32 s43, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; VI-NEXT: s_or_b32 s44, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s5, s77, 16 |
| ; VI-NEXT: s_or_b32 s45, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s5, s74, 16 |
| ; VI-NEXT: s_or_b32 s46, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; VI-NEXT: s_or_b32 s47, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; VI-NEXT: s_or_b32 s48, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; VI-NEXT: s_or_b32 s49, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s76 |
| ; VI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; VI-NEXT: s_or_b32 s50, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s73 |
| ; VI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; VI-NEXT: s_or_b32 s51, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s14 |
| ; VI-NEXT: s_lshl_b32 s5, s15, 16 |
| ; VI-NEXT: s_or_b32 s52, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s11 |
| ; VI-NEXT: s_lshl_b32 s5, s12, 16 |
| ; VI-NEXT: s_or_b32 s53, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s8 |
| ; VI-NEXT: s_lshl_b32 s5, s9, 16 |
| ; VI-NEXT: s_or_b32 s54, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s6 |
| ; VI-NEXT: s_lshl_b32 s5, s7, 16 |
| ; VI-NEXT: s_or_b32 s55, s4, s5 |
| ; VI-NEXT: s_cbranch_execnz .LBB19_4 |
| ; VI-NEXT: .LBB19_2: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v19, 0x200 |
| ; VI-NEXT: v_mov_b32_e32 v0, s35 |
| ; VI-NEXT: v_mov_b32_e32 v2, s34 |
| ; VI-NEXT: v_add_f16_sdwa v0, v0, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v1, s16, v19 |
| ; VI-NEXT: v_add_f16_sdwa v2, v2, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, s17, v19 |
| ; VI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; VI-NEXT: v_or_b32_e32 v1, v3, v2 |
| ; VI-NEXT: v_mov_b32_e32 v2, s31 |
| ; VI-NEXT: v_add_f16_sdwa v2, v2, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, s18, v19 |
| ; VI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; VI-NEXT: v_mov_b32_e32 v3, s30 |
| ; VI-NEXT: v_add_f16_sdwa v3, v3, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v4, s19, v19 |
| ; VI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; VI-NEXT: v_mov_b32_e32 v4, s91 |
| ; VI-NEXT: v_add_f16_sdwa v4, v4, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v5, s20, v19 |
| ; VI-NEXT: v_or_b32_e32 v4, v5, v4 |
| ; VI-NEXT: v_mov_b32_e32 v5, s90 |
| ; VI-NEXT: v_add_f16_sdwa v5, v5, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v6, s21, v19 |
| ; VI-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; VI-NEXT: v_mov_b32_e32 v6, s89 |
| ; VI-NEXT: v_add_f16_sdwa v6, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v7, s22, v19 |
| ; VI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; VI-NEXT: v_mov_b32_e32 v7, s88 |
| ; VI-NEXT: v_add_f16_sdwa v7, v7, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v8, s23, v19 |
| ; VI-NEXT: v_or_b32_e32 v7, v8, v7 |
| ; VI-NEXT: v_mov_b32_e32 v8, s79 |
| ; VI-NEXT: v_add_f16_sdwa v8, v8, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v9, s24, v19 |
| ; VI-NEXT: v_or_b32_e32 v8, v9, v8 |
| ; VI-NEXT: v_mov_b32_e32 v9, s77 |
| ; VI-NEXT: v_add_f16_sdwa v9, v9, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v10, s25, v19 |
| ; VI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; VI-NEXT: v_mov_b32_e32 v10, s74 |
| ; VI-NEXT: v_add_f16_sdwa v10, v10, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v11, s26, v19 |
| ; VI-NEXT: v_or_b32_e32 v10, v11, v10 |
| ; VI-NEXT: v_mov_b32_e32 v11, s72 |
| ; VI-NEXT: v_add_f16_sdwa v11, v11, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v12, s27, v19 |
| ; VI-NEXT: v_or_b32_e32 v11, v12, v11 |
| ; VI-NEXT: v_mov_b32_e32 v12, s13 |
| ; VI-NEXT: v_add_f16_sdwa v12, v12, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v13, s28, v19 |
| ; VI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; VI-NEXT: v_mov_b32_e32 v13, s10 |
| ; VI-NEXT: v_add_f16_sdwa v13, v13, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v14, s29, v19 |
| ; VI-NEXT: v_or_b32_e32 v13, v14, v13 |
| ; VI-NEXT: v_mov_b32_e32 v14, s78 |
| ; VI-NEXT: v_add_f16_sdwa v14, v14, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v15, s76, v19 |
| ; VI-NEXT: v_or_b32_e32 v14, v15, v14 |
| ; VI-NEXT: v_mov_b32_e32 v15, s75 |
| ; VI-NEXT: v_add_f16_sdwa v15, v15, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v16, s73, v19 |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: v_mov_b32_e32 v16, s15 |
| ; VI-NEXT: v_add_f16_sdwa v16, v16, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v17, s14, v19 |
| ; VI-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; VI-NEXT: v_mov_b32_e32 v17, s12 |
| ; VI-NEXT: v_add_f16_sdwa v17, v17, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v18, s11, v19 |
| ; VI-NEXT: v_or_b32_e32 v17, v18, v17 |
| ; VI-NEXT: v_mov_b32_e32 v18, s9 |
| ; VI-NEXT: v_add_f16_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v20, s8, v19 |
| ; VI-NEXT: v_or_b32_e32 v18, v20, v18 |
| ; VI-NEXT: v_mov_b32_e32 v20, s7 |
| ; VI-NEXT: v_add_f16_sdwa v20, v20, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v19, s6, v19 |
| ; VI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; VI-NEXT: s_branch .LBB19_5 |
| ; VI-NEXT: .LBB19_3: |
| ; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; VI-NEXT: s_branch .LBB19_2 |
| ; VI-NEXT: .LBB19_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_mov_b32_e32 v20, s56 |
| ; VI-NEXT: v_mov_b32_e32 v21, s57 |
| ; VI-NEXT: v_mov_b32_e32 v22, s58 |
| ; VI-NEXT: v_mov_b32_e32 v23, s59 |
| ; VI-NEXT: v_mov_b32_e32 v24, s60 |
| ; VI-NEXT: v_mov_b32_e32 v25, s61 |
| ; VI-NEXT: v_mov_b32_e32 v26, s62 |
| ; VI-NEXT: v_mov_b32_e32 v27, s63 |
| ; VI-NEXT: v_mov_b32_e32 v28, s64 |
| ; VI-NEXT: v_mov_b32_e32 v29, s65 |
| ; VI-NEXT: v_mov_b32_e32 v30, s66 |
| ; VI-NEXT: v_mov_b32_e32 v31, s67 |
| ; VI-NEXT: .LBB19_5: ; %end |
| ; VI-NEXT: v_readlane_b32 s67, v32, 19 |
| ; VI-NEXT: v_readlane_b32 s66, v32, 18 |
| ; VI-NEXT: v_readlane_b32 s65, v32, 17 |
| ; VI-NEXT: v_readlane_b32 s64, v32, 16 |
| ; VI-NEXT: v_readlane_b32 s55, v32, 15 |
| ; VI-NEXT: v_readlane_b32 s54, v32, 14 |
| ; VI-NEXT: v_readlane_b32 s53, v32, 13 |
| ; VI-NEXT: v_readlane_b32 s52, v32, 12 |
| ; VI-NEXT: v_readlane_b32 s51, v32, 11 |
| ; VI-NEXT: v_readlane_b32 s50, v32, 10 |
| ; VI-NEXT: v_readlane_b32 s49, v32, 9 |
| ; VI-NEXT: v_readlane_b32 s48, v32, 8 |
| ; VI-NEXT: v_readlane_b32 s39, v32, 7 |
| ; VI-NEXT: v_readlane_b32 s38, v32, 6 |
| ; VI-NEXT: v_readlane_b32 s37, v32, 5 |
| ; VI-NEXT: v_readlane_b32 s36, v32, 4 |
| ; VI-NEXT: v_readlane_b32 s35, v32, 3 |
| ; VI-NEXT: v_readlane_b32 s34, v32, 2 |
| ; VI-NEXT: v_readlane_b32 s31, v32, 1 |
| ; VI-NEXT: v_readlane_b32 s30, v32, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40f16_to_v20i32_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_readfirstlane_b32 s56, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s58, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s60, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s62, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s72, v1 |
| ; GFX9-NEXT: v_readfirstlane_b32 s74, v0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: s_lshr_b32 s4, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s5, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s6, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s7, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s8, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s9, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s10, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s11, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s12, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s16, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s56, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s58, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s60, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s62, 16 |
| ; GFX9-NEXT: s_lshr_b32 s73, s72, 16 |
| ; GFX9-NEXT: s_lshr_b32 s75, s74, 16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s42, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_cmp_lg_u32 s42, 0 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s50, s74, s75 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s51, s72, s73 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s52, s62, s63 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s53, s60, s61 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s54, s58, s59 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s55, s56, s57 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB19_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB19_4 |
| ; GFX9-NEXT: .LBB19_2: ; %cmp.true |
| ; GFX9-NEXT: v_mov_b32_e32 v19, 0x200 |
| ; GFX9-NEXT: v_pk_add_f16 v0, s36, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, s37, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, s38, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, s39, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, s40, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, s41, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, s42, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, s43, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, s44, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, s45, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, s46, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, s47, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v12, s48, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, s49, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, s50, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v15, s51, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v16, s52, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v17, s53, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v18, s54, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v19, s55, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_branch .LBB19_5 |
| ; GFX9-NEXT: .LBB19_3: |
| ; GFX9-NEXT: s_branch .LBB19_2 |
| ; GFX9-NEXT: .LBB19_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB19_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v40f16_to_v20i32_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s45, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s46, v0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s56, v2 |
| ; GFX11-NEXT: s_lshr_b32 s41, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s6, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s5, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s4, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: s_lshr_b32 s58, s45, 16 |
| ; GFX11-NEXT: s_lshr_b32 s59, s46, 16 |
| ; GFX11-NEXT: s_mov_b32 s40, 0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s56, 0 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s57 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s47 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s44 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s18, s46, s59 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s19, s45, s58 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB19_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB19_4 |
| ; GFX11-NEXT: .LBB19_2: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v18, 0x200, s18 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v19, 0x200, s19 op_sel_hi:[0,1] |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB19_3: |
| ; GFX11-NEXT: s_branch .LBB19_2 |
| ; GFX11-NEXT: .LBB19_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31 |
| ; 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 <40 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <40 x half> %a1 to <20 x i32> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x half> %a to <20 x i32> |
| br label %end |
| |
| end: |
| %phi = phi <20 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x i32> %phi |
| } |
| |
| define <10 x i64> @bitcast_v20f32_to_v10i64(<20 x float> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v20f32_to_v10i64: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB20_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_add_f32_e32 v19, 1.0, v19 |
| ; SI-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; SI-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; SI-NEXT: v_add_f32_e32 v16, 1.0, v16 |
| ; SI-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; SI-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; SI-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; SI-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; SI-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; SI-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; SI-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; SI-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; SI-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; SI-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; SI-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; SI-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; SI-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; SI-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; SI-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; SI-NEXT: .LBB20_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20f32_to_v10i64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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 v19, 1.0, v19 |
| ; VI-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; VI-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; 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_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_v20f32_to_v10i64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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 v19, 1.0, v19 |
| ; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; 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_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_v20f32_to_v10i64: |
| ; 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, v20 |
| ; 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 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18 |
| ; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16 |
| ; 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 <20 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <20 x float> %a1 to <10 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x float> %a to <10 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <10 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x i64> %phi |
| } |
| |
| define inreg <10 x i64> @bitcast_v20f32_to_v10i64_scalar(<20 x float> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v20f32_to_v10i64_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; SI-NEXT: s_mov_b32 s49, s29 |
| ; SI-NEXT: s_mov_b32 s48, s28 |
| ; SI-NEXT: s_mov_b32 s47, s27 |
| ; SI-NEXT: s_mov_b32 s46, s26 |
| ; SI-NEXT: s_mov_b32 s45, s25 |
| ; SI-NEXT: s_mov_b32 s44, s24 |
| ; SI-NEXT: s_mov_b32 s43, s23 |
| ; SI-NEXT: s_mov_b32 s42, s22 |
| ; SI-NEXT: s_mov_b32 s41, s21 |
| ; SI-NEXT: s_mov_b32 s40, s20 |
| ; SI-NEXT: s_mov_b32 s39, s19 |
| ; SI-NEXT: s_mov_b32 s38, s18 |
| ; SI-NEXT: s_mov_b32 s37, s17 |
| ; SI-NEXT: s_mov_b32 s36, s16 |
| ; SI-NEXT: v_readfirstlane_b32 s55, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s54, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s53, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s52, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s51, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s50, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB21_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB21_4 |
| ; SI-NEXT: .LBB21_2: ; %cmp.true |
| ; SI-NEXT: v_add_f32_e64 v19, s55, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v18, s54, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v17, s53, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v16, s52, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v15, s51, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v14, s50, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v13, s49, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v12, s48, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v11, s47, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v10, s46, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v9, s45, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v8, s44, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v7, s43, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v6, s42, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v5, s41, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v4, s40, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v3, s39, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v2, s38, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v1, s37, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v0, s36, 1.0 |
| ; SI-NEXT: s_branch .LBB21_5 |
| ; SI-NEXT: .LBB21_3: |
| ; SI-NEXT: s_branch .LBB21_2 |
| ; SI-NEXT: .LBB21_4: |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_mov_b32_e32 v20, s56 |
| ; SI-NEXT: v_mov_b32_e32 v21, s57 |
| ; SI-NEXT: v_mov_b32_e32 v22, s58 |
| ; SI-NEXT: v_mov_b32_e32 v23, s59 |
| ; SI-NEXT: v_mov_b32_e32 v24, s60 |
| ; SI-NEXT: v_mov_b32_e32 v25, s61 |
| ; SI-NEXT: v_mov_b32_e32 v26, s62 |
| ; SI-NEXT: v_mov_b32_e32 v27, s63 |
| ; SI-NEXT: v_mov_b32_e32 v28, s64 |
| ; SI-NEXT: v_mov_b32_e32 v29, s65 |
| ; SI-NEXT: v_mov_b32_e32 v30, s66 |
| ; SI-NEXT: v_mov_b32_e32 v31, s67 |
| ; SI-NEXT: .LBB21_5: ; %end |
| ; SI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20f32_to_v10i64_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; VI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; VI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; VI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; VI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; VI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; VI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; VI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; VI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; VI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; VI-NEXT: s_mov_b32 s49, s29 |
| ; VI-NEXT: s_mov_b32 s48, s28 |
| ; VI-NEXT: s_mov_b32 s47, s27 |
| ; VI-NEXT: s_mov_b32 s46, s26 |
| ; VI-NEXT: s_mov_b32 s45, s25 |
| ; VI-NEXT: s_mov_b32 s44, s24 |
| ; VI-NEXT: s_mov_b32 s43, s23 |
| ; VI-NEXT: s_mov_b32 s42, s22 |
| ; VI-NEXT: s_mov_b32 s41, s21 |
| ; VI-NEXT: s_mov_b32 s40, s20 |
| ; VI-NEXT: s_mov_b32 s39, s19 |
| ; VI-NEXT: s_mov_b32 s38, s18 |
| ; VI-NEXT: s_mov_b32 s37, s17 |
| ; VI-NEXT: s_mov_b32 s36, s16 |
| ; VI-NEXT: v_readfirstlane_b32 s55, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s54, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s53, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s52, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s51, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s50, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB21_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB21_4 |
| ; VI-NEXT: .LBB21_2: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e64 v19, s55, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v18, s54, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v17, s53, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v16, s52, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v15, s51, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v14, s50, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v13, s49, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v12, s48, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v11, s47, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v10, s46, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v9, s45, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v8, s44, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v7, s43, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v6, s42, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v5, s41, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v4, s40, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v3, s39, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v2, s38, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v1, s37, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v0, s36, 1.0 |
| ; VI-NEXT: s_branch .LBB21_5 |
| ; VI-NEXT: .LBB21_3: |
| ; VI-NEXT: s_branch .LBB21_2 |
| ; VI-NEXT: .LBB21_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_mov_b32_e32 v20, s56 |
| ; VI-NEXT: v_mov_b32_e32 v21, s57 |
| ; VI-NEXT: v_mov_b32_e32 v22, s58 |
| ; VI-NEXT: v_mov_b32_e32 v23, s59 |
| ; VI-NEXT: v_mov_b32_e32 v24, s60 |
| ; VI-NEXT: v_mov_b32_e32 v25, s61 |
| ; VI-NEXT: v_mov_b32_e32 v26, s62 |
| ; VI-NEXT: v_mov_b32_e32 v27, s63 |
| ; VI-NEXT: v_mov_b32_e32 v28, s64 |
| ; VI-NEXT: v_mov_b32_e32 v29, s65 |
| ; VI-NEXT: v_mov_b32_e32 v30, s66 |
| ; VI-NEXT: v_mov_b32_e32 v31, s67 |
| ; VI-NEXT: .LBB21_5: ; %end |
| ; VI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; VI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; VI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; VI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; VI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; VI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; VI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; VI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; VI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; VI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; VI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; VI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v20f32_to_v10i64_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_mov_b32 s49, s29 |
| ; GFX9-NEXT: s_mov_b32 s48, s28 |
| ; GFX9-NEXT: s_mov_b32 s47, s27 |
| ; GFX9-NEXT: s_mov_b32 s46, s26 |
| ; GFX9-NEXT: s_mov_b32 s45, s25 |
| ; GFX9-NEXT: s_mov_b32 s44, s24 |
| ; GFX9-NEXT: s_mov_b32 s43, s23 |
| ; GFX9-NEXT: s_mov_b32 s42, s22 |
| ; GFX9-NEXT: s_mov_b32 s41, s21 |
| ; GFX9-NEXT: s_mov_b32 s40, s20 |
| ; GFX9-NEXT: s_mov_b32 s39, s19 |
| ; GFX9-NEXT: s_mov_b32 s38, s18 |
| ; GFX9-NEXT: s_mov_b32 s37, s17 |
| ; GFX9-NEXT: s_mov_b32 s36, s16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s55, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s54, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s53, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s52, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s51, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s50, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB21_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB21_4 |
| ; GFX9-NEXT: .LBB21_2: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e64 v19, s55, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v18, s54, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v17, s53, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v16, s52, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v15, s51, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v14, s50, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v13, s49, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v12, s48, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v11, s47, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v10, s46, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v9, s45, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v8, s44, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v7, s43, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v6, s42, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v5, s41, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v4, s40, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v3, s39, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v2, s38, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v1, s37, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v0, s36, 1.0 |
| ; GFX9-NEXT: s_branch .LBB21_5 |
| ; GFX9-NEXT: .LBB21_3: |
| ; GFX9-NEXT: s_branch .LBB21_2 |
| ; GFX9-NEXT: .LBB21_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB21_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v20f32_to_v10i64_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_xor_saveexec_b32 s4, -1 |
| ; GFX11-NEXT: scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill |
| ; GFX11-NEXT: s_mov_b32 exec_lo, s4 |
| ; GFX11-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX11-NEXT: s_mov_b32 s36, s0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s0, v2 |
| ; GFX11-NEXT: s_mov_b32 s47, s23 |
| ; GFX11-NEXT: s_mov_b32 s46, s22 |
| ; GFX11-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX11-NEXT: s_mov_b32 s45, s21 |
| ; GFX11-NEXT: s_mov_b32 s44, s20 |
| ; GFX11-NEXT: s_mov_b32 s43, s19 |
| ; GFX11-NEXT: s_mov_b32 s42, s18 |
| ; GFX11-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX11-NEXT: s_mov_b32 s41, s17 |
| ; GFX11-NEXT: s_mov_b32 s40, s16 |
| ; GFX11-NEXT: s_mov_b32 s38, s2 |
| ; GFX11-NEXT: s_mov_b32 s37, s1 |
| ; GFX11-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX11-NEXT: s_mov_b32 s39, s3 |
| ; GFX11-NEXT: s_cmp_lg_u32 s0, 0 |
| ; GFX11-NEXT: s_mov_b32 s0, 0 |
| ; GFX11-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX11-NEXT: s_mov_b32 s48, s24 |
| ; GFX11-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX11-NEXT: s_mov_b32 s49, s25 |
| ; GFX11-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX11-NEXT: s_mov_b32 s50, s26 |
| ; GFX11-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX11-NEXT: s_mov_b32 s51, s27 |
| ; GFX11-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX11-NEXT: s_mov_b32 s52, s28 |
| ; GFX11-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX11-NEXT: s_mov_b32 s53, s29 |
| ; GFX11-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX11-NEXT: v_readfirstlane_b32 s54, v0 |
| ; GFX11-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX11-NEXT: v_readfirstlane_b32 s55, v1 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB21_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB21_4 |
| ; GFX11-NEXT: .LBB21_2: ; %cmp.true |
| ; GFX11-NEXT: v_add_f32_e64 v19, s55, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v18, s54, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v17, s53, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v16, s52, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v15, s51, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v14, s50, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v13, s49, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v12, s48, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v11, s47, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v10, s46, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v9, s45, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v8, s44, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v7, s43, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v6, s42, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v5, s41, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v4, s40, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v3, s39, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v2, s38, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v1, s37, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v0, s36, 1.0 |
| ; GFX11-NEXT: s_branch .LBB21_5 |
| ; GFX11-NEXT: .LBB21_3: |
| ; GFX11-NEXT: s_branch .LBB21_2 |
| ; GFX11-NEXT: .LBB21_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67 |
| ; GFX11-NEXT: .LBB21_5: ; %end |
| ; GFX11-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX11-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX11-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX11-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX11-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX11-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX11-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX11-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX11-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX11-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX11-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX11-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX11-NEXT: s_xor_saveexec_b32 s0, -1 |
| ; GFX11-NEXT: scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload |
| ; GFX11-NEXT: s_mov_b32 exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; 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 <20 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <20 x float> %a1 to <10 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x float> %a to <10 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <10 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x i64> %phi |
| } |
| |
| define <20 x float> @bitcast_v10i64_to_v20f32(<10 x i64> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v10i64_to_v20f32: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB22_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; SI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; SI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; SI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; SI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; SI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; SI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; SI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; SI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; SI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; SI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; SI-NEXT: .LBB22_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10i64_to_v20f32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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_u32_e32 v18, vcc, 3, v18 |
| ; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc |
| ; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16 |
| ; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc |
| ; 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: .LBB22_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10i64_to_v20f32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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_co_u32_e32 v18, vcc, 3, v18 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, 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: 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: .LBB22_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v10i64_to_v20f32: |
| ; 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, v20 |
| ; 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_add_co_u32 v18, vcc_lo, v18, 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 v19, null, 0, v19, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo |
| ; 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: .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 = add <10 x i64> %a, splat (i64 3) |
| %a2 = bitcast <10 x i64> %a1 to <20 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x i64> %a to <20 x float> |
| br label %end |
| |
| end: |
| %phi = phi <20 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x float> %phi |
| } |
| |
| define inreg <20 x float> @bitcast_v10i64_to_v20f32_scalar(<10 x i64> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v10i64_to_v20f32_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB23_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB23_3 |
| ; SI-NEXT: .LBB23_2: ; %cmp.true |
| ; SI-NEXT: s_add_u32 s7, s7, 3 |
| ; SI-NEXT: s_addc_u32 s6, s6, 0 |
| ; SI-NEXT: s_add_u32 s9, s9, 3 |
| ; SI-NEXT: s_addc_u32 s8, s8, 0 |
| ; SI-NEXT: s_add_u32 s11, s11, 3 |
| ; SI-NEXT: s_addc_u32 s10, s10, 0 |
| ; SI-NEXT: s_add_u32 s28, s28, 3 |
| ; SI-NEXT: s_addc_u32 s29, s29, 0 |
| ; SI-NEXT: s_add_u32 s26, s26, 3 |
| ; SI-NEXT: s_addc_u32 s27, s27, 0 |
| ; SI-NEXT: s_add_u32 s24, s24, 3 |
| ; SI-NEXT: s_addc_u32 s25, s25, 0 |
| ; SI-NEXT: s_add_u32 s22, s22, 3 |
| ; SI-NEXT: s_addc_u32 s23, s23, 0 |
| ; SI-NEXT: s_add_u32 s20, s20, 3 |
| ; SI-NEXT: s_addc_u32 s21, s21, 0 |
| ; SI-NEXT: s_add_u32 s18, s18, 3 |
| ; SI-NEXT: s_addc_u32 s19, s19, 0 |
| ; SI-NEXT: s_add_u32 s16, s16, 3 |
| ; SI-NEXT: s_addc_u32 s17, s17, 0 |
| ; SI-NEXT: .LBB23_3: ; %end |
| ; SI-NEXT: v_mov_b32_e32 v0, s16 |
| ; SI-NEXT: v_mov_b32_e32 v1, s17 |
| ; SI-NEXT: v_mov_b32_e32 v2, s18 |
| ; SI-NEXT: v_mov_b32_e32 v3, s19 |
| ; SI-NEXT: v_mov_b32_e32 v4, s20 |
| ; SI-NEXT: v_mov_b32_e32 v5, s21 |
| ; SI-NEXT: v_mov_b32_e32 v6, s22 |
| ; SI-NEXT: v_mov_b32_e32 v7, s23 |
| ; SI-NEXT: v_mov_b32_e32 v8, s24 |
| ; SI-NEXT: v_mov_b32_e32 v9, s25 |
| ; SI-NEXT: v_mov_b32_e32 v10, s26 |
| ; SI-NEXT: v_mov_b32_e32 v11, s27 |
| ; SI-NEXT: v_mov_b32_e32 v12, s28 |
| ; SI-NEXT: v_mov_b32_e32 v13, s29 |
| ; SI-NEXT: v_mov_b32_e32 v14, s11 |
| ; SI-NEXT: v_mov_b32_e32 v15, s10 |
| ; SI-NEXT: v_mov_b32_e32 v16, s9 |
| ; SI-NEXT: v_mov_b32_e32 v17, s8 |
| ; SI-NEXT: v_mov_b32_e32 v18, s7 |
| ; SI-NEXT: v_mov_b32_e32 v19, s6 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB23_4: |
| ; SI-NEXT: s_branch .LBB23_2 |
| ; |
| ; VI-LABEL: bitcast_v10i64_to_v20f32_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB23_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB23_3 |
| ; VI-NEXT: .LBB23_2: ; %cmp.true |
| ; VI-NEXT: s_add_u32 s7, s7, 3 |
| ; VI-NEXT: s_addc_u32 s6, s6, 0 |
| ; VI-NEXT: s_add_u32 s9, s9, 3 |
| ; VI-NEXT: s_addc_u32 s8, s8, 0 |
| ; VI-NEXT: s_add_u32 s11, s11, 3 |
| ; VI-NEXT: s_addc_u32 s10, s10, 0 |
| ; VI-NEXT: s_add_u32 s28, s28, 3 |
| ; VI-NEXT: s_addc_u32 s29, s29, 0 |
| ; VI-NEXT: s_add_u32 s26, s26, 3 |
| ; VI-NEXT: s_addc_u32 s27, s27, 0 |
| ; VI-NEXT: s_add_u32 s24, s24, 3 |
| ; VI-NEXT: s_addc_u32 s25, s25, 0 |
| ; VI-NEXT: s_add_u32 s22, s22, 3 |
| ; VI-NEXT: s_addc_u32 s23, s23, 0 |
| ; VI-NEXT: s_add_u32 s20, s20, 3 |
| ; VI-NEXT: s_addc_u32 s21, s21, 0 |
| ; VI-NEXT: s_add_u32 s18, s18, 3 |
| ; VI-NEXT: s_addc_u32 s19, s19, 0 |
| ; VI-NEXT: s_add_u32 s16, s16, 3 |
| ; VI-NEXT: s_addc_u32 s17, s17, 0 |
| ; VI-NEXT: .LBB23_3: ; %end |
| ; VI-NEXT: v_mov_b32_e32 v0, s16 |
| ; VI-NEXT: v_mov_b32_e32 v1, s17 |
| ; VI-NEXT: v_mov_b32_e32 v2, s18 |
| ; VI-NEXT: v_mov_b32_e32 v3, s19 |
| ; VI-NEXT: v_mov_b32_e32 v4, s20 |
| ; VI-NEXT: v_mov_b32_e32 v5, s21 |
| ; VI-NEXT: v_mov_b32_e32 v6, s22 |
| ; VI-NEXT: v_mov_b32_e32 v7, s23 |
| ; VI-NEXT: v_mov_b32_e32 v8, s24 |
| ; VI-NEXT: v_mov_b32_e32 v9, s25 |
| ; VI-NEXT: v_mov_b32_e32 v10, s26 |
| ; VI-NEXT: v_mov_b32_e32 v11, s27 |
| ; VI-NEXT: v_mov_b32_e32 v12, s28 |
| ; VI-NEXT: v_mov_b32_e32 v13, s29 |
| ; VI-NEXT: v_mov_b32_e32 v14, s11 |
| ; VI-NEXT: v_mov_b32_e32 v15, s10 |
| ; VI-NEXT: v_mov_b32_e32 v16, s9 |
| ; VI-NEXT: v_mov_b32_e32 v17, s8 |
| ; VI-NEXT: v_mov_b32_e32 v18, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB23_4: |
| ; VI-NEXT: s_branch .LBB23_2 |
| ; |
| ; GFX9-LABEL: bitcast_v10i64_to_v20f32_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s10, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s11, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB23_4 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB23_3 |
| ; GFX9-NEXT: .LBB23_2: ; %cmp.true |
| ; GFX9-NEXT: s_add_u32 s7, s7, 3 |
| ; GFX9-NEXT: s_addc_u32 s6, s6, 0 |
| ; GFX9-NEXT: s_add_u32 s9, s9, 3 |
| ; GFX9-NEXT: s_addc_u32 s8, s8, 0 |
| ; GFX9-NEXT: s_add_u32 s11, s11, 3 |
| ; GFX9-NEXT: s_addc_u32 s10, s10, 0 |
| ; GFX9-NEXT: s_add_u32 s28, s28, 3 |
| ; GFX9-NEXT: s_addc_u32 s29, s29, 0 |
| ; GFX9-NEXT: s_add_u32 s26, s26, 3 |
| ; GFX9-NEXT: s_addc_u32 s27, s27, 0 |
| ; GFX9-NEXT: s_add_u32 s24, s24, 3 |
| ; GFX9-NEXT: s_addc_u32 s25, s25, 0 |
| ; GFX9-NEXT: s_add_u32 s22, s22, 3 |
| ; GFX9-NEXT: s_addc_u32 s23, s23, 0 |
| ; GFX9-NEXT: s_add_u32 s20, s20, 3 |
| ; GFX9-NEXT: s_addc_u32 s21, s21, 0 |
| ; GFX9-NEXT: s_add_u32 s18, s18, 3 |
| ; GFX9-NEXT: s_addc_u32 s19, s19, 0 |
| ; GFX9-NEXT: s_add_u32 s16, s16, 3 |
| ; GFX9-NEXT: s_addc_u32 s17, s17, 0 |
| ; GFX9-NEXT: .LBB23_3: ; %end |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s28 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s29 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s6 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; GFX9-NEXT: .LBB23_4: |
| ; GFX9-NEXT: s_branch .LBB23_2 |
| ; |
| ; GFX11-LABEL: bitcast_v10i64_to_v20f32_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB23_4 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB23_3 |
| ; GFX11-NEXT: .LBB23_2: ; %cmp.true |
| ; GFX11-NEXT: s_add_u32 s5, s5, 3 |
| ; GFX11-NEXT: s_addc_u32 s4, s4, 0 |
| ; GFX11-NEXT: s_add_u32 s28, s28, 3 |
| ; GFX11-NEXT: s_addc_u32 s29, s29, 0 |
| ; GFX11-NEXT: s_add_u32 s26, s26, 3 |
| ; GFX11-NEXT: s_addc_u32 s27, s27, 0 |
| ; GFX11-NEXT: s_add_u32 s24, s24, 3 |
| ; GFX11-NEXT: s_addc_u32 s25, s25, 0 |
| ; GFX11-NEXT: s_add_u32 s22, s22, 3 |
| ; GFX11-NEXT: s_addc_u32 s23, s23, 0 |
| ; GFX11-NEXT: s_add_u32 s20, s20, 3 |
| ; GFX11-NEXT: s_addc_u32 s21, s21, 0 |
| ; GFX11-NEXT: s_add_u32 s18, s18, 3 |
| ; GFX11-NEXT: s_addc_u32 s19, s19, 0 |
| ; GFX11-NEXT: s_add_u32 s16, s16, 3 |
| ; GFX11-NEXT: s_addc_u32 s17, s17, 0 |
| ; GFX11-NEXT: s_add_u32 s2, s2, 3 |
| ; GFX11-NEXT: s_addc_u32 s3, s3, 0 |
| ; GFX11-NEXT: s_add_u32 s0, s0, 3 |
| ; GFX11-NEXT: s_addc_u32 s1, s1, 0 |
| ; GFX11-NEXT: .LBB23_3: ; %end |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB23_4: |
| ; GFX11-NEXT: s_branch .LBB23_2 |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <10 x i64> %a, splat (i64 3) |
| %a2 = bitcast <10 x i64> %a1 to <20 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x i64> %a to <20 x float> |
| br label %end |
| |
| end: |
| %phi = phi <20 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x float> %phi |
| } |
| |
| define <10 x double> @bitcast_v20f32_to_v10f64(<20 x float> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v20f32_to_v10f64: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB24_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_add_f32_e32 v19, 1.0, v19 |
| ; SI-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; SI-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; SI-NEXT: v_add_f32_e32 v16, 1.0, v16 |
| ; SI-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; SI-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; SI-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; SI-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; SI-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; SI-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; SI-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; SI-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; SI-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; SI-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; SI-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; SI-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; SI-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; SI-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; SI-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; SI-NEXT: .LBB24_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20f32_to_v10f64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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 .LBB24_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e32 v19, 1.0, v19 |
| ; VI-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; VI-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; 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_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: .LBB24_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v20f32_to_v10f64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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 .LBB24_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19 |
| ; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; 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_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: .LBB24_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v20f32_to_v10f64: |
| ; 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, v20 |
| ; 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 .LBB24_2 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18 |
| ; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16 |
| ; 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: .LBB24_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 <20 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <20 x float> %a1 to <10 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x float> %a to <10 x double> |
| br label %end |
| |
| end: |
| %phi = phi <10 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x double> %phi |
| } |
| |
| define inreg <10 x double> @bitcast_v20f32_to_v10f64_scalar(<20 x float> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v20f32_to_v10f64_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; SI-NEXT: s_mov_b32 s49, s29 |
| ; SI-NEXT: s_mov_b32 s48, s28 |
| ; SI-NEXT: s_mov_b32 s47, s27 |
| ; SI-NEXT: s_mov_b32 s46, s26 |
| ; SI-NEXT: s_mov_b32 s45, s25 |
| ; SI-NEXT: s_mov_b32 s44, s24 |
| ; SI-NEXT: s_mov_b32 s43, s23 |
| ; SI-NEXT: s_mov_b32 s42, s22 |
| ; SI-NEXT: s_mov_b32 s41, s21 |
| ; SI-NEXT: s_mov_b32 s40, s20 |
| ; SI-NEXT: s_mov_b32 s39, s19 |
| ; SI-NEXT: s_mov_b32 s38, s18 |
| ; SI-NEXT: s_mov_b32 s37, s17 |
| ; SI-NEXT: s_mov_b32 s36, s16 |
| ; SI-NEXT: v_readfirstlane_b32 s55, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s54, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s53, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s52, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s51, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s50, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB25_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB25_4 |
| ; SI-NEXT: .LBB25_2: ; %cmp.true |
| ; SI-NEXT: v_add_f32_e64 v19, s55, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v18, s54, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v17, s53, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v16, s52, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v15, s51, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v14, s50, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v13, s49, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v12, s48, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v11, s47, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v10, s46, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v9, s45, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v8, s44, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v7, s43, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v6, s42, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v5, s41, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v4, s40, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v3, s39, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v2, s38, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v1, s37, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v0, s36, 1.0 |
| ; SI-NEXT: s_branch .LBB25_5 |
| ; SI-NEXT: .LBB25_3: |
| ; SI-NEXT: s_branch .LBB25_2 |
| ; SI-NEXT: .LBB25_4: |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_mov_b32_e32 v20, s56 |
| ; SI-NEXT: v_mov_b32_e32 v21, s57 |
| ; SI-NEXT: v_mov_b32_e32 v22, s58 |
| ; SI-NEXT: v_mov_b32_e32 v23, s59 |
| ; SI-NEXT: v_mov_b32_e32 v24, s60 |
| ; SI-NEXT: v_mov_b32_e32 v25, s61 |
| ; SI-NEXT: v_mov_b32_e32 v26, s62 |
| ; SI-NEXT: v_mov_b32_e32 v27, s63 |
| ; SI-NEXT: v_mov_b32_e32 v28, s64 |
| ; SI-NEXT: v_mov_b32_e32 v29, s65 |
| ; SI-NEXT: v_mov_b32_e32 v30, s66 |
| ; SI-NEXT: v_mov_b32_e32 v31, s67 |
| ; SI-NEXT: .LBB25_5: ; %end |
| ; SI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20f32_to_v10f64_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; VI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; VI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; VI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; VI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; VI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; VI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; VI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; VI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; VI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; VI-NEXT: s_mov_b32 s49, s29 |
| ; VI-NEXT: s_mov_b32 s48, s28 |
| ; VI-NEXT: s_mov_b32 s47, s27 |
| ; VI-NEXT: s_mov_b32 s46, s26 |
| ; VI-NEXT: s_mov_b32 s45, s25 |
| ; VI-NEXT: s_mov_b32 s44, s24 |
| ; VI-NEXT: s_mov_b32 s43, s23 |
| ; VI-NEXT: s_mov_b32 s42, s22 |
| ; VI-NEXT: s_mov_b32 s41, s21 |
| ; VI-NEXT: s_mov_b32 s40, s20 |
| ; VI-NEXT: s_mov_b32 s39, s19 |
| ; VI-NEXT: s_mov_b32 s38, s18 |
| ; VI-NEXT: s_mov_b32 s37, s17 |
| ; VI-NEXT: s_mov_b32 s36, s16 |
| ; VI-NEXT: v_readfirstlane_b32 s55, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s54, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s53, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s52, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s51, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s50, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB25_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB25_4 |
| ; VI-NEXT: .LBB25_2: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e64 v19, s55, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v18, s54, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v17, s53, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v16, s52, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v15, s51, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v14, s50, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v13, s49, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v12, s48, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v11, s47, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v10, s46, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v9, s45, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v8, s44, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v7, s43, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v6, s42, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v5, s41, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v4, s40, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v3, s39, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v2, s38, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v1, s37, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v0, s36, 1.0 |
| ; VI-NEXT: s_branch .LBB25_5 |
| ; VI-NEXT: .LBB25_3: |
| ; VI-NEXT: s_branch .LBB25_2 |
| ; VI-NEXT: .LBB25_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_mov_b32_e32 v20, s56 |
| ; VI-NEXT: v_mov_b32_e32 v21, s57 |
| ; VI-NEXT: v_mov_b32_e32 v22, s58 |
| ; VI-NEXT: v_mov_b32_e32 v23, s59 |
| ; VI-NEXT: v_mov_b32_e32 v24, s60 |
| ; VI-NEXT: v_mov_b32_e32 v25, s61 |
| ; VI-NEXT: v_mov_b32_e32 v26, s62 |
| ; VI-NEXT: v_mov_b32_e32 v27, s63 |
| ; VI-NEXT: v_mov_b32_e32 v28, s64 |
| ; VI-NEXT: v_mov_b32_e32 v29, s65 |
| ; VI-NEXT: v_mov_b32_e32 v30, s66 |
| ; VI-NEXT: v_mov_b32_e32 v31, s67 |
| ; VI-NEXT: .LBB25_5: ; %end |
| ; VI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; VI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; VI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; VI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; VI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; VI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; VI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; VI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; VI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; VI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; VI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; VI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v20f32_to_v10f64_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_mov_b32 s49, s29 |
| ; GFX9-NEXT: s_mov_b32 s48, s28 |
| ; GFX9-NEXT: s_mov_b32 s47, s27 |
| ; GFX9-NEXT: s_mov_b32 s46, s26 |
| ; GFX9-NEXT: s_mov_b32 s45, s25 |
| ; GFX9-NEXT: s_mov_b32 s44, s24 |
| ; GFX9-NEXT: s_mov_b32 s43, s23 |
| ; GFX9-NEXT: s_mov_b32 s42, s22 |
| ; GFX9-NEXT: s_mov_b32 s41, s21 |
| ; GFX9-NEXT: s_mov_b32 s40, s20 |
| ; GFX9-NEXT: s_mov_b32 s39, s19 |
| ; GFX9-NEXT: s_mov_b32 s38, s18 |
| ; GFX9-NEXT: s_mov_b32 s37, s17 |
| ; GFX9-NEXT: s_mov_b32 s36, s16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s55, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s54, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s53, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s52, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s51, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s50, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB25_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB25_4 |
| ; GFX9-NEXT: .LBB25_2: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e64 v19, s55, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v18, s54, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v17, s53, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v16, s52, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v15, s51, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v14, s50, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v13, s49, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v12, s48, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v11, s47, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v10, s46, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v9, s45, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v8, s44, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v7, s43, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v6, s42, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v5, s41, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v4, s40, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v3, s39, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v2, s38, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v1, s37, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v0, s36, 1.0 |
| ; GFX9-NEXT: s_branch .LBB25_5 |
| ; GFX9-NEXT: .LBB25_3: |
| ; GFX9-NEXT: s_branch .LBB25_2 |
| ; GFX9-NEXT: .LBB25_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB25_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v20f32_to_v10f64_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_xor_saveexec_b32 s4, -1 |
| ; GFX11-NEXT: scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill |
| ; GFX11-NEXT: s_mov_b32 exec_lo, s4 |
| ; GFX11-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX11-NEXT: s_mov_b32 s36, s0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s0, v2 |
| ; GFX11-NEXT: s_mov_b32 s47, s23 |
| ; GFX11-NEXT: s_mov_b32 s46, s22 |
| ; GFX11-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX11-NEXT: s_mov_b32 s45, s21 |
| ; GFX11-NEXT: s_mov_b32 s44, s20 |
| ; GFX11-NEXT: s_mov_b32 s43, s19 |
| ; GFX11-NEXT: s_mov_b32 s42, s18 |
| ; GFX11-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX11-NEXT: s_mov_b32 s41, s17 |
| ; GFX11-NEXT: s_mov_b32 s40, s16 |
| ; GFX11-NEXT: s_mov_b32 s38, s2 |
| ; GFX11-NEXT: s_mov_b32 s37, s1 |
| ; GFX11-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX11-NEXT: s_mov_b32 s39, s3 |
| ; GFX11-NEXT: s_cmp_lg_u32 s0, 0 |
| ; GFX11-NEXT: s_mov_b32 s0, 0 |
| ; GFX11-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX11-NEXT: s_mov_b32 s48, s24 |
| ; GFX11-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX11-NEXT: s_mov_b32 s49, s25 |
| ; GFX11-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX11-NEXT: s_mov_b32 s50, s26 |
| ; GFX11-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX11-NEXT: s_mov_b32 s51, s27 |
| ; GFX11-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX11-NEXT: s_mov_b32 s52, s28 |
| ; GFX11-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX11-NEXT: s_mov_b32 s53, s29 |
| ; GFX11-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX11-NEXT: v_readfirstlane_b32 s54, v0 |
| ; GFX11-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX11-NEXT: v_readfirstlane_b32 s55, v1 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB25_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB25_4 |
| ; GFX11-NEXT: .LBB25_2: ; %cmp.true |
| ; GFX11-NEXT: v_add_f32_e64 v19, s55, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v18, s54, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v17, s53, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v16, s52, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v15, s51, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v14, s50, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v13, s49, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v12, s48, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v11, s47, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v10, s46, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v9, s45, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v8, s44, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v7, s43, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v6, s42, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v5, s41, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v4, s40, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v3, s39, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v2, s38, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v1, s37, 1.0 |
| ; GFX11-NEXT: v_add_f32_e64 v0, s36, 1.0 |
| ; GFX11-NEXT: s_branch .LBB25_5 |
| ; GFX11-NEXT: .LBB25_3: |
| ; GFX11-NEXT: s_branch .LBB25_2 |
| ; GFX11-NEXT: .LBB25_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67 |
| ; GFX11-NEXT: .LBB25_5: ; %end |
| ; GFX11-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX11-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX11-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX11-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX11-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX11-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX11-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX11-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX11-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX11-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX11-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX11-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX11-NEXT: s_xor_saveexec_b32 s0, -1 |
| ; GFX11-NEXT: scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload |
| ; GFX11-NEXT: s_mov_b32 exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; 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 <20 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <20 x float> %a1 to <10 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x float> %a to <10 x double> |
| br label %end |
| |
| end: |
| %phi = phi <10 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x double> %phi |
| } |
| |
| define <20 x float> @bitcast_v10f64_to_v20f32(<10 x double> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v10f64_to_v20f32: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB26_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0 |
| ; SI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; SI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; SI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; SI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; SI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; SI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; SI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; SI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; SI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; SI-NEXT: .LBB26_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10f64_to_v20f32: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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_f64 v[18:19], v[18:19], 1.0 |
| ; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; 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: .LBB26_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10f64_to_v20f32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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_f64 v[18:19], v[18:19], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; 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: .LBB26_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v10f64_to_v20f32: |
| ; 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, v20 |
| ; 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_f64 v[18:19], v[18:19], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; 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: .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 = fadd <10 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <10 x double> %a1 to <20 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x double> %a to <20 x float> |
| br label %end |
| |
| end: |
| %phi = phi <20 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x float> %phi |
| } |
| |
| define inreg <20 x float> @bitcast_v10f64_to_v20f32_scalar(<10 x double> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v10f64_to_v20f32_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; SI-NEXT: s_mov_b32 s49, s29 |
| ; SI-NEXT: s_mov_b32 s48, s28 |
| ; SI-NEXT: s_mov_b32 s47, s27 |
| ; SI-NEXT: s_mov_b32 s46, s26 |
| ; SI-NEXT: s_mov_b32 s45, s25 |
| ; SI-NEXT: s_mov_b32 s44, s24 |
| ; SI-NEXT: s_mov_b32 s43, s23 |
| ; SI-NEXT: s_mov_b32 s42, s22 |
| ; SI-NEXT: s_mov_b32 s41, s21 |
| ; SI-NEXT: s_mov_b32 s40, s20 |
| ; SI-NEXT: s_mov_b32 s39, s19 |
| ; SI-NEXT: s_mov_b32 s38, s18 |
| ; SI-NEXT: s_mov_b32 s37, s17 |
| ; SI-NEXT: s_mov_b32 s36, s16 |
| ; SI-NEXT: v_readfirstlane_b32 s55, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s54, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s53, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s52, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s51, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s50, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB27_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB27_4 |
| ; SI-NEXT: .LBB27_2: ; %cmp.true |
| ; SI-NEXT: v_add_f64 v[18:19], s[54:55], 1.0 |
| ; SI-NEXT: v_add_f64 v[16:17], s[52:53], 1.0 |
| ; SI-NEXT: v_add_f64 v[14:15], s[50:51], 1.0 |
| ; SI-NEXT: v_add_f64 v[12:13], s[48:49], 1.0 |
| ; SI-NEXT: v_add_f64 v[10:11], s[46:47], 1.0 |
| ; SI-NEXT: v_add_f64 v[8:9], s[44:45], 1.0 |
| ; SI-NEXT: v_add_f64 v[6:7], s[42:43], 1.0 |
| ; SI-NEXT: v_add_f64 v[4:5], s[40:41], 1.0 |
| ; SI-NEXT: v_add_f64 v[2:3], s[38:39], 1.0 |
| ; SI-NEXT: v_add_f64 v[0:1], s[36:37], 1.0 |
| ; SI-NEXT: s_branch .LBB27_5 |
| ; SI-NEXT: .LBB27_3: |
| ; SI-NEXT: s_branch .LBB27_2 |
| ; SI-NEXT: .LBB27_4: |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_mov_b32_e32 v20, s56 |
| ; SI-NEXT: v_mov_b32_e32 v21, s57 |
| ; SI-NEXT: v_mov_b32_e32 v22, s58 |
| ; SI-NEXT: v_mov_b32_e32 v23, s59 |
| ; SI-NEXT: v_mov_b32_e32 v24, s60 |
| ; SI-NEXT: v_mov_b32_e32 v25, s61 |
| ; SI-NEXT: v_mov_b32_e32 v26, s62 |
| ; SI-NEXT: v_mov_b32_e32 v27, s63 |
| ; SI-NEXT: v_mov_b32_e32 v28, s64 |
| ; SI-NEXT: v_mov_b32_e32 v29, s65 |
| ; SI-NEXT: v_mov_b32_e32 v30, s66 |
| ; SI-NEXT: v_mov_b32_e32 v31, s67 |
| ; SI-NEXT: .LBB27_5: ; %end |
| ; SI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10f64_to_v20f32_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; VI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; VI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; VI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; VI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; VI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; VI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; VI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; VI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; VI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; VI-NEXT: s_mov_b32 s49, s29 |
| ; VI-NEXT: s_mov_b32 s48, s28 |
| ; VI-NEXT: s_mov_b32 s47, s27 |
| ; VI-NEXT: s_mov_b32 s46, s26 |
| ; VI-NEXT: s_mov_b32 s45, s25 |
| ; VI-NEXT: s_mov_b32 s44, s24 |
| ; VI-NEXT: s_mov_b32 s43, s23 |
| ; VI-NEXT: s_mov_b32 s42, s22 |
| ; VI-NEXT: s_mov_b32 s41, s21 |
| ; VI-NEXT: s_mov_b32 s40, s20 |
| ; VI-NEXT: s_mov_b32 s39, s19 |
| ; VI-NEXT: s_mov_b32 s38, s18 |
| ; VI-NEXT: s_mov_b32 s37, s17 |
| ; VI-NEXT: s_mov_b32 s36, s16 |
| ; VI-NEXT: v_readfirstlane_b32 s55, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s54, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s53, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s52, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s51, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s50, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB27_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB27_4 |
| ; VI-NEXT: .LBB27_2: ; %cmp.true |
| ; VI-NEXT: v_add_f64 v[18:19], s[54:55], 1.0 |
| ; VI-NEXT: v_add_f64 v[16:17], s[52:53], 1.0 |
| ; VI-NEXT: v_add_f64 v[14:15], s[50:51], 1.0 |
| ; VI-NEXT: v_add_f64 v[12:13], s[48:49], 1.0 |
| ; VI-NEXT: v_add_f64 v[10:11], s[46:47], 1.0 |
| ; VI-NEXT: v_add_f64 v[8:9], s[44:45], 1.0 |
| ; VI-NEXT: v_add_f64 v[6:7], s[42:43], 1.0 |
| ; VI-NEXT: v_add_f64 v[4:5], s[40:41], 1.0 |
| ; VI-NEXT: v_add_f64 v[2:3], s[38:39], 1.0 |
| ; VI-NEXT: v_add_f64 v[0:1], s[36:37], 1.0 |
| ; VI-NEXT: s_branch .LBB27_5 |
| ; VI-NEXT: .LBB27_3: |
| ; VI-NEXT: s_branch .LBB27_2 |
| ; VI-NEXT: .LBB27_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_mov_b32_e32 v20, s56 |
| ; VI-NEXT: v_mov_b32_e32 v21, s57 |
| ; VI-NEXT: v_mov_b32_e32 v22, s58 |
| ; VI-NEXT: v_mov_b32_e32 v23, s59 |
| ; VI-NEXT: v_mov_b32_e32 v24, s60 |
| ; VI-NEXT: v_mov_b32_e32 v25, s61 |
| ; VI-NEXT: v_mov_b32_e32 v26, s62 |
| ; VI-NEXT: v_mov_b32_e32 v27, s63 |
| ; VI-NEXT: v_mov_b32_e32 v28, s64 |
| ; VI-NEXT: v_mov_b32_e32 v29, s65 |
| ; VI-NEXT: v_mov_b32_e32 v30, s66 |
| ; VI-NEXT: v_mov_b32_e32 v31, s67 |
| ; VI-NEXT: .LBB27_5: ; %end |
| ; VI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; VI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; VI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; VI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; VI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; VI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; VI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; VI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; VI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; VI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; VI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; VI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10f64_to_v20f32_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_mov_b32 s49, s29 |
| ; GFX9-NEXT: s_mov_b32 s48, s28 |
| ; GFX9-NEXT: s_mov_b32 s47, s27 |
| ; GFX9-NEXT: s_mov_b32 s46, s26 |
| ; GFX9-NEXT: s_mov_b32 s45, s25 |
| ; GFX9-NEXT: s_mov_b32 s44, s24 |
| ; GFX9-NEXT: s_mov_b32 s43, s23 |
| ; GFX9-NEXT: s_mov_b32 s42, s22 |
| ; GFX9-NEXT: s_mov_b32 s41, s21 |
| ; GFX9-NEXT: s_mov_b32 s40, s20 |
| ; GFX9-NEXT: s_mov_b32 s39, s19 |
| ; GFX9-NEXT: s_mov_b32 s38, s18 |
| ; GFX9-NEXT: s_mov_b32 s37, s17 |
| ; GFX9-NEXT: s_mov_b32 s36, s16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s55, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s54, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s53, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s52, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s51, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s50, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB27_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB27_4 |
| ; GFX9-NEXT: .LBB27_2: ; %cmp.true |
| ; GFX9-NEXT: v_add_f64 v[18:19], s[54:55], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[16:17], s[52:53], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[14:15], s[50:51], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[12:13], s[48:49], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[10:11], s[46:47], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[8:9], s[44:45], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[6:7], s[42:43], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[4:5], s[40:41], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[2:3], s[38:39], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[0:1], s[36:37], 1.0 |
| ; GFX9-NEXT: s_branch .LBB27_5 |
| ; GFX9-NEXT: .LBB27_3: |
| ; GFX9-NEXT: s_branch .LBB27_2 |
| ; GFX9-NEXT: .LBB27_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB27_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v10f64_to_v20f32_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_xor_saveexec_b32 s4, -1 |
| ; GFX11-NEXT: scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill |
| ; GFX11-NEXT: s_mov_b32 exec_lo, s4 |
| ; GFX11-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX11-NEXT: s_mov_b32 s36, s0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s0, v2 |
| ; GFX11-NEXT: s_mov_b32 s47, s23 |
| ; GFX11-NEXT: s_mov_b32 s46, s22 |
| ; GFX11-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX11-NEXT: s_mov_b32 s45, s21 |
| ; GFX11-NEXT: s_mov_b32 s44, s20 |
| ; GFX11-NEXT: s_mov_b32 s43, s19 |
| ; GFX11-NEXT: s_mov_b32 s42, s18 |
| ; GFX11-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX11-NEXT: s_mov_b32 s41, s17 |
| ; GFX11-NEXT: s_mov_b32 s40, s16 |
| ; GFX11-NEXT: s_mov_b32 s38, s2 |
| ; GFX11-NEXT: s_mov_b32 s37, s1 |
| ; GFX11-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX11-NEXT: s_mov_b32 s39, s3 |
| ; GFX11-NEXT: s_cmp_lg_u32 s0, 0 |
| ; GFX11-NEXT: s_mov_b32 s0, 0 |
| ; GFX11-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX11-NEXT: s_mov_b32 s48, s24 |
| ; GFX11-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX11-NEXT: s_mov_b32 s49, s25 |
| ; GFX11-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX11-NEXT: s_mov_b32 s50, s26 |
| ; GFX11-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX11-NEXT: s_mov_b32 s51, s27 |
| ; GFX11-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX11-NEXT: s_mov_b32 s52, s28 |
| ; GFX11-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX11-NEXT: s_mov_b32 s53, s29 |
| ; GFX11-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX11-NEXT: v_readfirstlane_b32 s54, v0 |
| ; GFX11-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX11-NEXT: v_readfirstlane_b32 s55, v1 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB27_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB27_4 |
| ; GFX11-NEXT: .LBB27_2: ; %cmp.true |
| ; GFX11-NEXT: v_add_f64 v[18:19], s[54:55], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[16:17], s[52:53], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[14:15], s[50:51], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[12:13], s[48:49], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[10:11], s[46:47], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[8:9], s[44:45], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[6:7], s[42:43], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[4:5], s[40:41], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[2:3], s[38:39], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[0:1], s[36:37], 1.0 |
| ; GFX11-NEXT: s_branch .LBB27_5 |
| ; GFX11-NEXT: .LBB27_3: |
| ; GFX11-NEXT: s_branch .LBB27_2 |
| ; GFX11-NEXT: .LBB27_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67 |
| ; GFX11-NEXT: .LBB27_5: ; %end |
| ; GFX11-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX11-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX11-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX11-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX11-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX11-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX11-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX11-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX11-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX11-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX11-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX11-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX11-NEXT: s_xor_saveexec_b32 s0, -1 |
| ; GFX11-NEXT: scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload |
| ; GFX11-NEXT: s_mov_b32 exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; 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 <10 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <10 x double> %a1 to <20 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x double> %a to <20 x float> |
| br label %end |
| |
| end: |
| %phi = phi <20 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x float> %phi |
| } |
| |
| define <40 x i16> @bitcast_v20f32_to_v40i16(<20 x float> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v20f32_to_v40i16: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr31 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr29 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr26 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr25 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr24 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr23 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; implicit-def: $vgpr22 |
| ; SI-NEXT: ; implicit-def: $vgpr30 |
| ; SI-NEXT: ; implicit-def: $vgpr21 |
| ; SI-NEXT: ; implicit-def: $vgpr28 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr27 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB28_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v31, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v34, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB28_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB28_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; SI-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; SI-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; SI-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; SI-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; SI-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; SI-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; SI-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; SI-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; SI-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; SI-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; SI-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; SI-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; SI-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; SI-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; SI-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; SI-NEXT: v_add_f32_e32 v16, 1.0, v16 |
| ; SI-NEXT: v_add_f32_e32 v19, 1.0, v19 |
| ; SI-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v31, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v34, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB28_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v34 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v31 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v29 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v25 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v35 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v23 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v32 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v22 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v30 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v28 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v27 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v34 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v31 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v29 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v26 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v25 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v24 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20f32_to_v40i16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; 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: 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 .LBB28_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB28_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB28_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e32 v19, 1.0, v19 |
| ; VI-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; VI-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; 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_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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB28_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v20f32_to_v40i16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; 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 .LBB28_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB28_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB28_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19 |
| ; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; 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_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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB28_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_mov_b32 s4, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v39, v0, s4 |
| ; GFX9-NEXT: v_perm_b32 v1, v38, v1, s4 |
| ; GFX9-NEXT: v_perm_b32 v2, v37, v2, s4 |
| ; GFX9-NEXT: v_perm_b32 v3, v36, v3, s4 |
| ; GFX9-NEXT: v_perm_b32 v4, v35, v4, s4 |
| ; GFX9-NEXT: v_perm_b32 v5, v34, v5, s4 |
| ; GFX9-NEXT: v_perm_b32 v6, v33, v6, s4 |
| ; GFX9-NEXT: v_perm_b32 v7, v32, v7, s4 |
| ; GFX9-NEXT: v_perm_b32 v8, v31, v8, s4 |
| ; GFX9-NEXT: v_perm_b32 v9, v30, v9, s4 |
| ; GFX9-NEXT: v_perm_b32 v10, v29, v10, s4 |
| ; GFX9-NEXT: v_perm_b32 v11, v28, v11, s4 |
| ; GFX9-NEXT: v_perm_b32 v12, v27, v12, s4 |
| ; GFX9-NEXT: v_perm_b32 v13, v26, v13, s4 |
| ; GFX9-NEXT: v_perm_b32 v14, v25, v14, s4 |
| ; GFX9-NEXT: v_perm_b32 v15, v24, v15, s4 |
| ; GFX9-NEXT: v_perm_b32 v16, v23, v16, s4 |
| ; GFX9-NEXT: v_perm_b32 v17, v22, v17, s4 |
| ; GFX9-NEXT: v_perm_b32 v18, v21, v18, s4 |
| ; GFX9-NEXT: v_perm_b32 v19, v20, v19, s4 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v20f32_to_v40i16: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB28_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0 |
| ; GFX11-TRUE16-NEXT: .LBB28_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v20f32_to_v40i16: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB28_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB28_2: ; %Flow |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB28_4 |
| ; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB28_4: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v39, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v37, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v36, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v35, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v34, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v33, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v32, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v31, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v30, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v29, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v28, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v27, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v26, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v25, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v24, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v23, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v22, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v21, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v20, v19, 0x5040100 |
| ; GFX11-FAKE16-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 <20 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <20 x float> %a1 to <40 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x float> %a to <40 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <40 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x i16> %phi |
| } |
| |
| define inreg <40 x i16> @bitcast_v20f32_to_v40i16_scalar(<20 x float> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v20f32_to_v40i16_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s8, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s5, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s8, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB29_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_lshr_b32 s89, s5, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s72, s17, 16 |
| ; SI-NEXT: s_lshr_b64 s[10:11], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b64 s[12:13], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[14:15], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[40:41], s[28:29], 16 |
| ; SI-NEXT: s_lshr_b64 s[42:43], s[26:27], 16 |
| ; SI-NEXT: s_lshr_b64 s[44:45], s[24:25], 16 |
| ; SI-NEXT: s_lshr_b64 s[46:47], s[22:23], 16 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[20:21], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[18:19], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[16:17], 16 |
| ; SI-NEXT: s_cbranch_execnz .LBB29_4 |
| ; SI-NEXT: .LBB29_2: ; %cmp.true |
| ; SI-NEXT: v_add_f32_e64 v19, s5, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v18, s4, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v17, s7, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v16, s6, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[20:21], v[18:19], 16 |
| ; SI-NEXT: v_add_f32_e64 v15, s9, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v14, s8, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[21:22], v[16:17], 16 |
| ; SI-NEXT: v_add_f32_e64 v13, s29, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v12, s28, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[22:23], v[14:15], 16 |
| ; SI-NEXT: v_add_f32_e64 v11, s27, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v10, s26, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[23:24], v[12:13], 16 |
| ; SI-NEXT: v_add_f32_e64 v9, s25, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v8, s24, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[24:25], v[10:11], 16 |
| ; SI-NEXT: v_add_f32_e64 v7, s23, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v6, s22, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[25:26], v[8:9], 16 |
| ; SI-NEXT: v_add_f32_e64 v5, s21, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v4, s20, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[26:27], v[6:7], 16 |
| ; SI-NEXT: v_add_f32_e64 v3, s19, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v2, s18, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[27:28], v[4:5], 16 |
| ; SI-NEXT: v_add_f32_e64 v1, s17, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v0, s16, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[28:29], v[2:3], 16 |
| ; SI-NEXT: v_lshr_b64 v[29:30], v[0:1], 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v1 |
| ; SI-NEXT: s_branch .LBB29_5 |
| ; SI-NEXT: .LBB29_3: |
| ; SI-NEXT: ; implicit-def: $sgpr60 |
| ; SI-NEXT: ; implicit-def: $sgpr72 |
| ; SI-NEXT: ; implicit-def: $sgpr58 |
| ; SI-NEXT: ; implicit-def: $sgpr73 |
| ; SI-NEXT: ; implicit-def: $sgpr56 |
| ; SI-NEXT: ; implicit-def: $sgpr74 |
| ; SI-NEXT: ; implicit-def: $sgpr46 |
| ; SI-NEXT: ; implicit-def: $sgpr75 |
| ; SI-NEXT: ; implicit-def: $sgpr44 |
| ; SI-NEXT: ; implicit-def: $sgpr76 |
| ; SI-NEXT: ; implicit-def: $sgpr42 |
| ; SI-NEXT: ; implicit-def: $sgpr77 |
| ; SI-NEXT: ; implicit-def: $sgpr40 |
| ; SI-NEXT: ; implicit-def: $sgpr78 |
| ; SI-NEXT: ; implicit-def: $sgpr14 |
| ; SI-NEXT: ; implicit-def: $sgpr79 |
| ; SI-NEXT: ; implicit-def: $sgpr12 |
| ; SI-NEXT: ; implicit-def: $sgpr88 |
| ; SI-NEXT: ; implicit-def: $sgpr10 |
| ; SI-NEXT: ; implicit-def: $sgpr89 |
| ; SI-NEXT: s_branch .LBB29_2 |
| ; SI-NEXT: .LBB29_4: |
| ; SI-NEXT: v_mov_b32_e32 v0, s16 |
| ; SI-NEXT: v_mov_b32_e32 v1, s17 |
| ; SI-NEXT: v_mov_b32_e32 v2, s18 |
| ; SI-NEXT: v_mov_b32_e32 v3, s19 |
| ; SI-NEXT: v_mov_b32_e32 v4, s20 |
| ; SI-NEXT: v_mov_b32_e32 v5, s21 |
| ; SI-NEXT: v_mov_b32_e32 v6, s22 |
| ; SI-NEXT: v_mov_b32_e32 v7, s23 |
| ; SI-NEXT: v_mov_b32_e32 v8, s24 |
| ; SI-NEXT: v_mov_b32_e32 v9, s25 |
| ; SI-NEXT: v_mov_b32_e32 v10, s26 |
| ; SI-NEXT: v_mov_b32_e32 v11, s27 |
| ; SI-NEXT: v_mov_b32_e32 v12, s28 |
| ; SI-NEXT: v_mov_b32_e32 v13, s29 |
| ; SI-NEXT: v_mov_b32_e32 v14, s8 |
| ; SI-NEXT: v_mov_b32_e32 v15, s9 |
| ; SI-NEXT: v_mov_b32_e32 v16, s6 |
| ; SI-NEXT: v_mov_b32_e32 v17, s7 |
| ; SI-NEXT: v_mov_b32_e32 v18, s4 |
| ; SI-NEXT: v_mov_b32_e32 v19, s5 |
| ; SI-NEXT: v_mov_b32_e32 v48, s72 |
| ; SI-NEXT: v_mov_b32_e32 v39, s73 |
| ; SI-NEXT: v_mov_b32_e32 v38, s74 |
| ; SI-NEXT: v_mov_b32_e32 v37, s75 |
| ; SI-NEXT: v_mov_b32_e32 v36, s76 |
| ; SI-NEXT: v_mov_b32_e32 v35, s77 |
| ; SI-NEXT: v_mov_b32_e32 v34, s78 |
| ; SI-NEXT: v_mov_b32_e32 v33, s79 |
| ; SI-NEXT: v_mov_b32_e32 v32, s88 |
| ; SI-NEXT: v_mov_b32_e32 v31, s89 |
| ; SI-NEXT: v_mov_b32_e32 v20, s10 |
| ; SI-NEXT: v_mov_b32_e32 v21, s12 |
| ; SI-NEXT: v_mov_b32_e32 v22, s14 |
| ; SI-NEXT: v_mov_b32_e32 v23, s40 |
| ; SI-NEXT: v_mov_b32_e32 v24, s42 |
| ; SI-NEXT: v_mov_b32_e32 v25, s44 |
| ; SI-NEXT: v_mov_b32_e32 v26, s46 |
| ; SI-NEXT: v_mov_b32_e32 v27, s56 |
| ; SI-NEXT: v_mov_b32_e32 v28, s58 |
| ; SI-NEXT: v_mov_b32_e32 v29, s60 |
| ; SI-NEXT: .LBB29_5: ; %end |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v29 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v48 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v28 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v27 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v25 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v35 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v23 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v34 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v22 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v32 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v31 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v29 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v28 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v27 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v26 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v25 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v24 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20f32_to_v40i16_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB29_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_lshr_b32 s12, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s10, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s46, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: s_cbranch_execnz .LBB29_4 |
| ; VI-NEXT: .LBB29_2: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e64 v19, s6, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v18, s7, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v17, s8, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v16, s9, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v15, s10, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v14, s11, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v13, s29, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v12, s28, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v11, s27, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v10, s26, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v9, s25, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v8, s24, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v7, s23, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v6, s22, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v5, s21, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v4, s20, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v3, s19, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v2, s18, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v1, s17, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v0, s16, 1.0 |
| ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: s_branch .LBB29_5 |
| ; VI-NEXT: .LBB29_3: |
| ; VI-NEXT: ; implicit-def: $sgpr63 |
| ; VI-NEXT: ; implicit-def: $sgpr62 |
| ; VI-NEXT: ; implicit-def: $sgpr61 |
| ; VI-NEXT: ; implicit-def: $sgpr60 |
| ; VI-NEXT: ; implicit-def: $sgpr59 |
| ; VI-NEXT: ; implicit-def: $sgpr58 |
| ; VI-NEXT: ; implicit-def: $sgpr57 |
| ; VI-NEXT: ; implicit-def: $sgpr56 |
| ; VI-NEXT: ; implicit-def: $sgpr47 |
| ; VI-NEXT: ; implicit-def: $sgpr46 |
| ; VI-NEXT: ; implicit-def: $sgpr45 |
| ; VI-NEXT: ; implicit-def: $sgpr44 |
| ; VI-NEXT: ; implicit-def: $sgpr43 |
| ; VI-NEXT: ; implicit-def: $sgpr42 |
| ; VI-NEXT: ; implicit-def: $sgpr41 |
| ; VI-NEXT: ; implicit-def: $sgpr40 |
| ; VI-NEXT: ; implicit-def: $sgpr15 |
| ; VI-NEXT: ; implicit-def: $sgpr14 |
| ; VI-NEXT: ; implicit-def: $sgpr13 |
| ; VI-NEXT: ; implicit-def: $sgpr12 |
| ; VI-NEXT: s_branch .LBB29_2 |
| ; VI-NEXT: .LBB29_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s16 |
| ; VI-NEXT: v_mov_b32_e32 v1, s17 |
| ; VI-NEXT: v_mov_b32_e32 v2, s18 |
| ; VI-NEXT: v_mov_b32_e32 v3, s19 |
| ; VI-NEXT: v_mov_b32_e32 v4, s20 |
| ; VI-NEXT: v_mov_b32_e32 v5, s21 |
| ; VI-NEXT: v_mov_b32_e32 v6, s22 |
| ; VI-NEXT: v_mov_b32_e32 v7, s23 |
| ; VI-NEXT: v_mov_b32_e32 v8, s24 |
| ; VI-NEXT: v_mov_b32_e32 v9, s25 |
| ; VI-NEXT: v_mov_b32_e32 v10, s26 |
| ; VI-NEXT: v_mov_b32_e32 v11, s27 |
| ; VI-NEXT: v_mov_b32_e32 v12, s28 |
| ; VI-NEXT: v_mov_b32_e32 v13, s29 |
| ; VI-NEXT: v_mov_b32_e32 v14, s11 |
| ; VI-NEXT: v_mov_b32_e32 v15, s10 |
| ; VI-NEXT: v_mov_b32_e32 v16, s9 |
| ; VI-NEXT: v_mov_b32_e32 v17, s8 |
| ; VI-NEXT: v_mov_b32_e32 v18, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: v_mov_b32_e32 v39, s63 |
| ; VI-NEXT: v_mov_b32_e32 v38, s62 |
| ; VI-NEXT: v_mov_b32_e32 v37, s61 |
| ; VI-NEXT: v_mov_b32_e32 v36, s60 |
| ; VI-NEXT: v_mov_b32_e32 v35, s59 |
| ; VI-NEXT: v_mov_b32_e32 v34, s58 |
| ; VI-NEXT: v_mov_b32_e32 v33, s57 |
| ; VI-NEXT: v_mov_b32_e32 v32, s56 |
| ; VI-NEXT: v_mov_b32_e32 v31, s47 |
| ; VI-NEXT: v_mov_b32_e32 v30, s46 |
| ; VI-NEXT: v_mov_b32_e32 v29, s45 |
| ; VI-NEXT: v_mov_b32_e32 v28, s44 |
| ; VI-NEXT: v_mov_b32_e32 v27, s43 |
| ; VI-NEXT: v_mov_b32_e32 v26, s42 |
| ; VI-NEXT: v_mov_b32_e32 v25, s41 |
| ; VI-NEXT: v_mov_b32_e32 v24, s40 |
| ; VI-NEXT: v_mov_b32_e32 v23, s15 |
| ; VI-NEXT: v_mov_b32_e32 v22, s14 |
| ; VI-NEXT: v_mov_b32_e32 v21, s13 |
| ; VI-NEXT: v_mov_b32_e32 v20, s12 |
| ; VI-NEXT: .LBB29_5: ; %end |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v20f32_to_v40i16_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s10, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s11, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB29_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_lshr_b32 s12, s6, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s7, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s8, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s9, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s10, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s11, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s43, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s58, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s60, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s62, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s16, 16 |
| ; GFX9-NEXT: s_cbranch_execnz .LBB29_4 |
| ; GFX9-NEXT: .LBB29_2: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e64 v19, s6, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v18, s7, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v17, s8, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v16, s9, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v15, s10, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v14, s11, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v13, s29, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v12, s28, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v11, s27, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v10, s26, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v9, s25, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v8, s24, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v7, s23, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v6, s22, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v5, s21, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v4, s20, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v3, s19, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v2, s18, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v1, s17, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v0, s16, 1.0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: s_branch .LBB29_5 |
| ; GFX9-NEXT: .LBB29_3: |
| ; GFX9-NEXT: ; implicit-def: $sgpr63 |
| ; GFX9-NEXT: ; implicit-def: $sgpr62 |
| ; GFX9-NEXT: ; implicit-def: $sgpr61 |
| ; GFX9-NEXT: ; implicit-def: $sgpr60 |
| ; GFX9-NEXT: ; implicit-def: $sgpr59 |
| ; GFX9-NEXT: ; implicit-def: $sgpr58 |
| ; GFX9-NEXT: ; implicit-def: $sgpr57 |
| ; GFX9-NEXT: ; implicit-def: $sgpr56 |
| ; GFX9-NEXT: ; implicit-def: $sgpr47 |
| ; GFX9-NEXT: ; implicit-def: $sgpr46 |
| ; GFX9-NEXT: ; implicit-def: $sgpr45 |
| ; GFX9-NEXT: ; implicit-def: $sgpr44 |
| ; GFX9-NEXT: ; implicit-def: $sgpr43 |
| ; GFX9-NEXT: ; implicit-def: $sgpr42 |
| ; GFX9-NEXT: ; implicit-def: $sgpr41 |
| ; GFX9-NEXT: ; implicit-def: $sgpr40 |
| ; GFX9-NEXT: ; implicit-def: $sgpr15 |
| ; GFX9-NEXT: ; implicit-def: $sgpr14 |
| ; GFX9-NEXT: ; implicit-def: $sgpr13 |
| ; GFX9-NEXT: ; implicit-def: $sgpr12 |
| ; GFX9-NEXT: s_branch .LBB29_2 |
| ; GFX9-NEXT: .LBB29_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s28 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s29 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s6 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v33, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v32, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s15 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s14 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s13 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s12 |
| ; GFX9-NEXT: .LBB29_5: ; %end |
| ; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX9-NEXT: v_lshl_or_b32 v0, v39, 16, v0 |
| ; GFX9-NEXT: v_lshl_or_b32 v1, v38, 16, v1 |
| ; GFX9-NEXT: v_lshl_or_b32 v2, v37, 16, v2 |
| ; GFX9-NEXT: v_lshl_or_b32 v3, v36, 16, v3 |
| ; GFX9-NEXT: v_lshl_or_b32 v4, v35, 16, v4 |
| ; GFX9-NEXT: v_lshl_or_b32 v5, v34, 16, v5 |
| ; GFX9-NEXT: v_lshl_or_b32 v6, v33, 16, v6 |
| ; GFX9-NEXT: v_lshl_or_b32 v7, v32, 16, v7 |
| ; GFX9-NEXT: v_lshl_or_b32 v8, v31, 16, v8 |
| ; GFX9-NEXT: v_lshl_or_b32 v9, v30, 16, v9 |
| ; GFX9-NEXT: v_lshl_or_b32 v10, v29, 16, v10 |
| ; GFX9-NEXT: v_lshl_or_b32 v11, v28, 16, v11 |
| ; GFX9-NEXT: v_lshl_or_b32 v12, v27, 16, v12 |
| ; GFX9-NEXT: v_lshl_or_b32 v13, v26, 16, v13 |
| ; GFX9-NEXT: v_lshl_or_b32 v14, v25, 16, v14 |
| ; GFX9-NEXT: v_lshl_or_b32 v15, v24, 16, v15 |
| ; GFX9-NEXT: v_lshl_or_b32 v16, v23, 16, v16 |
| ; GFX9-NEXT: v_lshl_or_b32 v17, v22, 16, v17 |
| ; GFX9-NEXT: v_lshl_or_b32 v18, v21, 16, v18 |
| ; GFX9-NEXT: v_lshl_or_b32 v19, v20, 16, v19 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v20f32_to_v40i16_scalar: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-TRUE16-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB29_3 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s4, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s5, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s29, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s28, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s27, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s26, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s24, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s23, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s22, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s21, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s20, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s19, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s18, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s17, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s16, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s47, s3, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s56, s2, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s57, s1, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s58, s0, 16 |
| ; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB29_4 |
| ; GFX11-TRUE16-NEXT: .LBB29_2: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v19, s4, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v18, s5, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v17, s29, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v16, s28, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v15, s27, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v14, s26, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v13, s25, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v12, s24, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v11, s23, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v10, s22, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v9, s21, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v8, s20, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v7, s19, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v6, s18, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v5, s17, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v4, s16, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v3, s3, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v2, s2, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v1, s1, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v0, s0, 1.0 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-TRUE16-NEXT: s_branch .LBB29_5 |
| ; GFX11-TRUE16-NEXT: .LBB29_3: |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr58 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr57 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr56 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr47 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr46 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr45 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr44 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr43 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr41 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr40 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr15 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr14 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr13 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr12 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr11 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr10 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr9 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr8 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr7 |
| ; GFX11-TRUE16-NEXT: s_branch .LBB29_2 |
| ; GFX11-TRUE16-NEXT: .LBB29_4: |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v39, s58 :: v_dual_mov_b32 v38, s57 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v37, s56 :: v_dual_mov_b32 v36, s47 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v35, s46 :: v_dual_mov_b32 v34, s45 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v33, s44 :: v_dual_mov_b32 v32, s43 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, s42 :: v_dual_mov_b32 v30, s41 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, s40 :: v_dual_mov_b32 v28, s15 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, s14 :: v_dual_mov_b32 v26, s13 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, s12 :: v_dual_mov_b32 v24, s11 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, s10 :: v_dual_mov_b32 v22, s9 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, s8 :: v_dual_mov_b32 v20, s7 |
| ; GFX11-TRUE16-NEXT: .LBB29_5: ; %end |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v39.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v38.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v37.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v36.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v35.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v34.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v33.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v32.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v31.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v30.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v29.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v28.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v27.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v26.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v25.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v24.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v23.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v22.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v21.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v20.l |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v20f32_to_v40i16_scalar: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-FAKE16-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-FAKE16-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB29_3 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s4, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s5, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s29, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s28, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s27, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s26, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s25, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s24, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s23, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s40, s22, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s21, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s20, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s19, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s18, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s17, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s16, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s47, s3, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s56, s2, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s57, s1, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s58, s0, 16 |
| ; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB29_4 |
| ; GFX11-FAKE16-NEXT: .LBB29_2: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v15, s4, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v16, s5, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v17, s29, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v18, s28, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v19, s27, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v10, s26, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v11, s25, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v12, s24, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v13, s23, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v14, s22, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v5, s21, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v6, s20, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v7, s19, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v8, s18, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v9, s17, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v0, s16, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v1, s3, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v2, s2, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v3, s1, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v4, s0, 1.0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v4 |
| ; GFX11-FAKE16-NEXT: s_branch .LBB29_5 |
| ; GFX11-FAKE16-NEXT: .LBB29_3: |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr58 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr57 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr56 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr47 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr46 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr45 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr44 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr43 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr41 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr40 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr15 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr14 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr13 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr12 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr11 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr10 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr9 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr8 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr7 |
| ; GFX11-FAKE16-NEXT: s_branch .LBB29_2 |
| ; GFX11-FAKE16-NEXT: .LBB29_4: |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v3, s1 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v1, s3 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s16 :: v_dual_mov_b32 v9, s17 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v6, s20 :: v_dual_mov_b32 v5, s21 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s22 :: v_dual_mov_b32 v13, s23 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v11, s25 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s26 :: v_dual_mov_b32 v19, s27 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v18, s28 :: v_dual_mov_b32 v17, s29 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v15, s4 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v39, s58 :: v_dual_mov_b32 v38, s57 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, s56 :: v_dual_mov_b32 v36, s47 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v35, s46 :: v_dual_mov_b32 v34, s45 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v33, s44 :: v_dual_mov_b32 v32, s43 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v31, s42 :: v_dual_mov_b32 v30, s41 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v29, s40 :: v_dual_mov_b32 v28, s15 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v27, s14 :: v_dual_mov_b32 v26, s13 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v25, s12 :: v_dual_mov_b32 v24, s11 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v23, s10 :: v_dual_mov_b32 v22, s9 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v21, s8 :: v_dual_mov_b32 v20, s7 |
| ; GFX11-FAKE16-NEXT: .LBB29_5: ; %end |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v1 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v0 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v39, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v38, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v36, 16, v48 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v35, 16, v49 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v6 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xffff, v5 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v34, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v33, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v31, 16, v35 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v30, 16, v36 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v30, 0xffff, v11 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xffff, v10 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v29, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v28, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v26, 16, v30 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v25, 16, v31 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v16 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v15 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v24, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v23, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v21, 16, v25 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v20, 16, v26 |
| ; GFX11-FAKE16-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 <20 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <20 x float> %a1 to <40 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x float> %a to <40 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <40 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x i16> %phi |
| } |
| |
| define <20 x float> @bitcast_v40i16_to_v20f32(<40 x i16> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v40i16_to_v20f32: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: v_mov_b32_e32 v32, v19 |
| ; SI-NEXT: v_mov_b32_e32 v33, v18 |
| ; SI-NEXT: v_mov_b32_e32 v34, v17 |
| ; SI-NEXT: v_mov_b32_e32 v35, v16 |
| ; SI-NEXT: v_mov_b32_e32 v36, v15 |
| ; SI-NEXT: v_mov_b32_e32 v37, v14 |
| ; SI-NEXT: v_mov_b32_e32 v38, v13 |
| ; SI-NEXT: v_mov_b32_e32 v39, v12 |
| ; SI-NEXT: v_mov_b32_e32 v48, v11 |
| ; SI-NEXT: v_mov_b32_e32 v49, v10 |
| ; SI-NEXT: v_mov_b32_e32 v50, v9 |
| ; SI-NEXT: v_mov_b32_e32 v51, v8 |
| ; SI-NEXT: v_mov_b32_e32 v52, v7 |
| ; SI-NEXT: v_mov_b32_e32 v53, v6 |
| ; SI-NEXT: v_mov_b32_e32 v54, v5 |
| ; SI-NEXT: v_mov_b32_e32 v55, v4 |
| ; SI-NEXT: v_mov_b32_e32 v40, v3 |
| ; SI-NEXT: v_mov_b32_e32 v41, v2 |
| ; SI-NEXT: v_mov_b32_e32 v42, v1 |
| ; SI-NEXT: v_mov_b32_e32 v43, v0 |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v33 |
| ; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v34 |
| ; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v35 |
| ; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v36 |
| ; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v37 |
| ; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v38 |
| ; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v39 |
| ; SI-NEXT: v_lshrrev_b32_e32 v8, 16, v48 |
| ; SI-NEXT: v_lshrrev_b32_e32 v9, 16, v49 |
| ; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v50 |
| ; SI-NEXT: v_lshrrev_b32_e32 v11, 16, v51 |
| ; SI-NEXT: v_lshrrev_b32_e32 v12, 16, v52 |
| ; SI-NEXT: v_lshrrev_b32_e32 v13, 16, v53 |
| ; SI-NEXT: v_lshrrev_b32_e32 v14, 16, v54 |
| ; SI-NEXT: v_lshrrev_b32_e32 v15, 16, v55 |
| ; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v40 |
| ; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v41 |
| ; SI-NEXT: v_lshrrev_b32_e32 v18, 16, v42 |
| ; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v43 |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_waitcnt expcnt(4) |
| ; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v56, 16, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v47, 16, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v12 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB30_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v39 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v43 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v42 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v41 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v40 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v55 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v54 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v53 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v52 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v51 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v50 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v49 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v48 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v59 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v58 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v57 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v56 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v47 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v46 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v45 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v44 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v63 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v62 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v61 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v60 |
| ; SI-NEXT: ; implicit-def: $vgpr43 |
| ; SI-NEXT: ; implicit-def: $vgpr42 |
| ; SI-NEXT: ; implicit-def: $vgpr41 |
| ; SI-NEXT: ; implicit-def: $vgpr40 |
| ; SI-NEXT: ; implicit-def: $vgpr55 |
| ; SI-NEXT: ; implicit-def: $vgpr54 |
| ; SI-NEXT: ; implicit-def: $vgpr53 |
| ; SI-NEXT: ; implicit-def: $vgpr52 |
| ; SI-NEXT: ; implicit-def: $vgpr51 |
| ; SI-NEXT: ; implicit-def: $vgpr50 |
| ; SI-NEXT: ; implicit-def: $vgpr49 |
| ; SI-NEXT: ; implicit-def: $vgpr48 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr59 |
| ; SI-NEXT: ; implicit-def: $vgpr58 |
| ; SI-NEXT: ; implicit-def: $vgpr57 |
| ; SI-NEXT: ; implicit-def: $vgpr56 |
| ; SI-NEXT: ; implicit-def: $vgpr47 |
| ; SI-NEXT: ; implicit-def: $vgpr46 |
| ; SI-NEXT: ; implicit-def: $vgpr45 |
| ; SI-NEXT: ; implicit-def: $vgpr44 |
| ; SI-NEXT: ; implicit-def: $vgpr63 |
| ; SI-NEXT: ; implicit-def: $vgpr62 |
| ; SI-NEXT: ; implicit-def: $vgpr61 |
| ; SI-NEXT: ; implicit-def: $vgpr60 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v38 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v37 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v36 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v35 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v34 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v33 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v19 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v32 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: .LBB30_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB30_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v39 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v43 |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v42 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v41 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v40 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v55 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v54 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v53 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v52 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v51 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v50 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v49 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v48 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_or_b32_e32 v0, v59, v0 |
| ; SI-NEXT: s_mov_b32 s6, 0x30000 |
| ; SI-NEXT: v_or_b32_e32 v1, v58, v1 |
| ; SI-NEXT: v_or_b32_e32 v2, v57, v2 |
| ; SI-NEXT: v_or_b32_e32 v3, v56, v3 |
| ; SI-NEXT: v_or_b32_e32 v4, v47, v4 |
| ; SI-NEXT: v_or_b32_e32 v5, v46, v5 |
| ; SI-NEXT: v_or_b32_e32 v6, v45, v6 |
| ; SI-NEXT: v_or_b32_e32 v7, v44, v7 |
| ; SI-NEXT: v_or_b32_e32 v8, v63, v8 |
| ; SI-NEXT: v_or_b32_e32 v9, v62, v9 |
| ; SI-NEXT: v_or_b32_e32 v10, v61, v10 |
| ; SI-NEXT: v_or_b32_e32 v11, v60, v11 |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0 |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, s6, v3 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, s6, v5 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, s6, v7 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v38 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_or_b32_e32 v13, v14, v13 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v37 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_or_b32_e32 v14, v15, v14 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v36 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v35 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v34 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_or_b32_e32 v17, v18, v17 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v33 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_or_b32_e32 v18, v19, v18 |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v32 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_or_b32_e32 v19, v20, v19 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, s6, v11 |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, s6, v13 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v14 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, s6, v15 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, s6, v16 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, s6, v17 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v18 |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 0x30000, v19 |
| ; SI-NEXT: .LBB30_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40i16_to_v20f32: |
| ; 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:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: v_mov_b32_e32 v32, v19 |
| ; VI-NEXT: v_mov_b32_e32 v33, v18 |
| ; VI-NEXT: v_mov_b32_e32 v34, v17 |
| ; VI-NEXT: v_mov_b32_e32 v35, v16 |
| ; VI-NEXT: v_mov_b32_e32 v36, v15 |
| ; VI-NEXT: v_mov_b32_e32 v37, v14 |
| ; VI-NEXT: v_mov_b32_e32 v38, v13 |
| ; VI-NEXT: v_mov_b32_e32 v39, v12 |
| ; VI-NEXT: v_mov_b32_e32 v48, v11 |
| ; VI-NEXT: v_mov_b32_e32 v49, v10 |
| ; VI-NEXT: v_mov_b32_e32 v50, v9 |
| ; VI-NEXT: v_mov_b32_e32 v51, v8 |
| ; VI-NEXT: v_mov_b32_e32 v52, v7 |
| ; VI-NEXT: v_mov_b32_e32 v53, v6 |
| ; VI-NEXT: v_mov_b32_e32 v54, v5 |
| ; VI-NEXT: v_mov_b32_e32 v55, v4 |
| ; VI-NEXT: v_mov_b32_e32 v40, v3 |
| ; VI-NEXT: v_mov_b32_e32 v41, v2 |
| ; VI-NEXT: v_mov_b32_e32 v42, v1 |
| ; VI-NEXT: v_mov_b32_e32 v43, v0 |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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 .LBB30_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_mov_b32_e32 v19, 16 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v0, v19, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v1, v19, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v2, v19, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v3, v19, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v4, v19, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v5, v19, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v6, v19, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v7, v19, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v8, v19, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v9, v19, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v10, v19, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v11, v19, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v12, v19, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v13, v19, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v14, v19, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v15, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v16, v19, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v17, v19, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v18, v19, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v19, v19, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v40, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v55, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v54, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v53, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v52, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v51, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v50, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v49, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v48, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v39, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v38, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v37, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v36, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v35, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v34, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v33, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v32, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; 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: .LBB30_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB30_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v19, 3 |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v43 |
| ; VI-NEXT: v_add_u16_sdwa v1, v43, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v42 |
| ; VI-NEXT: v_add_u16_sdwa v2, v42, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v1, v1, v2 |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v41 |
| ; VI-NEXT: v_add_u16_sdwa v3, v41, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v40 |
| ; VI-NEXT: v_add_u16_sdwa v4, v40, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v55 |
| ; VI-NEXT: v_add_u16_sdwa v5, v55, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v54 |
| ; VI-NEXT: v_add_u16_sdwa v6, v54, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v53 |
| ; VI-NEXT: v_add_u16_sdwa v7, v53, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v52 |
| ; VI-NEXT: v_add_u16_sdwa v8, v52, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v51 |
| ; VI-NEXT: v_add_u16_sdwa v9, v51, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v50 |
| ; VI-NEXT: v_add_u16_sdwa v10, v50, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v49 |
| ; VI-NEXT: v_add_u16_sdwa v11, v49, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v48 |
| ; VI-NEXT: v_add_u16_sdwa v12, v48, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v39 |
| ; VI-NEXT: v_add_u16_sdwa v13, v39, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v38 |
| ; VI-NEXT: v_add_u16_sdwa v14, v38, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v37 |
| ; VI-NEXT: v_add_u16_sdwa v15, v37, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; VI-NEXT: v_add_u16_e32 v15, 3, v36 |
| ; VI-NEXT: v_add_u16_sdwa v16, v36, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v35 |
| ; VI-NEXT: v_add_u16_sdwa v17, v35, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; VI-NEXT: v_add_u16_e32 v17, 3, v34 |
| ; VI-NEXT: v_add_u16_sdwa v18, v34, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; VI-NEXT: v_add_u16_e32 v18, 3, v33 |
| ; VI-NEXT: v_add_u16_sdwa v20, v33, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; VI-NEXT: v_add_u16_e32 v20, 3, v32 |
| ; VI-NEXT: v_add_u16_sdwa v19, v32, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v19, v20, v19 |
| ; VI-NEXT: .LBB30_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40i16_to_v20f32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_mov_b32_e32 v32, v19 |
| ; 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: v_mov_b32_e32 v33, v18 |
| ; GFX9-NEXT: v_mov_b32_e32 v43, v0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, v17 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, v16 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, v15 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, v14 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, v13 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, v12 |
| ; GFX9-NEXT: v_mov_b32_e32 v48, v11 |
| ; GFX9-NEXT: v_mov_b32_e32 v49, v10 |
| ; GFX9-NEXT: v_mov_b32_e32 v50, v9 |
| ; GFX9-NEXT: v_mov_b32_e32 v51, v8 |
| ; GFX9-NEXT: v_mov_b32_e32 v52, v7 |
| ; GFX9-NEXT: v_mov_b32_e32 v53, v6 |
| ; GFX9-NEXT: v_mov_b32_e32 v54, v5 |
| ; GFX9-NEXT: v_mov_b32_e32 v55, v4 |
| ; GFX9-NEXT: v_mov_b32_e32 v40, v3 |
| ; GFX9-NEXT: v_mov_b32_e32 v41, v2 |
| ; GFX9-NEXT: v_mov_b32_e32 v42, v1 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v48 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v49 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v50 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v51 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v52 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v53 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v54 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v55 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43 |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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 .LBB30_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v35 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v34 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v33 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v32 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; 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: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: .LBB30_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB30_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_pk_add_u16 v0, v0, 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 v2, v2, 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 v4, v4, 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 v6, v6, 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 v8, v8, 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 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; GFX9-NEXT: v_pk_add_u16 v12, v12, 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 v14, v14, 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 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB30_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; 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-TRUE16-LABEL: bitcast_v40i16_to_v20f32: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB30_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: .LBB30_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v40i16_to_v20f32: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v37, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v39, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v48, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v36, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v35, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v34, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v33, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v32, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v31, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v30, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v29, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v28, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v27, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v26, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v25, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v24, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v23, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v22, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v21, v19, 0x5040100 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB30_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: .LBB30_2: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-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 <40 x i16> %a, splat (i16 3) |
| %a2 = bitcast <40 x i16> %a1 to <20 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x i16> %a to <20 x float> |
| br label %end |
| |
| end: |
| %phi = phi <20 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x float> %phi |
| } |
| |
| define inreg <20 x float> @bitcast_v40i16_to_v20f32_scalar(<40 x i16> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v40i16_to_v20f32_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v20, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v20, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v20, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v20, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v20, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v20, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v20, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v20, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v20, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v20, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v20, s54, 10 |
| ; SI-NEXT: v_writelane_b32 v20, s55, 11 |
| ; SI-NEXT: v_writelane_b32 v20, s64, 12 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s12, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s15, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s74, v1 |
| ; SI-NEXT: v_readfirstlane_b32 s77, v0 |
| ; SI-NEXT: v_writelane_b32 v20, s65, 13 |
| ; SI-NEXT: s_lshr_b32 s11, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s13, s28, 16 |
| ; SI-NEXT: s_lshr_b32 s72, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s26, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; SI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; SI-NEXT: s_lshr_b32 s92, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s93, s18, 16 |
| ; SI-NEXT: s_lshr_b32 s94, s17, 16 |
| ; SI-NEXT: s_lshr_b32 s95, s16, 16 |
| ; SI-NEXT: s_lshr_b32 s6, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s8, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s10, s12, 16 |
| ; SI-NEXT: s_lshr_b32 s14, s15, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s74, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s77, 16 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v20, s66, 14 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_writelane_b32 v20, s67, 15 |
| ; SI-NEXT: s_cbranch_scc0 .LBB31_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s95, 16 |
| ; SI-NEXT: s_or_b32 s36, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s94, 16 |
| ; SI-NEXT: s_or_b32 s37, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s93, 16 |
| ; SI-NEXT: s_or_b32 s38, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s92, 16 |
| ; SI-NEXT: s_or_b32 s39, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; SI-NEXT: s_or_b32 s40, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; SI-NEXT: s_or_b32 s41, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; SI-NEXT: s_or_b32 s42, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; SI-NEXT: s_or_b32 s43, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; SI-NEXT: s_or_b32 s44, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; SI-NEXT: s_or_b32 s45, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; SI-NEXT: s_or_b32 s46, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; SI-NEXT: s_or_b32 s47, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; SI-NEXT: s_or_b32 s48, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; SI-NEXT: s_or_b32 s49, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s77, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; SI-NEXT: s_or_b32 s50, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s74, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; SI-NEXT: s_or_b32 s51, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s15, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; SI-NEXT: s_or_b32 s52, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; SI-NEXT: s_or_b32 s53, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; SI-NEXT: s_or_b32 s54, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; SI-NEXT: s_or_b32 s55, s4, s5 |
| ; SI-NEXT: s_cbranch_execnz .LBB31_3 |
| ; SI-NEXT: .LBB31_2: ; %cmp.true |
| ; SI-NEXT: s_add_i32 s16, s16, 3 |
| ; SI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s95, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s17, s17, 3 |
| ; SI-NEXT: s_add_i32 s36, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s94, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s18, s18, 3 |
| ; SI-NEXT: s_add_i32 s37, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s93, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s19, s19, 3 |
| ; SI-NEXT: s_add_i32 s38, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s92, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s20, s20, 3 |
| ; SI-NEXT: s_add_i32 s39, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s21, s21, 3 |
| ; SI-NEXT: s_add_i32 s40, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s22, s22, 3 |
| ; SI-NEXT: s_add_i32 s41, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s23, s23, 3 |
| ; SI-NEXT: s_add_i32 s42, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s24, s24, 3 |
| ; SI-NEXT: s_add_i32 s43, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s25, s25, 3 |
| ; SI-NEXT: s_add_i32 s44, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s26, s26, 3 |
| ; SI-NEXT: s_add_i32 s45, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s27, s27, 3 |
| ; SI-NEXT: s_add_i32 s46, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s28, s28, 3 |
| ; SI-NEXT: s_add_i32 s47, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s29, s29, 3 |
| ; SI-NEXT: s_add_i32 s48, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s77, s77, 3 |
| ; SI-NEXT: s_add_i32 s49, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s77, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s74, s74, 3 |
| ; SI-NEXT: s_add_i32 s50, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s74, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s15, s15, 3 |
| ; SI-NEXT: s_add_i32 s51, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s15, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s12, s12, 3 |
| ; SI-NEXT: s_add_i32 s52, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s9, s9, 3 |
| ; SI-NEXT: s_add_i32 s53, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s7, s7, 3 |
| ; SI-NEXT: s_add_i32 s54, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s55, s4, 0x30000 |
| ; SI-NEXT: .LBB31_3: ; %end |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_readlane_b32 s67, v20, 15 |
| ; SI-NEXT: v_readlane_b32 s66, v20, 14 |
| ; SI-NEXT: v_readlane_b32 s65, v20, 13 |
| ; SI-NEXT: v_readlane_b32 s64, v20, 12 |
| ; SI-NEXT: v_readlane_b32 s55, v20, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v20, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v20, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v20, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v20, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v20, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v20, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v20, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v20, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v20, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v20, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v20, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB31_4: |
| ; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; SI-NEXT: s_branch .LBB31_2 |
| ; |
| ; VI-LABEL: bitcast_v40i16_to_v20f32_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v20, s30, 0 |
| ; VI-NEXT: v_writelane_b32 v20, s31, 1 |
| ; VI-NEXT: v_writelane_b32 v20, s34, 2 |
| ; VI-NEXT: v_writelane_b32 v20, s35, 3 |
| ; VI-NEXT: v_writelane_b32 v20, s36, 4 |
| ; VI-NEXT: v_writelane_b32 v20, s37, 5 |
| ; VI-NEXT: v_writelane_b32 v20, s38, 6 |
| ; VI-NEXT: v_writelane_b32 v20, s39, 7 |
| ; VI-NEXT: v_writelane_b32 v20, s48, 8 |
| ; VI-NEXT: v_writelane_b32 v20, s49, 9 |
| ; VI-NEXT: v_writelane_b32 v20, s50, 10 |
| ; VI-NEXT: v_writelane_b32 v20, s51, 11 |
| ; VI-NEXT: v_writelane_b32 v20, s52, 12 |
| ; VI-NEXT: v_writelane_b32 v20, s53, 13 |
| ; VI-NEXT: v_writelane_b32 v20, s54, 14 |
| ; VI-NEXT: v_writelane_b32 v20, s55, 15 |
| ; VI-NEXT: v_writelane_b32 v20, s64, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s12, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s15, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s74, v1 |
| ; VI-NEXT: v_readfirstlane_b32 s77, v0 |
| ; VI-NEXT: v_writelane_b32 v20, s65, 17 |
| ; VI-NEXT: s_lshr_b32 s11, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s72, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s75, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s78, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s30, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s31, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s34, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s35, s16, 16 |
| ; VI-NEXT: s_lshr_b32 s6, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s8, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s10, s12, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s15, 16 |
| ; VI-NEXT: s_lshr_b32 s73, s74, 16 |
| ; VI-NEXT: s_lshr_b32 s76, s77, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v20, s66, 18 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_writelane_b32 v20, s67, 19 |
| ; VI-NEXT: s_cbranch_scc0 .LBB31_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s35, 16 |
| ; VI-NEXT: s_or_b32 s36, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s5, s34, 16 |
| ; VI-NEXT: s_or_b32 s37, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s5, s31, 16 |
| ; VI-NEXT: s_or_b32 s38, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s5, s30, 16 |
| ; VI-NEXT: s_or_b32 s39, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; VI-NEXT: s_or_b32 s40, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; VI-NEXT: s_or_b32 s41, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; VI-NEXT: s_or_b32 s42, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; VI-NEXT: s_or_b32 s43, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; VI-NEXT: s_or_b32 s44, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; VI-NEXT: s_or_b32 s45, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; VI-NEXT: s_or_b32 s46, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; VI-NEXT: s_or_b32 s47, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; VI-NEXT: s_or_b32 s48, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; VI-NEXT: s_or_b32 s49, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s77 |
| ; VI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; VI-NEXT: s_or_b32 s50, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s74 |
| ; VI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; VI-NEXT: s_or_b32 s51, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s15 |
| ; VI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; VI-NEXT: s_or_b32 s52, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s12 |
| ; VI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; VI-NEXT: s_or_b32 s53, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s9 |
| ; VI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; VI-NEXT: s_or_b32 s54, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s7 |
| ; VI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; VI-NEXT: s_or_b32 s55, s4, s5 |
| ; VI-NEXT: s_cbranch_execnz .LBB31_3 |
| ; VI-NEXT: .LBB31_2: ; %cmp.true |
| ; VI-NEXT: s_add_i32 s16, s16, 3 |
| ; VI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s35, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s17, s17, 3 |
| ; VI-NEXT: s_add_i32 s36, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s34, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s18, s18, 3 |
| ; VI-NEXT: s_add_i32 s37, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s31, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s19, s19, 3 |
| ; VI-NEXT: s_add_i32 s38, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s30, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s20, s20, 3 |
| ; VI-NEXT: s_add_i32 s39, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s21, s21, 3 |
| ; VI-NEXT: s_add_i32 s40, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s22, s22, 3 |
| ; VI-NEXT: s_add_i32 s41, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s23, s23, 3 |
| ; VI-NEXT: s_add_i32 s42, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s24, s24, 3 |
| ; VI-NEXT: s_add_i32 s43, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s25, s25, 3 |
| ; VI-NEXT: s_add_i32 s44, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s26, s26, 3 |
| ; VI-NEXT: s_add_i32 s45, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s27, s27, 3 |
| ; VI-NEXT: s_add_i32 s46, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s28, s28, 3 |
| ; VI-NEXT: s_add_i32 s47, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s29, s29, 3 |
| ; VI-NEXT: s_add_i32 s48, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s77, s77, 3 |
| ; VI-NEXT: s_add_i32 s49, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s77, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s74, s74, 3 |
| ; VI-NEXT: s_add_i32 s50, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s74, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s15, s15, 3 |
| ; VI-NEXT: s_add_i32 s51, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s15, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s12, s12, 3 |
| ; VI-NEXT: s_add_i32 s52, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s9, s9, 3 |
| ; VI-NEXT: s_add_i32 s53, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s9, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s7, s7, 3 |
| ; VI-NEXT: s_add_i32 s54, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s7, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s55, s4, 0x30000 |
| ; VI-NEXT: .LBB31_3: ; %end |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_readlane_b32 s67, v20, 19 |
| ; VI-NEXT: v_readlane_b32 s66, v20, 18 |
| ; VI-NEXT: v_readlane_b32 s65, v20, 17 |
| ; VI-NEXT: v_readlane_b32 s64, v20, 16 |
| ; VI-NEXT: v_readlane_b32 s55, v20, 15 |
| ; VI-NEXT: v_readlane_b32 s54, v20, 14 |
| ; VI-NEXT: v_readlane_b32 s53, v20, 13 |
| ; VI-NEXT: v_readlane_b32 s52, v20, 12 |
| ; VI-NEXT: v_readlane_b32 s51, v20, 11 |
| ; VI-NEXT: v_readlane_b32 s50, v20, 10 |
| ; VI-NEXT: v_readlane_b32 s49, v20, 9 |
| ; VI-NEXT: v_readlane_b32 s48, v20, 8 |
| ; VI-NEXT: v_readlane_b32 s39, v20, 7 |
| ; VI-NEXT: v_readlane_b32 s38, v20, 6 |
| ; VI-NEXT: v_readlane_b32 s37, v20, 5 |
| ; VI-NEXT: v_readlane_b32 s36, v20, 4 |
| ; VI-NEXT: v_readlane_b32 s35, v20, 3 |
| ; VI-NEXT: v_readlane_b32 s34, v20, 2 |
| ; VI-NEXT: v_readlane_b32 s31, v20, 1 |
| ; VI-NEXT: v_readlane_b32 s30, v20, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB31_4: |
| ; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; VI-NEXT: s_branch .LBB31_2 |
| ; |
| ; GFX9-LABEL: bitcast_v40i16_to_v20f32_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_readfirstlane_b32 s56, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s58, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s60, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s62, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s72, v1 |
| ; GFX9-NEXT: v_readfirstlane_b32 s74, v0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: s_lshr_b32 s4, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s5, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s6, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s7, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s8, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s9, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s10, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s11, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s12, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s16, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s56, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s58, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s60, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s62, 16 |
| ; GFX9-NEXT: s_lshr_b32 s73, s72, 16 |
| ; GFX9-NEXT: s_lshr_b32 s75, s74, 16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s42, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_cmp_lg_u32 s42, 0 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s50, s74, s75 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s51, s72, s73 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s52, s62, s63 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s53, s60, s61 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s54, s58, s59 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s55, s56, s57 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB31_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB31_4 |
| ; GFX9-NEXT: .LBB31_2: ; %cmp.true |
| ; GFX9-NEXT: v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_branch .LBB31_5 |
| ; GFX9-NEXT: .LBB31_3: |
| ; GFX9-NEXT: s_branch .LBB31_2 |
| ; GFX9-NEXT: .LBB31_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB31_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v40i16_to_v20f32_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s45, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s46, v0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s56, v2 |
| ; GFX11-NEXT: s_lshr_b32 s41, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s6, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s5, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s4, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: s_lshr_b32 s58, s45, 16 |
| ; GFX11-NEXT: s_lshr_b32 s59, s46, 16 |
| ; GFX11-NEXT: s_mov_b32 s40, 0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s56, 0 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s57 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s47 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s44 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s18, s46, s59 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s19, s45, s58 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB31_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB31_4 |
| ; GFX11-NEXT: .LBB31_2: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v18, s18, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v19, s19, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB31_3: |
| ; GFX11-NEXT: s_branch .LBB31_2 |
| ; GFX11-NEXT: .LBB31_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31 |
| ; 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 <40 x i16> %a, splat (i16 3) |
| %a2 = bitcast <40 x i16> %a1 to <20 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x i16> %a to <20 x float> |
| br label %end |
| |
| end: |
| %phi = phi <20 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x float> %phi |
| } |
| |
| define <40 x half> @bitcast_v20f32_to_v40f16(<20 x float> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v20f32_to_v40f16: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr31 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr29 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr26 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr25 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr24 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr23 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; implicit-def: $vgpr22 |
| ; SI-NEXT: ; implicit-def: $vgpr30 |
| ; SI-NEXT: ; implicit-def: $vgpr21 |
| ; SI-NEXT: ; implicit-def: $vgpr28 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr27 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB32_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v31, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v34, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB32_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB32_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: v_add_f32_e32 v1, 1.0, v1 |
| ; SI-NEXT: v_add_f32_e32 v0, 1.0, v0 |
| ; SI-NEXT: v_add_f32_e32 v3, 1.0, v3 |
| ; SI-NEXT: v_add_f32_e32 v2, 1.0, v2 |
| ; SI-NEXT: v_add_f32_e32 v5, 1.0, v5 |
| ; SI-NEXT: v_add_f32_e32 v4, 1.0, v4 |
| ; SI-NEXT: v_add_f32_e32 v7, 1.0, v7 |
| ; SI-NEXT: v_add_f32_e32 v6, 1.0, v6 |
| ; SI-NEXT: v_add_f32_e32 v9, 1.0, v9 |
| ; SI-NEXT: v_add_f32_e32 v8, 1.0, v8 |
| ; SI-NEXT: v_add_f32_e32 v11, 1.0, v11 |
| ; SI-NEXT: v_add_f32_e32 v10, 1.0, v10 |
| ; SI-NEXT: v_add_f32_e32 v13, 1.0, v13 |
| ; SI-NEXT: v_add_f32_e32 v12, 1.0, v12 |
| ; SI-NEXT: v_add_f32_e32 v15, 1.0, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 1.0, v14 |
| ; SI-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; SI-NEXT: v_add_f32_e32 v16, 1.0, v16 |
| ; SI-NEXT: v_add_f32_e32 v19, 1.0, v19 |
| ; SI-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v31, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v34, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB32_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v34 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v31 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v29 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v25 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v35 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v23 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v32 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v22 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v30 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v28 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v27 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v34 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v31 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v29 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v26 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v25 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v24 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20f32_to_v40f16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; 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: 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 .LBB32_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB32_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB32_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e32 v19, 1.0, v19 |
| ; VI-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; VI-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; 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_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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB32_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v20f32_to_v40f16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; 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 .LBB32_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB32_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB32_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19 |
| ; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18 |
| ; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17 |
| ; 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_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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB32_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_mov_b32 s4, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v39, v0, s4 |
| ; GFX9-NEXT: v_perm_b32 v1, v38, v1, s4 |
| ; GFX9-NEXT: v_perm_b32 v2, v37, v2, s4 |
| ; GFX9-NEXT: v_perm_b32 v3, v36, v3, s4 |
| ; GFX9-NEXT: v_perm_b32 v4, v35, v4, s4 |
| ; GFX9-NEXT: v_perm_b32 v5, v34, v5, s4 |
| ; GFX9-NEXT: v_perm_b32 v6, v33, v6, s4 |
| ; GFX9-NEXT: v_perm_b32 v7, v32, v7, s4 |
| ; GFX9-NEXT: v_perm_b32 v8, v31, v8, s4 |
| ; GFX9-NEXT: v_perm_b32 v9, v30, v9, s4 |
| ; GFX9-NEXT: v_perm_b32 v10, v29, v10, s4 |
| ; GFX9-NEXT: v_perm_b32 v11, v28, v11, s4 |
| ; GFX9-NEXT: v_perm_b32 v12, v27, v12, s4 |
| ; GFX9-NEXT: v_perm_b32 v13, v26, v13, s4 |
| ; GFX9-NEXT: v_perm_b32 v14, v25, v14, s4 |
| ; GFX9-NEXT: v_perm_b32 v15, v24, v15, s4 |
| ; GFX9-NEXT: v_perm_b32 v16, v23, v16, s4 |
| ; GFX9-NEXT: v_perm_b32 v17, v22, v17, s4 |
| ; GFX9-NEXT: v_perm_b32 v18, v21, v18, s4 |
| ; GFX9-NEXT: v_perm_b32 v19, v20, v19, s4 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v20f32_to_v40f16: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB32_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2 |
| ; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0 |
| ; GFX11-TRUE16-NEXT: .LBB32_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v20f32_to_v40f16: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB32_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB32_2: ; %Flow |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB32_4 |
| ; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2 |
| ; GFX11-FAKE16-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB32_4: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v39, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v37, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v36, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v35, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v34, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v33, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v32, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v31, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v30, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v29, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v28, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v27, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v26, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v25, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v24, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v23, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v22, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v21, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v20, v19, 0x5040100 |
| ; GFX11-FAKE16-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 <20 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <20 x float> %a1 to <40 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x float> %a to <40 x half> |
| br label %end |
| |
| end: |
| %phi = phi <40 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x half> %phi |
| } |
| |
| define inreg <40 x half> @bitcast_v20f32_to_v40f16_scalar(<20 x float> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v20f32_to_v40f16_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s8, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s5, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s8, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB33_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_lshr_b32 s89, s5, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s72, s17, 16 |
| ; SI-NEXT: s_lshr_b64 s[10:11], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b64 s[12:13], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[14:15], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[40:41], s[28:29], 16 |
| ; SI-NEXT: s_lshr_b64 s[42:43], s[26:27], 16 |
| ; SI-NEXT: s_lshr_b64 s[44:45], s[24:25], 16 |
| ; SI-NEXT: s_lshr_b64 s[46:47], s[22:23], 16 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[20:21], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[18:19], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[16:17], 16 |
| ; SI-NEXT: s_cbranch_execnz .LBB33_4 |
| ; SI-NEXT: .LBB33_2: ; %cmp.true |
| ; SI-NEXT: v_add_f32_e64 v19, s5, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v18, s4, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v17, s7, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v16, s6, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[20:21], v[18:19], 16 |
| ; SI-NEXT: v_add_f32_e64 v15, s9, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v14, s8, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[21:22], v[16:17], 16 |
| ; SI-NEXT: v_add_f32_e64 v13, s29, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v12, s28, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[22:23], v[14:15], 16 |
| ; SI-NEXT: v_add_f32_e64 v11, s27, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v10, s26, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[23:24], v[12:13], 16 |
| ; SI-NEXT: v_add_f32_e64 v9, s25, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v8, s24, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[24:25], v[10:11], 16 |
| ; SI-NEXT: v_add_f32_e64 v7, s23, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v6, s22, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[25:26], v[8:9], 16 |
| ; SI-NEXT: v_add_f32_e64 v5, s21, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v4, s20, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[26:27], v[6:7], 16 |
| ; SI-NEXT: v_add_f32_e64 v3, s19, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v2, s18, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[27:28], v[4:5], 16 |
| ; SI-NEXT: v_add_f32_e64 v1, s17, 1.0 |
| ; SI-NEXT: v_add_f32_e64 v0, s16, 1.0 |
| ; SI-NEXT: v_lshr_b64 v[28:29], v[2:3], 16 |
| ; SI-NEXT: v_lshr_b64 v[29:30], v[0:1], 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v1 |
| ; SI-NEXT: s_branch .LBB33_5 |
| ; SI-NEXT: .LBB33_3: |
| ; SI-NEXT: ; implicit-def: $sgpr60 |
| ; SI-NEXT: ; implicit-def: $sgpr72 |
| ; SI-NEXT: ; implicit-def: $sgpr58 |
| ; SI-NEXT: ; implicit-def: $sgpr73 |
| ; SI-NEXT: ; implicit-def: $sgpr56 |
| ; SI-NEXT: ; implicit-def: $sgpr74 |
| ; SI-NEXT: ; implicit-def: $sgpr46 |
| ; SI-NEXT: ; implicit-def: $sgpr75 |
| ; SI-NEXT: ; implicit-def: $sgpr44 |
| ; SI-NEXT: ; implicit-def: $sgpr76 |
| ; SI-NEXT: ; implicit-def: $sgpr42 |
| ; SI-NEXT: ; implicit-def: $sgpr77 |
| ; SI-NEXT: ; implicit-def: $sgpr40 |
| ; SI-NEXT: ; implicit-def: $sgpr78 |
| ; SI-NEXT: ; implicit-def: $sgpr14 |
| ; SI-NEXT: ; implicit-def: $sgpr79 |
| ; SI-NEXT: ; implicit-def: $sgpr12 |
| ; SI-NEXT: ; implicit-def: $sgpr88 |
| ; SI-NEXT: ; implicit-def: $sgpr10 |
| ; SI-NEXT: ; implicit-def: $sgpr89 |
| ; SI-NEXT: s_branch .LBB33_2 |
| ; SI-NEXT: .LBB33_4: |
| ; SI-NEXT: v_mov_b32_e32 v0, s16 |
| ; SI-NEXT: v_mov_b32_e32 v1, s17 |
| ; SI-NEXT: v_mov_b32_e32 v2, s18 |
| ; SI-NEXT: v_mov_b32_e32 v3, s19 |
| ; SI-NEXT: v_mov_b32_e32 v4, s20 |
| ; SI-NEXT: v_mov_b32_e32 v5, s21 |
| ; SI-NEXT: v_mov_b32_e32 v6, s22 |
| ; SI-NEXT: v_mov_b32_e32 v7, s23 |
| ; SI-NEXT: v_mov_b32_e32 v8, s24 |
| ; SI-NEXT: v_mov_b32_e32 v9, s25 |
| ; SI-NEXT: v_mov_b32_e32 v10, s26 |
| ; SI-NEXT: v_mov_b32_e32 v11, s27 |
| ; SI-NEXT: v_mov_b32_e32 v12, s28 |
| ; SI-NEXT: v_mov_b32_e32 v13, s29 |
| ; SI-NEXT: v_mov_b32_e32 v14, s8 |
| ; SI-NEXT: v_mov_b32_e32 v15, s9 |
| ; SI-NEXT: v_mov_b32_e32 v16, s6 |
| ; SI-NEXT: v_mov_b32_e32 v17, s7 |
| ; SI-NEXT: v_mov_b32_e32 v18, s4 |
| ; SI-NEXT: v_mov_b32_e32 v19, s5 |
| ; SI-NEXT: v_mov_b32_e32 v48, s72 |
| ; SI-NEXT: v_mov_b32_e32 v39, s73 |
| ; SI-NEXT: v_mov_b32_e32 v38, s74 |
| ; SI-NEXT: v_mov_b32_e32 v37, s75 |
| ; SI-NEXT: v_mov_b32_e32 v36, s76 |
| ; SI-NEXT: v_mov_b32_e32 v35, s77 |
| ; SI-NEXT: v_mov_b32_e32 v34, s78 |
| ; SI-NEXT: v_mov_b32_e32 v33, s79 |
| ; SI-NEXT: v_mov_b32_e32 v32, s88 |
| ; SI-NEXT: v_mov_b32_e32 v31, s89 |
| ; SI-NEXT: v_mov_b32_e32 v20, s10 |
| ; SI-NEXT: v_mov_b32_e32 v21, s12 |
| ; SI-NEXT: v_mov_b32_e32 v22, s14 |
| ; SI-NEXT: v_mov_b32_e32 v23, s40 |
| ; SI-NEXT: v_mov_b32_e32 v24, s42 |
| ; SI-NEXT: v_mov_b32_e32 v25, s44 |
| ; SI-NEXT: v_mov_b32_e32 v26, s46 |
| ; SI-NEXT: v_mov_b32_e32 v27, s56 |
| ; SI-NEXT: v_mov_b32_e32 v28, s58 |
| ; SI-NEXT: v_mov_b32_e32 v29, s60 |
| ; SI-NEXT: .LBB33_5: ; %end |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v29 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v48 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v28 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v27 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v25 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v35 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v23 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v34 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v22 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v32 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v31 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v29 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v28 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v27 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v26 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v25 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v24 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v20f32_to_v40f16_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB33_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_lshr_b32 s12, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s10, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s46, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: s_cbranch_execnz .LBB33_4 |
| ; VI-NEXT: .LBB33_2: ; %cmp.true |
| ; VI-NEXT: v_add_f32_e64 v19, s6, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v18, s7, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v17, s8, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v16, s9, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v15, s10, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v14, s11, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v13, s29, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v12, s28, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v11, s27, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v10, s26, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v9, s25, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v8, s24, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v7, s23, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v6, s22, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v5, s21, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v4, s20, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v3, s19, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v2, s18, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v1, s17, 1.0 |
| ; VI-NEXT: v_add_f32_e64 v0, s16, 1.0 |
| ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: s_branch .LBB33_5 |
| ; VI-NEXT: .LBB33_3: |
| ; VI-NEXT: ; implicit-def: $sgpr63 |
| ; VI-NEXT: ; implicit-def: $sgpr62 |
| ; VI-NEXT: ; implicit-def: $sgpr61 |
| ; VI-NEXT: ; implicit-def: $sgpr60 |
| ; VI-NEXT: ; implicit-def: $sgpr59 |
| ; VI-NEXT: ; implicit-def: $sgpr58 |
| ; VI-NEXT: ; implicit-def: $sgpr57 |
| ; VI-NEXT: ; implicit-def: $sgpr56 |
| ; VI-NEXT: ; implicit-def: $sgpr47 |
| ; VI-NEXT: ; implicit-def: $sgpr46 |
| ; VI-NEXT: ; implicit-def: $sgpr45 |
| ; VI-NEXT: ; implicit-def: $sgpr44 |
| ; VI-NEXT: ; implicit-def: $sgpr43 |
| ; VI-NEXT: ; implicit-def: $sgpr42 |
| ; VI-NEXT: ; implicit-def: $sgpr41 |
| ; VI-NEXT: ; implicit-def: $sgpr40 |
| ; VI-NEXT: ; implicit-def: $sgpr15 |
| ; VI-NEXT: ; implicit-def: $sgpr14 |
| ; VI-NEXT: ; implicit-def: $sgpr13 |
| ; VI-NEXT: ; implicit-def: $sgpr12 |
| ; VI-NEXT: s_branch .LBB33_2 |
| ; VI-NEXT: .LBB33_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s16 |
| ; VI-NEXT: v_mov_b32_e32 v1, s17 |
| ; VI-NEXT: v_mov_b32_e32 v2, s18 |
| ; VI-NEXT: v_mov_b32_e32 v3, s19 |
| ; VI-NEXT: v_mov_b32_e32 v4, s20 |
| ; VI-NEXT: v_mov_b32_e32 v5, s21 |
| ; VI-NEXT: v_mov_b32_e32 v6, s22 |
| ; VI-NEXT: v_mov_b32_e32 v7, s23 |
| ; VI-NEXT: v_mov_b32_e32 v8, s24 |
| ; VI-NEXT: v_mov_b32_e32 v9, s25 |
| ; VI-NEXT: v_mov_b32_e32 v10, s26 |
| ; VI-NEXT: v_mov_b32_e32 v11, s27 |
| ; VI-NEXT: v_mov_b32_e32 v12, s28 |
| ; VI-NEXT: v_mov_b32_e32 v13, s29 |
| ; VI-NEXT: v_mov_b32_e32 v14, s11 |
| ; VI-NEXT: v_mov_b32_e32 v15, s10 |
| ; VI-NEXT: v_mov_b32_e32 v16, s9 |
| ; VI-NEXT: v_mov_b32_e32 v17, s8 |
| ; VI-NEXT: v_mov_b32_e32 v18, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: v_mov_b32_e32 v39, s63 |
| ; VI-NEXT: v_mov_b32_e32 v38, s62 |
| ; VI-NEXT: v_mov_b32_e32 v37, s61 |
| ; VI-NEXT: v_mov_b32_e32 v36, s60 |
| ; VI-NEXT: v_mov_b32_e32 v35, s59 |
| ; VI-NEXT: v_mov_b32_e32 v34, s58 |
| ; VI-NEXT: v_mov_b32_e32 v33, s57 |
| ; VI-NEXT: v_mov_b32_e32 v32, s56 |
| ; VI-NEXT: v_mov_b32_e32 v31, s47 |
| ; VI-NEXT: v_mov_b32_e32 v30, s46 |
| ; VI-NEXT: v_mov_b32_e32 v29, s45 |
| ; VI-NEXT: v_mov_b32_e32 v28, s44 |
| ; VI-NEXT: v_mov_b32_e32 v27, s43 |
| ; VI-NEXT: v_mov_b32_e32 v26, s42 |
| ; VI-NEXT: v_mov_b32_e32 v25, s41 |
| ; VI-NEXT: v_mov_b32_e32 v24, s40 |
| ; VI-NEXT: v_mov_b32_e32 v23, s15 |
| ; VI-NEXT: v_mov_b32_e32 v22, s14 |
| ; VI-NEXT: v_mov_b32_e32 v21, s13 |
| ; VI-NEXT: v_mov_b32_e32 v20, s12 |
| ; VI-NEXT: .LBB33_5: ; %end |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v20f32_to_v40f16_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s10, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s11, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB33_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_lshr_b32 s12, s6, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s7, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s8, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s9, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s10, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s11, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s43, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s58, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s60, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s62, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s16, 16 |
| ; GFX9-NEXT: s_cbranch_execnz .LBB33_4 |
| ; GFX9-NEXT: .LBB33_2: ; %cmp.true |
| ; GFX9-NEXT: v_add_f32_e64 v19, s6, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v18, s7, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v17, s8, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v16, s9, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v15, s10, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v14, s11, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v13, s29, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v12, s28, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v11, s27, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v10, s26, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v9, s25, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v8, s24, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v7, s23, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v6, s22, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v5, s21, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v4, s20, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v3, s19, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v2, s18, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v1, s17, 1.0 |
| ; GFX9-NEXT: v_add_f32_e64 v0, s16, 1.0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: s_branch .LBB33_5 |
| ; GFX9-NEXT: .LBB33_3: |
| ; GFX9-NEXT: ; implicit-def: $sgpr63 |
| ; GFX9-NEXT: ; implicit-def: $sgpr62 |
| ; GFX9-NEXT: ; implicit-def: $sgpr61 |
| ; GFX9-NEXT: ; implicit-def: $sgpr60 |
| ; GFX9-NEXT: ; implicit-def: $sgpr59 |
| ; GFX9-NEXT: ; implicit-def: $sgpr58 |
| ; GFX9-NEXT: ; implicit-def: $sgpr57 |
| ; GFX9-NEXT: ; implicit-def: $sgpr56 |
| ; GFX9-NEXT: ; implicit-def: $sgpr47 |
| ; GFX9-NEXT: ; implicit-def: $sgpr46 |
| ; GFX9-NEXT: ; implicit-def: $sgpr45 |
| ; GFX9-NEXT: ; implicit-def: $sgpr44 |
| ; GFX9-NEXT: ; implicit-def: $sgpr43 |
| ; GFX9-NEXT: ; implicit-def: $sgpr42 |
| ; GFX9-NEXT: ; implicit-def: $sgpr41 |
| ; GFX9-NEXT: ; implicit-def: $sgpr40 |
| ; GFX9-NEXT: ; implicit-def: $sgpr15 |
| ; GFX9-NEXT: ; implicit-def: $sgpr14 |
| ; GFX9-NEXT: ; implicit-def: $sgpr13 |
| ; GFX9-NEXT: ; implicit-def: $sgpr12 |
| ; GFX9-NEXT: s_branch .LBB33_2 |
| ; GFX9-NEXT: .LBB33_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s28 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s29 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s6 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v33, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v32, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s15 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s14 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s13 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s12 |
| ; GFX9-NEXT: .LBB33_5: ; %end |
| ; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX9-NEXT: v_lshl_or_b32 v0, v39, 16, v0 |
| ; GFX9-NEXT: v_lshl_or_b32 v1, v38, 16, v1 |
| ; GFX9-NEXT: v_lshl_or_b32 v2, v37, 16, v2 |
| ; GFX9-NEXT: v_lshl_or_b32 v3, v36, 16, v3 |
| ; GFX9-NEXT: v_lshl_or_b32 v4, v35, 16, v4 |
| ; GFX9-NEXT: v_lshl_or_b32 v5, v34, 16, v5 |
| ; GFX9-NEXT: v_lshl_or_b32 v6, v33, 16, v6 |
| ; GFX9-NEXT: v_lshl_or_b32 v7, v32, 16, v7 |
| ; GFX9-NEXT: v_lshl_or_b32 v8, v31, 16, v8 |
| ; GFX9-NEXT: v_lshl_or_b32 v9, v30, 16, v9 |
| ; GFX9-NEXT: v_lshl_or_b32 v10, v29, 16, v10 |
| ; GFX9-NEXT: v_lshl_or_b32 v11, v28, 16, v11 |
| ; GFX9-NEXT: v_lshl_or_b32 v12, v27, 16, v12 |
| ; GFX9-NEXT: v_lshl_or_b32 v13, v26, 16, v13 |
| ; GFX9-NEXT: v_lshl_or_b32 v14, v25, 16, v14 |
| ; GFX9-NEXT: v_lshl_or_b32 v15, v24, 16, v15 |
| ; GFX9-NEXT: v_lshl_or_b32 v16, v23, 16, v16 |
| ; GFX9-NEXT: v_lshl_or_b32 v17, v22, 16, v17 |
| ; GFX9-NEXT: v_lshl_or_b32 v18, v21, 16, v18 |
| ; GFX9-NEXT: v_lshl_or_b32 v19, v20, 16, v19 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v20f32_to_v40f16_scalar: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-TRUE16-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB33_3 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s4, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s5, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s29, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s28, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s27, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s26, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s24, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s23, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s22, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s21, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s20, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s19, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s18, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s17, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s16, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s47, s3, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s56, s2, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s57, s1, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s58, s0, 16 |
| ; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB33_4 |
| ; GFX11-TRUE16-NEXT: .LBB33_2: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v19, s4, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v18, s5, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v17, s29, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v16, s28, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v15, s27, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v14, s26, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v13, s25, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v12, s24, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v11, s23, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v10, s22, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v9, s21, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v8, s20, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v7, s19, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v6, s18, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v5, s17, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v4, s16, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v3, s3, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v2, s2, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v1, s1, 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f32_e64 v0, s0, 1.0 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-TRUE16-NEXT: s_branch .LBB33_5 |
| ; GFX11-TRUE16-NEXT: .LBB33_3: |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr58 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr57 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr56 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr47 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr46 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr45 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr44 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr43 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr41 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr40 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr15 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr14 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr13 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr12 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr11 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr10 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr9 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr8 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr7 |
| ; GFX11-TRUE16-NEXT: s_branch .LBB33_2 |
| ; GFX11-TRUE16-NEXT: .LBB33_4: |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v39, s58 :: v_dual_mov_b32 v38, s57 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v37, s56 :: v_dual_mov_b32 v36, s47 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v35, s46 :: v_dual_mov_b32 v34, s45 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v33, s44 :: v_dual_mov_b32 v32, s43 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, s42 :: v_dual_mov_b32 v30, s41 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, s40 :: v_dual_mov_b32 v28, s15 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, s14 :: v_dual_mov_b32 v26, s13 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, s12 :: v_dual_mov_b32 v24, s11 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, s10 :: v_dual_mov_b32 v22, s9 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, s8 :: v_dual_mov_b32 v20, s7 |
| ; GFX11-TRUE16-NEXT: .LBB33_5: ; %end |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v39.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v38.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v37.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v36.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v35.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v34.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v33.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v32.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v31.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v30.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v29.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v28.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v27.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v26.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v25.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v24.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v23.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v22.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v21.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v20.l |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v20f32_to_v40f16_scalar: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-FAKE16-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-FAKE16-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB33_3 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s4, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s5, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s29, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s28, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s27, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s26, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s25, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s24, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s23, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s40, s22, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s21, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s20, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s19, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s18, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s17, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s16, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s47, s3, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s56, s2, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s57, s1, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s58, s0, 16 |
| ; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB33_4 |
| ; GFX11-FAKE16-NEXT: .LBB33_2: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v15, s4, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v16, s5, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v17, s29, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v18, s28, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v19, s27, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v10, s26, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v11, s25, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v12, s24, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v13, s23, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v14, s22, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v5, s21, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v6, s20, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v7, s19, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v8, s18, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v9, s17, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v0, s16, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v1, s3, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v2, s2, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v3, s1, 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f32_e64 v4, s0, 1.0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v4 |
| ; GFX11-FAKE16-NEXT: s_branch .LBB33_5 |
| ; GFX11-FAKE16-NEXT: .LBB33_3: |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr58 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr57 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr56 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr47 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr46 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr45 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr44 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr43 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr41 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr40 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr15 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr14 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr13 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr12 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr11 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr10 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr9 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr8 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr7 |
| ; GFX11-FAKE16-NEXT: s_branch .LBB33_2 |
| ; GFX11-FAKE16-NEXT: .LBB33_4: |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v3, s1 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v1, s3 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s16 :: v_dual_mov_b32 v9, s17 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v6, s20 :: v_dual_mov_b32 v5, s21 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s22 :: v_dual_mov_b32 v13, s23 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v11, s25 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s26 :: v_dual_mov_b32 v19, s27 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v18, s28 :: v_dual_mov_b32 v17, s29 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v15, s4 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v39, s58 :: v_dual_mov_b32 v38, s57 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, s56 :: v_dual_mov_b32 v36, s47 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v35, s46 :: v_dual_mov_b32 v34, s45 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v33, s44 :: v_dual_mov_b32 v32, s43 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v31, s42 :: v_dual_mov_b32 v30, s41 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v29, s40 :: v_dual_mov_b32 v28, s15 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v27, s14 :: v_dual_mov_b32 v26, s13 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v25, s12 :: v_dual_mov_b32 v24, s11 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v23, s10 :: v_dual_mov_b32 v22, s9 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v21, s8 :: v_dual_mov_b32 v20, s7 |
| ; GFX11-FAKE16-NEXT: .LBB33_5: ; %end |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v1 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v0 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v39, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v38, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v36, 16, v48 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v35, 16, v49 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v6 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xffff, v5 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v34, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v33, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v31, 16, v35 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v30, 16, v36 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v30, 0xffff, v11 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xffff, v10 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v29, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v28, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v26, 16, v30 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v25, 16, v31 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v16 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v15 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v24, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v23, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v21, 16, v25 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v20, 16, v26 |
| ; GFX11-FAKE16-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 <20 x float> %a, splat (float 1.000000e+00) |
| %a2 = bitcast <20 x float> %a1 to <40 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <20 x float> %a to <40 x half> |
| br label %end |
| |
| end: |
| %phi = phi <40 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x half> %phi |
| } |
| |
| define <20 x float> @bitcast_v40f16_to_v20f32(<40 x half> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v40f16_to_v20f32: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_mov_b32_e32 v32, v19 |
| ; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: v_mov_b32_e32 v33, v18 |
| ; SI-NEXT: v_mov_b32_e32 v43, v0 |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; SI-NEXT: v_mov_b32_e32 v34, v17 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; SI-NEXT: v_mov_b32_e32 v35, v16 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; SI-NEXT: v_mov_b32_e32 v36, v15 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v35 |
| ; SI-NEXT: v_mov_b32_e32 v37, v14 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v36 |
| ; SI-NEXT: v_mov_b32_e32 v38, v13 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v37 |
| ; SI-NEXT: v_mov_b32_e32 v39, v12 |
| ; SI-NEXT: v_mov_b32_e32 v48, v11 |
| ; SI-NEXT: v_mov_b32_e32 v49, v10 |
| ; SI-NEXT: v_mov_b32_e32 v50, v9 |
| ; SI-NEXT: v_mov_b32_e32 v51, v8 |
| ; SI-NEXT: v_mov_b32_e32 v52, v7 |
| ; SI-NEXT: v_mov_b32_e32 v53, v6 |
| ; SI-NEXT: v_mov_b32_e32 v54, v5 |
| ; SI-NEXT: v_mov_b32_e32 v55, v4 |
| ; SI-NEXT: v_mov_b32_e32 v40, v3 |
| ; SI-NEXT: v_mov_b32_e32 v41, v2 |
| ; SI-NEXT: v_mov_b32_e32 v42, v1 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v38 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v39 |
| ; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v48 |
| ; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v49 |
| ; SI-NEXT: v_lshrrev_b32_e32 v62, 16, v50 |
| ; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v51 |
| ; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v52 |
| ; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v53 |
| ; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v54 |
| ; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v55 |
| ; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v40 |
| ; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v41 |
| ; SI-NEXT: v_lshrrev_b32_e32 v58, 16, v42 |
| ; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v43 |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB34_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v43 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v59 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v42 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v58 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; SI-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v41 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v57 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v40 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v56 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v55 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v47 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v54 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v46 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v53 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v45 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v52 |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v44 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v51 |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v63 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v50 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v62 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v49 |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v61 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v60 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v39 |
| ; SI-NEXT: ; implicit-def: $vgpr43 |
| ; SI-NEXT: ; implicit-def: $vgpr42 |
| ; SI-NEXT: ; implicit-def: $vgpr41 |
| ; SI-NEXT: ; implicit-def: $vgpr40 |
| ; SI-NEXT: ; implicit-def: $vgpr55 |
| ; SI-NEXT: ; implicit-def: $vgpr54 |
| ; SI-NEXT: ; implicit-def: $vgpr53 |
| ; SI-NEXT: ; implicit-def: $vgpr52 |
| ; SI-NEXT: ; implicit-def: $vgpr51 |
| ; SI-NEXT: ; implicit-def: $vgpr50 |
| ; SI-NEXT: ; implicit-def: $vgpr49 |
| ; SI-NEXT: ; implicit-def: $vgpr48 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr59 |
| ; SI-NEXT: ; implicit-def: $vgpr58 |
| ; SI-NEXT: ; implicit-def: $vgpr57 |
| ; SI-NEXT: ; implicit-def: $vgpr56 |
| ; SI-NEXT: ; implicit-def: $vgpr47 |
| ; SI-NEXT: ; implicit-def: $vgpr46 |
| ; SI-NEXT: ; implicit-def: $vgpr45 |
| ; SI-NEXT: ; implicit-def: $vgpr44 |
| ; SI-NEXT: ; implicit-def: $vgpr63 |
| ; SI-NEXT: ; implicit-def: $vgpr62 |
| ; SI-NEXT: ; implicit-def: $vgpr61 |
| ; SI-NEXT: ; implicit-def: $vgpr60 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v38 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v37 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v36 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v35 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v34 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v33 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v19 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v32 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: .LBB34_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB34_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v0, v59 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, v58 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, v43 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v42 |
| ; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; SI-NEXT: v_or_b32_e32 v1, v3, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, v57 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v41 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, v40 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, v55 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v56 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, v54 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, v53 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, v47 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, v44 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, v52 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; SI-NEXT: v_or_b32_e32 v4, v5, v4 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, v46 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, v51 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, v61 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, v45 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, v49 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; SI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, v63 |
| ; SI-NEXT: v_or_b32_e32 v7, v9, v7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, v62 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; SI-NEXT: v_or_b32_e32 v8, v10, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, v50 |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, v38 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, v37 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v11 |
| ; SI-NEXT: v_or_b32_e32 v10, v12, v10 |
| ; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, v60 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, v35 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, v34 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; SI-NEXT: v_or_b32_e32 v11, v13, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, v39 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v21, v32 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, v12 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; SI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v14 |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: v_or_b32_e32 v13, v15, v13 |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v20, v20 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; SI-NEXT: v_or_b32_e32 v14, v16, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, v36 |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v17 |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: v_or_b32_e32 v16, v18, v16 |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_or_b32_e32 v17, v19, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, v33 |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_or_b32_e32 v18, v19, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v19, v21, v19 |
| ; SI-NEXT: .LBB34_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40f16_to_v20f32: |
| ; 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:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: v_mov_b32_e32 v32, v19 |
| ; VI-NEXT: v_mov_b32_e32 v33, v18 |
| ; VI-NEXT: v_mov_b32_e32 v34, v17 |
| ; VI-NEXT: v_mov_b32_e32 v35, v16 |
| ; VI-NEXT: v_mov_b32_e32 v36, v15 |
| ; VI-NEXT: v_mov_b32_e32 v37, v14 |
| ; VI-NEXT: v_mov_b32_e32 v38, v13 |
| ; VI-NEXT: v_mov_b32_e32 v39, v12 |
| ; VI-NEXT: v_mov_b32_e32 v48, v11 |
| ; VI-NEXT: v_mov_b32_e32 v49, v10 |
| ; VI-NEXT: v_mov_b32_e32 v50, v9 |
| ; VI-NEXT: v_mov_b32_e32 v51, v8 |
| ; VI-NEXT: v_mov_b32_e32 v52, v7 |
| ; VI-NEXT: v_mov_b32_e32 v53, v6 |
| ; VI-NEXT: v_mov_b32_e32 v54, v5 |
| ; VI-NEXT: v_mov_b32_e32 v55, v4 |
| ; VI-NEXT: v_mov_b32_e32 v40, v3 |
| ; VI-NEXT: v_mov_b32_e32 v41, v2 |
| ; VI-NEXT: v_mov_b32_e32 v42, v1 |
| ; VI-NEXT: v_mov_b32_e32 v43, v0 |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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_mov_b32_e32 v19, 16 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v0, v19, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v1, v19, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v2, v19, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v3, v19, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v4, v19, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v5, v19, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v6, v19, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v7, v19, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v8, v19, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v9, v19, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v10, v19, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v11, v19, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v12, v19, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v13, v19, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v14, v19, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v15, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v16, v19, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v17, v19, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v18, v19, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v19, v19, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v40, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v55, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v54, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v53, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v52, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v51, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v50, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v49, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v48, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v39, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v38, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v37, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v36, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v35, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v34, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v33, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v32, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; 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: .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_mov_b32_e32 v19, 0x200 |
| ; VI-NEXT: v_add_f16_sdwa v0, v43, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v1, 0x200, v43 |
| ; VI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; VI-NEXT: v_add_f16_sdwa v1, v42, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v2, 0x200, v42 |
| ; VI-NEXT: v_or_b32_e32 v1, v2, v1 |
| ; VI-NEXT: v_add_f16_sdwa v2, v41, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, 0x200, v41 |
| ; VI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; VI-NEXT: v_add_f16_sdwa v3, v40, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v4, 0x200, v40 |
| ; VI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; VI-NEXT: v_add_f16_sdwa v4, v55, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v5, 0x200, v55 |
| ; VI-NEXT: v_or_b32_e32 v4, v5, v4 |
| ; VI-NEXT: v_add_f16_sdwa v5, v54, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v6, 0x200, v54 |
| ; VI-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; VI-NEXT: v_add_f16_sdwa v6, v53, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v7, 0x200, v53 |
| ; VI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; VI-NEXT: v_add_f16_sdwa v7, v52, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v8, 0x200, v52 |
| ; VI-NEXT: v_or_b32_e32 v7, v8, v7 |
| ; VI-NEXT: v_add_f16_sdwa v8, v51, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v9, 0x200, v51 |
| ; VI-NEXT: v_or_b32_e32 v8, v9, v8 |
| ; VI-NEXT: v_add_f16_sdwa v9, v50, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v10, 0x200, v50 |
| ; VI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; VI-NEXT: v_add_f16_sdwa v10, v49, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v11, 0x200, v49 |
| ; VI-NEXT: v_or_b32_e32 v10, v11, v10 |
| ; VI-NEXT: v_add_f16_sdwa v11, v48, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v12, 0x200, v48 |
| ; VI-NEXT: v_or_b32_e32 v11, v12, v11 |
| ; VI-NEXT: v_add_f16_sdwa v12, v39, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v13, 0x200, v39 |
| ; VI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; VI-NEXT: v_add_f16_sdwa v13, v38, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v14, 0x200, v38 |
| ; VI-NEXT: v_or_b32_e32 v13, v14, v13 |
| ; VI-NEXT: v_add_f16_sdwa v14, v37, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v15, 0x200, v37 |
| ; VI-NEXT: v_or_b32_e32 v14, v15, v14 |
| ; VI-NEXT: v_add_f16_sdwa v15, v36, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v16, 0x200, v36 |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: v_add_f16_sdwa v16, v35, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v17, 0x200, v35 |
| ; VI-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; VI-NEXT: v_add_f16_sdwa v17, v34, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v18, 0x200, v34 |
| ; VI-NEXT: v_or_b32_e32 v17, v18, v17 |
| ; VI-NEXT: v_add_f16_sdwa v18, v33, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v20, 0x200, v33 |
| ; VI-NEXT: v_or_b32_e32 v18, v20, v18 |
| ; VI-NEXT: v_add_f16_sdwa v19, v32, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v20, 0x200, v32 |
| ; VI-NEXT: v_or_b32_e32 v19, v20, v19 |
| ; VI-NEXT: .LBB34_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40f16_to_v20f32: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_mov_b32_e32 v32, v19 |
| ; 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: v_mov_b32_e32 v33, v18 |
| ; GFX9-NEXT: v_mov_b32_e32 v43, v0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, v17 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, v16 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, v15 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, v14 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, v13 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, v12 |
| ; GFX9-NEXT: v_mov_b32_e32 v48, v11 |
| ; GFX9-NEXT: v_mov_b32_e32 v49, v10 |
| ; GFX9-NEXT: v_mov_b32_e32 v50, v9 |
| ; GFX9-NEXT: v_mov_b32_e32 v51, v8 |
| ; GFX9-NEXT: v_mov_b32_e32 v52, v7 |
| ; GFX9-NEXT: v_mov_b32_e32 v53, v6 |
| ; GFX9-NEXT: v_mov_b32_e32 v54, v5 |
| ; GFX9-NEXT: v_mov_b32_e32 v55, v4 |
| ; GFX9-NEXT: v_mov_b32_e32 v40, v3 |
| ; GFX9-NEXT: v_mov_b32_e32 v41, v2 |
| ; GFX9-NEXT: v_mov_b32_e32 v42, v1 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v48 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v49 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v50 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v51 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v52 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v53 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v54 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v55 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43 |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v35 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v34 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v33 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v32 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; 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: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; 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: buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: s_movk_i32 s7, 0x200 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_pk_add_f16 v0, v0, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, v1, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, v2, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, v3, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, v4, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, v5, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; GFX9-NEXT: v_pk_add_f16 v12, v12, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, v13, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, v14, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v15, v15, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v16, v16, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v17, v17, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v18, v18, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v19, v19, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB34_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; 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-TRUE16-LABEL: bitcast_v40f16_to_v20f32: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB34_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: .LBB34_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v40f16_to_v20f32: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v37, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v39, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v48, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v36, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v35, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v34, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v33, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v32, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v31, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v30, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v29, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v28, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v27, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v26, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v25, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v24, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v23, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v22, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v21, v19, 0x5040100 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB34_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: .LBB34_2: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-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 <40 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <40 x half> %a1 to <20 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x half> %a to <20 x float> |
| br label %end |
| |
| end: |
| %phi = phi <20 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x float> %phi |
| } |
| |
| define inreg <20 x float> @bitcast_v40f16_to_v20f32_scalar(<40 x half> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v40f16_to_v20f32_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; SI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; SI-NEXT: v_writelane_b32 v32, s64, 12 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s10, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s12, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s14, v1 |
| ; SI-NEXT: v_readfirstlane_b32 s73, v0 |
| ; SI-NEXT: v_writelane_b32 v32, s65, 13 |
| ; SI-NEXT: s_lshr_b32 s72, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s28, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s26, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; SI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; SI-NEXT: s_lshr_b32 s92, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s93, s18, 16 |
| ; SI-NEXT: s_lshr_b32 s94, s17, 16 |
| ; SI-NEXT: s_lshr_b32 s95, s16, 16 |
| ; SI-NEXT: s_lshr_b32 s7, s6, 16 |
| ; SI-NEXT: s_lshr_b32 s9, s8, 16 |
| ; SI-NEXT: s_lshr_b32 s11, s10, 16 |
| ; SI-NEXT: s_lshr_b32 s13, s12, 16 |
| ; SI-NEXT: s_lshr_b32 s15, s14, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s73, 16 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v32, s66, 14 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s67, 15 |
| ; SI-NEXT: s_cbranch_scc0 .LBB35_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s95, 16 |
| ; SI-NEXT: s_or_b32 s36, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s94, 16 |
| ; SI-NEXT: s_or_b32 s37, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s93, 16 |
| ; SI-NEXT: s_or_b32 s38, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s92, 16 |
| ; SI-NEXT: s_or_b32 s39, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; SI-NEXT: s_or_b32 s40, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; SI-NEXT: s_or_b32 s41, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; SI-NEXT: s_or_b32 s42, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; SI-NEXT: s_or_b32 s43, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; SI-NEXT: s_or_b32 s44, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; SI-NEXT: s_or_b32 s45, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s77, 16 |
| ; SI-NEXT: s_or_b32 s46, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; SI-NEXT: s_or_b32 s47, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; SI-NEXT: s_or_b32 s48, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; SI-NEXT: s_or_b32 s49, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s73, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s74, 16 |
| ; SI-NEXT: s_or_b32 s50, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s14, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s15, 16 |
| ; SI-NEXT: s_or_b32 s51, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; SI-NEXT: s_or_b32 s52, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s10, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; SI-NEXT: s_or_b32 s53, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s8, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s9, 16 |
| ; SI-NEXT: s_or_b32 s54, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s6, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s7, 16 |
| ; SI-NEXT: s_or_b32 s55, s4, s5 |
| ; SI-NEXT: s_cbranch_execnz .LBB35_4 |
| ; SI-NEXT: .LBB35_2: ; %cmp.true |
| ; SI-NEXT: v_cvt_f32_f16_e32 v0, s95 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, s16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, s94 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s17 |
| ; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, s18 |
| ; SI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, s93 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_or_b32_e32 v1, v3, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s92 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_or_b32_e32 v2, v4, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, s19 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s91 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, s20 |
| ; SI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s90 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, s21 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_or_b32_e32 v4, v6, v4 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, s89 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_or_b32_e32 v5, v7, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, s22 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, s88 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, s23 |
| ; SI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, s79 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, s24 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_or_b32_e32 v7, v9, v7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, s78 |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; SI-NEXT: v_or_b32_e32 v8, v10, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, s25 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, s77 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, s26 |
| ; SI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, s76 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, s27 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_or_b32_e32 v10, v12, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, s75 |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; SI-NEXT: v_or_b32_e32 v11, v13, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, s28 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, s72 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, s29 |
| ; SI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, s74 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, s73 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_or_b32_e32 v13, v15, v13 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, s15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; SI-NEXT: v_or_b32_e32 v14, v16, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, s14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, s13 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, s12 |
| ; SI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, s11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, s10 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_or_b32_e32 v16, v18, v16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, s9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_or_b32_e32 v17, v19, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, s8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v20, s7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v21, s6 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_or_b32_e32 v18, v19, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v19, v21, v19 |
| ; SI-NEXT: s_branch .LBB35_5 |
| ; SI-NEXT: .LBB35_3: |
| ; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; SI-NEXT: s_branch .LBB35_2 |
| ; SI-NEXT: .LBB35_4: |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_mov_b32_e32 v20, s56 |
| ; SI-NEXT: v_mov_b32_e32 v21, s57 |
| ; SI-NEXT: v_mov_b32_e32 v22, s58 |
| ; SI-NEXT: v_mov_b32_e32 v23, s59 |
| ; SI-NEXT: v_mov_b32_e32 v24, s60 |
| ; SI-NEXT: v_mov_b32_e32 v25, s61 |
| ; SI-NEXT: v_mov_b32_e32 v26, s62 |
| ; SI-NEXT: v_mov_b32_e32 v27, s63 |
| ; SI-NEXT: v_mov_b32_e32 v28, s64 |
| ; SI-NEXT: v_mov_b32_e32 v29, s65 |
| ; SI-NEXT: v_mov_b32_e32 v30, s66 |
| ; SI-NEXT: v_mov_b32_e32 v31, s67 |
| ; SI-NEXT: .LBB35_5: ; %end |
| ; SI-NEXT: v_readlane_b32 s67, v32, 15 |
| ; SI-NEXT: v_readlane_b32 s66, v32, 14 |
| ; SI-NEXT: v_readlane_b32 s65, v32, 13 |
| ; SI-NEXT: v_readlane_b32 s64, v32, 12 |
| ; SI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40f16_to_v20f32_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v32, s30, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s31, 1 |
| ; VI-NEXT: v_writelane_b32 v32, s34, 2 |
| ; VI-NEXT: v_writelane_b32 v32, s35, 3 |
| ; VI-NEXT: v_writelane_b32 v32, s36, 4 |
| ; VI-NEXT: v_writelane_b32 v32, s37, 5 |
| ; VI-NEXT: v_writelane_b32 v32, s38, 6 |
| ; VI-NEXT: v_writelane_b32 v32, s39, 7 |
| ; VI-NEXT: v_writelane_b32 v32, s48, 8 |
| ; VI-NEXT: v_writelane_b32 v32, s49, 9 |
| ; VI-NEXT: v_writelane_b32 v32, s50, 10 |
| ; VI-NEXT: v_writelane_b32 v32, s51, 11 |
| ; VI-NEXT: v_writelane_b32 v32, s52, 12 |
| ; VI-NEXT: v_writelane_b32 v32, s53, 13 |
| ; VI-NEXT: v_writelane_b32 v32, s54, 14 |
| ; VI-NEXT: v_writelane_b32 v32, s55, 15 |
| ; VI-NEXT: v_writelane_b32 v32, s64, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s14, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s73, v1 |
| ; VI-NEXT: v_readfirstlane_b32 s76, v0 |
| ; VI-NEXT: v_writelane_b32 v32, s65, 17 |
| ; VI-NEXT: s_lshr_b32 s10, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s72, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s74, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s77, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s30, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s31, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s34, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s35, s16, 16 |
| ; VI-NEXT: s_lshr_b32 s7, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s9, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s12, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s14, 16 |
| ; VI-NEXT: s_lshr_b32 s75, s73, 16 |
| ; VI-NEXT: s_lshr_b32 s78, s76, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v32, s66, 18 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s67, 19 |
| ; VI-NEXT: s_cbranch_scc0 .LBB35_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s35, 16 |
| ; VI-NEXT: s_or_b32 s36, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s5, s34, 16 |
| ; VI-NEXT: s_or_b32 s37, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s5, s31, 16 |
| ; VI-NEXT: s_or_b32 s38, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s5, s30, 16 |
| ; VI-NEXT: s_or_b32 s39, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; VI-NEXT: s_or_b32 s40, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; VI-NEXT: s_or_b32 s41, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; VI-NEXT: s_or_b32 s42, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; VI-NEXT: s_or_b32 s43, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; VI-NEXT: s_or_b32 s44, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s5, s77, 16 |
| ; VI-NEXT: s_or_b32 s45, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s5, s74, 16 |
| ; VI-NEXT: s_or_b32 s46, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; VI-NEXT: s_or_b32 s47, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; VI-NEXT: s_or_b32 s48, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; VI-NEXT: s_or_b32 s49, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s76 |
| ; VI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; VI-NEXT: s_or_b32 s50, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s73 |
| ; VI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; VI-NEXT: s_or_b32 s51, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s14 |
| ; VI-NEXT: s_lshl_b32 s5, s15, 16 |
| ; VI-NEXT: s_or_b32 s52, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s11 |
| ; VI-NEXT: s_lshl_b32 s5, s12, 16 |
| ; VI-NEXT: s_or_b32 s53, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s8 |
| ; VI-NEXT: s_lshl_b32 s5, s9, 16 |
| ; VI-NEXT: s_or_b32 s54, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s6 |
| ; VI-NEXT: s_lshl_b32 s5, s7, 16 |
| ; VI-NEXT: s_or_b32 s55, s4, s5 |
| ; VI-NEXT: s_cbranch_execnz .LBB35_4 |
| ; VI-NEXT: .LBB35_2: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v19, 0x200 |
| ; VI-NEXT: v_mov_b32_e32 v0, s35 |
| ; VI-NEXT: v_mov_b32_e32 v2, s34 |
| ; VI-NEXT: v_add_f16_sdwa v0, v0, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v1, s16, v19 |
| ; VI-NEXT: v_add_f16_sdwa v2, v2, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, s17, v19 |
| ; VI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; VI-NEXT: v_or_b32_e32 v1, v3, v2 |
| ; VI-NEXT: v_mov_b32_e32 v2, s31 |
| ; VI-NEXT: v_add_f16_sdwa v2, v2, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, s18, v19 |
| ; VI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; VI-NEXT: v_mov_b32_e32 v3, s30 |
| ; VI-NEXT: v_add_f16_sdwa v3, v3, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v4, s19, v19 |
| ; VI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; VI-NEXT: v_mov_b32_e32 v4, s91 |
| ; VI-NEXT: v_add_f16_sdwa v4, v4, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v5, s20, v19 |
| ; VI-NEXT: v_or_b32_e32 v4, v5, v4 |
| ; VI-NEXT: v_mov_b32_e32 v5, s90 |
| ; VI-NEXT: v_add_f16_sdwa v5, v5, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v6, s21, v19 |
| ; VI-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; VI-NEXT: v_mov_b32_e32 v6, s89 |
| ; VI-NEXT: v_add_f16_sdwa v6, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v7, s22, v19 |
| ; VI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; VI-NEXT: v_mov_b32_e32 v7, s88 |
| ; VI-NEXT: v_add_f16_sdwa v7, v7, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v8, s23, v19 |
| ; VI-NEXT: v_or_b32_e32 v7, v8, v7 |
| ; VI-NEXT: v_mov_b32_e32 v8, s79 |
| ; VI-NEXT: v_add_f16_sdwa v8, v8, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v9, s24, v19 |
| ; VI-NEXT: v_or_b32_e32 v8, v9, v8 |
| ; VI-NEXT: v_mov_b32_e32 v9, s77 |
| ; VI-NEXT: v_add_f16_sdwa v9, v9, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v10, s25, v19 |
| ; VI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; VI-NEXT: v_mov_b32_e32 v10, s74 |
| ; VI-NEXT: v_add_f16_sdwa v10, v10, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v11, s26, v19 |
| ; VI-NEXT: v_or_b32_e32 v10, v11, v10 |
| ; VI-NEXT: v_mov_b32_e32 v11, s72 |
| ; VI-NEXT: v_add_f16_sdwa v11, v11, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v12, s27, v19 |
| ; VI-NEXT: v_or_b32_e32 v11, v12, v11 |
| ; VI-NEXT: v_mov_b32_e32 v12, s13 |
| ; VI-NEXT: v_add_f16_sdwa v12, v12, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v13, s28, v19 |
| ; VI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; VI-NEXT: v_mov_b32_e32 v13, s10 |
| ; VI-NEXT: v_add_f16_sdwa v13, v13, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v14, s29, v19 |
| ; VI-NEXT: v_or_b32_e32 v13, v14, v13 |
| ; VI-NEXT: v_mov_b32_e32 v14, s78 |
| ; VI-NEXT: v_add_f16_sdwa v14, v14, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v15, s76, v19 |
| ; VI-NEXT: v_or_b32_e32 v14, v15, v14 |
| ; VI-NEXT: v_mov_b32_e32 v15, s75 |
| ; VI-NEXT: v_add_f16_sdwa v15, v15, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v16, s73, v19 |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: v_mov_b32_e32 v16, s15 |
| ; VI-NEXT: v_add_f16_sdwa v16, v16, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v17, s14, v19 |
| ; VI-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; VI-NEXT: v_mov_b32_e32 v17, s12 |
| ; VI-NEXT: v_add_f16_sdwa v17, v17, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v18, s11, v19 |
| ; VI-NEXT: v_or_b32_e32 v17, v18, v17 |
| ; VI-NEXT: v_mov_b32_e32 v18, s9 |
| ; VI-NEXT: v_add_f16_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v20, s8, v19 |
| ; VI-NEXT: v_or_b32_e32 v18, v20, v18 |
| ; VI-NEXT: v_mov_b32_e32 v20, s7 |
| ; VI-NEXT: v_add_f16_sdwa v20, v20, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v19, s6, v19 |
| ; VI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; VI-NEXT: s_branch .LBB35_5 |
| ; VI-NEXT: .LBB35_3: |
| ; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; VI-NEXT: s_branch .LBB35_2 |
| ; VI-NEXT: .LBB35_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_mov_b32_e32 v20, s56 |
| ; VI-NEXT: v_mov_b32_e32 v21, s57 |
| ; VI-NEXT: v_mov_b32_e32 v22, s58 |
| ; VI-NEXT: v_mov_b32_e32 v23, s59 |
| ; VI-NEXT: v_mov_b32_e32 v24, s60 |
| ; VI-NEXT: v_mov_b32_e32 v25, s61 |
| ; VI-NEXT: v_mov_b32_e32 v26, s62 |
| ; VI-NEXT: v_mov_b32_e32 v27, s63 |
| ; VI-NEXT: v_mov_b32_e32 v28, s64 |
| ; VI-NEXT: v_mov_b32_e32 v29, s65 |
| ; VI-NEXT: v_mov_b32_e32 v30, s66 |
| ; VI-NEXT: v_mov_b32_e32 v31, s67 |
| ; VI-NEXT: .LBB35_5: ; %end |
| ; VI-NEXT: v_readlane_b32 s67, v32, 19 |
| ; VI-NEXT: v_readlane_b32 s66, v32, 18 |
| ; VI-NEXT: v_readlane_b32 s65, v32, 17 |
| ; VI-NEXT: v_readlane_b32 s64, v32, 16 |
| ; VI-NEXT: v_readlane_b32 s55, v32, 15 |
| ; VI-NEXT: v_readlane_b32 s54, v32, 14 |
| ; VI-NEXT: v_readlane_b32 s53, v32, 13 |
| ; VI-NEXT: v_readlane_b32 s52, v32, 12 |
| ; VI-NEXT: v_readlane_b32 s51, v32, 11 |
| ; VI-NEXT: v_readlane_b32 s50, v32, 10 |
| ; VI-NEXT: v_readlane_b32 s49, v32, 9 |
| ; VI-NEXT: v_readlane_b32 s48, v32, 8 |
| ; VI-NEXT: v_readlane_b32 s39, v32, 7 |
| ; VI-NEXT: v_readlane_b32 s38, v32, 6 |
| ; VI-NEXT: v_readlane_b32 s37, v32, 5 |
| ; VI-NEXT: v_readlane_b32 s36, v32, 4 |
| ; VI-NEXT: v_readlane_b32 s35, v32, 3 |
| ; VI-NEXT: v_readlane_b32 s34, v32, 2 |
| ; VI-NEXT: v_readlane_b32 s31, v32, 1 |
| ; VI-NEXT: v_readlane_b32 s30, v32, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40f16_to_v20f32_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_readfirstlane_b32 s56, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s58, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s60, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s62, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s72, v1 |
| ; GFX9-NEXT: v_readfirstlane_b32 s74, v0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: s_lshr_b32 s4, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s5, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s6, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s7, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s8, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s9, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s10, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s11, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s12, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s16, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s56, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s58, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s60, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s62, 16 |
| ; GFX9-NEXT: s_lshr_b32 s73, s72, 16 |
| ; GFX9-NEXT: s_lshr_b32 s75, s74, 16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s42, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_cmp_lg_u32 s42, 0 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s50, s74, s75 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s51, s72, s73 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s52, s62, s63 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s53, s60, s61 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s54, s58, s59 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s55, s56, s57 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB35_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB35_4 |
| ; GFX9-NEXT: .LBB35_2: ; %cmp.true |
| ; GFX9-NEXT: v_mov_b32_e32 v19, 0x200 |
| ; GFX9-NEXT: v_pk_add_f16 v0, s36, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, s37, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, s38, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, s39, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, s40, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, s41, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, s42, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, s43, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, s44, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, s45, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, s46, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, s47, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v12, s48, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, s49, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, s50, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v15, s51, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v16, s52, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v17, s53, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v18, s54, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v19, s55, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_branch .LBB35_5 |
| ; GFX9-NEXT: .LBB35_3: |
| ; GFX9-NEXT: s_branch .LBB35_2 |
| ; GFX9-NEXT: .LBB35_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB35_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v40f16_to_v20f32_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s45, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s46, v0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s56, v2 |
| ; GFX11-NEXT: s_lshr_b32 s41, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s6, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s5, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s4, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: s_lshr_b32 s58, s45, 16 |
| ; GFX11-NEXT: s_lshr_b32 s59, s46, 16 |
| ; GFX11-NEXT: s_mov_b32 s40, 0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s56, 0 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s57 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s47 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s44 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s18, s46, s59 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s19, s45, s58 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB35_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB35_4 |
| ; GFX11-NEXT: .LBB35_2: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v18, 0x200, s18 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v19, 0x200, s19 op_sel_hi:[0,1] |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB35_3: |
| ; GFX11-NEXT: s_branch .LBB35_2 |
| ; GFX11-NEXT: .LBB35_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31 |
| ; 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 <40 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <40 x half> %a1 to <20 x float> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x half> %a to <20 x float> |
| br label %end |
| |
| end: |
| %phi = phi <20 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <20 x float> %phi |
| } |
| |
| define <10 x double> @bitcast_v10i64_to_v10f64(<10 x i64> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v10i64_to_v10f64: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB36_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; SI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; SI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; SI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; SI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; SI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; SI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; SI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; SI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; SI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; SI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc |
| ; SI-NEXT: .LBB36_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10i64_to_v10f64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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_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: v_add_u32_e32 v16, vcc, 3, v16 |
| ; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18 |
| ; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc |
| ; 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_v10i64_to_v10f64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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_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: v_add_co_u32_e32 v16, vcc, 3, v16 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc |
| ; 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_v10i64_to_v10f64: |
| ; 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, v20 |
| ; 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_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: v_add_co_u32 v16, vcc_lo, v16, 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 v17, null, 0, v17, vcc_lo |
| ; GFX11-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3 |
| ; GFX11-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo |
| ; 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 = add <10 x i64> %a, splat (i64 3) |
| %a2 = bitcast <10 x i64> %a1 to <10 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x i64> %a to <10 x double> |
| br label %end |
| |
| end: |
| %phi = phi <10 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x double> %phi |
| } |
| |
| define inreg <10 x double> @bitcast_v10i64_to_v10f64_scalar(<10 x i64> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v10i64_to_v10f64_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB37_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB37_3 |
| ; SI-NEXT: .LBB37_2: ; %cmp.true |
| ; SI-NEXT: s_add_u32 s16, s16, 3 |
| ; SI-NEXT: s_addc_u32 s17, s17, 0 |
| ; SI-NEXT: s_add_u32 s18, s18, 3 |
| ; SI-NEXT: s_addc_u32 s19, s19, 0 |
| ; SI-NEXT: s_add_u32 s20, s20, 3 |
| ; SI-NEXT: s_addc_u32 s21, s21, 0 |
| ; SI-NEXT: s_add_u32 s22, s22, 3 |
| ; SI-NEXT: s_addc_u32 s23, s23, 0 |
| ; SI-NEXT: s_add_u32 s24, s24, 3 |
| ; SI-NEXT: s_addc_u32 s25, s25, 0 |
| ; SI-NEXT: s_add_u32 s26, s26, 3 |
| ; SI-NEXT: s_addc_u32 s27, s27, 0 |
| ; SI-NEXT: s_add_u32 s28, s28, 3 |
| ; SI-NEXT: s_addc_u32 s29, s29, 0 |
| ; SI-NEXT: s_add_u32 s11, s11, 3 |
| ; SI-NEXT: s_addc_u32 s10, s10, 0 |
| ; SI-NEXT: s_add_u32 s9, s9, 3 |
| ; SI-NEXT: s_addc_u32 s8, s8, 0 |
| ; SI-NEXT: s_add_u32 s7, s7, 3 |
| ; SI-NEXT: s_addc_u32 s6, s6, 0 |
| ; SI-NEXT: .LBB37_3: ; %end |
| ; SI-NEXT: v_mov_b32_e32 v0, s16 |
| ; SI-NEXT: v_mov_b32_e32 v1, s17 |
| ; SI-NEXT: v_mov_b32_e32 v2, s18 |
| ; SI-NEXT: v_mov_b32_e32 v3, s19 |
| ; SI-NEXT: v_mov_b32_e32 v4, s20 |
| ; SI-NEXT: v_mov_b32_e32 v5, s21 |
| ; SI-NEXT: v_mov_b32_e32 v6, s22 |
| ; SI-NEXT: v_mov_b32_e32 v7, s23 |
| ; SI-NEXT: v_mov_b32_e32 v8, s24 |
| ; SI-NEXT: v_mov_b32_e32 v9, s25 |
| ; SI-NEXT: v_mov_b32_e32 v10, s26 |
| ; SI-NEXT: v_mov_b32_e32 v11, s27 |
| ; SI-NEXT: v_mov_b32_e32 v12, s28 |
| ; SI-NEXT: v_mov_b32_e32 v13, s29 |
| ; SI-NEXT: v_mov_b32_e32 v14, s11 |
| ; SI-NEXT: v_mov_b32_e32 v15, s10 |
| ; SI-NEXT: v_mov_b32_e32 v16, s9 |
| ; SI-NEXT: v_mov_b32_e32 v17, s8 |
| ; SI-NEXT: v_mov_b32_e32 v18, s7 |
| ; SI-NEXT: v_mov_b32_e32 v19, s6 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB37_4: |
| ; SI-NEXT: s_branch .LBB37_2 |
| ; |
| ; VI-LABEL: bitcast_v10i64_to_v10f64_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB37_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB37_3 |
| ; VI-NEXT: .LBB37_2: ; %cmp.true |
| ; VI-NEXT: s_add_u32 s16, s16, 3 |
| ; VI-NEXT: s_addc_u32 s17, s17, 0 |
| ; VI-NEXT: s_add_u32 s18, s18, 3 |
| ; VI-NEXT: s_addc_u32 s19, s19, 0 |
| ; VI-NEXT: s_add_u32 s20, s20, 3 |
| ; VI-NEXT: s_addc_u32 s21, s21, 0 |
| ; VI-NEXT: s_add_u32 s22, s22, 3 |
| ; VI-NEXT: s_addc_u32 s23, s23, 0 |
| ; VI-NEXT: s_add_u32 s24, s24, 3 |
| ; VI-NEXT: s_addc_u32 s25, s25, 0 |
| ; VI-NEXT: s_add_u32 s26, s26, 3 |
| ; VI-NEXT: s_addc_u32 s27, s27, 0 |
| ; VI-NEXT: s_add_u32 s28, s28, 3 |
| ; VI-NEXT: s_addc_u32 s29, s29, 0 |
| ; VI-NEXT: s_add_u32 s11, s11, 3 |
| ; VI-NEXT: s_addc_u32 s10, s10, 0 |
| ; VI-NEXT: s_add_u32 s9, s9, 3 |
| ; VI-NEXT: s_addc_u32 s8, s8, 0 |
| ; VI-NEXT: s_add_u32 s7, s7, 3 |
| ; VI-NEXT: s_addc_u32 s6, s6, 0 |
| ; VI-NEXT: .LBB37_3: ; %end |
| ; VI-NEXT: v_mov_b32_e32 v0, s16 |
| ; VI-NEXT: v_mov_b32_e32 v1, s17 |
| ; VI-NEXT: v_mov_b32_e32 v2, s18 |
| ; VI-NEXT: v_mov_b32_e32 v3, s19 |
| ; VI-NEXT: v_mov_b32_e32 v4, s20 |
| ; VI-NEXT: v_mov_b32_e32 v5, s21 |
| ; VI-NEXT: v_mov_b32_e32 v6, s22 |
| ; VI-NEXT: v_mov_b32_e32 v7, s23 |
| ; VI-NEXT: v_mov_b32_e32 v8, s24 |
| ; VI-NEXT: v_mov_b32_e32 v9, s25 |
| ; VI-NEXT: v_mov_b32_e32 v10, s26 |
| ; VI-NEXT: v_mov_b32_e32 v11, s27 |
| ; VI-NEXT: v_mov_b32_e32 v12, s28 |
| ; VI-NEXT: v_mov_b32_e32 v13, s29 |
| ; VI-NEXT: v_mov_b32_e32 v14, s11 |
| ; VI-NEXT: v_mov_b32_e32 v15, s10 |
| ; VI-NEXT: v_mov_b32_e32 v16, s9 |
| ; VI-NEXT: v_mov_b32_e32 v17, s8 |
| ; VI-NEXT: v_mov_b32_e32 v18, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB37_4: |
| ; VI-NEXT: s_branch .LBB37_2 |
| ; |
| ; GFX9-LABEL: bitcast_v10i64_to_v10f64_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s10, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s11, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB37_4 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB37_3 |
| ; GFX9-NEXT: .LBB37_2: ; %cmp.true |
| ; GFX9-NEXT: s_add_u32 s16, s16, 3 |
| ; GFX9-NEXT: s_addc_u32 s17, s17, 0 |
| ; GFX9-NEXT: s_add_u32 s18, s18, 3 |
| ; GFX9-NEXT: s_addc_u32 s19, s19, 0 |
| ; GFX9-NEXT: s_add_u32 s20, s20, 3 |
| ; GFX9-NEXT: s_addc_u32 s21, s21, 0 |
| ; GFX9-NEXT: s_add_u32 s22, s22, 3 |
| ; GFX9-NEXT: s_addc_u32 s23, s23, 0 |
| ; GFX9-NEXT: s_add_u32 s24, s24, 3 |
| ; GFX9-NEXT: s_addc_u32 s25, s25, 0 |
| ; GFX9-NEXT: s_add_u32 s26, s26, 3 |
| ; GFX9-NEXT: s_addc_u32 s27, s27, 0 |
| ; GFX9-NEXT: s_add_u32 s28, s28, 3 |
| ; GFX9-NEXT: s_addc_u32 s29, s29, 0 |
| ; GFX9-NEXT: s_add_u32 s11, s11, 3 |
| ; GFX9-NEXT: s_addc_u32 s10, s10, 0 |
| ; GFX9-NEXT: s_add_u32 s9, s9, 3 |
| ; GFX9-NEXT: s_addc_u32 s8, s8, 0 |
| ; GFX9-NEXT: s_add_u32 s7, s7, 3 |
| ; GFX9-NEXT: s_addc_u32 s6, s6, 0 |
| ; GFX9-NEXT: .LBB37_3: ; %end |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s28 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s29 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s6 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; GFX9-NEXT: .LBB37_4: |
| ; GFX9-NEXT: s_branch .LBB37_2 |
| ; |
| ; GFX11-LABEL: bitcast_v10i64_to_v10f64_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB37_4 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB37_3 |
| ; GFX11-NEXT: .LBB37_2: ; %cmp.true |
| ; GFX11-NEXT: s_add_u32 s0, s0, 3 |
| ; GFX11-NEXT: s_addc_u32 s1, s1, 0 |
| ; GFX11-NEXT: s_add_u32 s2, s2, 3 |
| ; GFX11-NEXT: s_addc_u32 s3, s3, 0 |
| ; GFX11-NEXT: s_add_u32 s16, s16, 3 |
| ; GFX11-NEXT: s_addc_u32 s17, s17, 0 |
| ; GFX11-NEXT: s_add_u32 s18, s18, 3 |
| ; GFX11-NEXT: s_addc_u32 s19, s19, 0 |
| ; GFX11-NEXT: s_add_u32 s20, s20, 3 |
| ; GFX11-NEXT: s_addc_u32 s21, s21, 0 |
| ; GFX11-NEXT: s_add_u32 s22, s22, 3 |
| ; GFX11-NEXT: s_addc_u32 s23, s23, 0 |
| ; GFX11-NEXT: s_add_u32 s24, s24, 3 |
| ; GFX11-NEXT: s_addc_u32 s25, s25, 0 |
| ; GFX11-NEXT: s_add_u32 s26, s26, 3 |
| ; GFX11-NEXT: s_addc_u32 s27, s27, 0 |
| ; GFX11-NEXT: s_add_u32 s28, s28, 3 |
| ; GFX11-NEXT: s_addc_u32 s29, s29, 0 |
| ; GFX11-NEXT: s_add_u32 s5, s5, 3 |
| ; GFX11-NEXT: s_addc_u32 s4, s4, 0 |
| ; GFX11-NEXT: .LBB37_3: ; %end |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB37_4: |
| ; GFX11-NEXT: s_branch .LBB37_2 |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <10 x i64> %a, splat (i64 3) |
| %a2 = bitcast <10 x i64> %a1 to <10 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x i64> %a to <10 x double> |
| br label %end |
| |
| end: |
| %phi = phi <10 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x double> %phi |
| } |
| |
| define <10 x i64> @bitcast_v10f64_to_v10i64(<10 x double> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v10f64_to_v10i64: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB38_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; SI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; SI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; SI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; SI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; SI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; SI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; SI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; SI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; SI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0 |
| ; SI-NEXT: .LBB38_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10f64_to_v10i64: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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[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: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; VI-NEXT: v_add_f64 v[18:19], v[18:19], 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_v10f64_to_v10i64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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[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: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 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_v10f64_to_v10i64: |
| ; 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, v20 |
| ; 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[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: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[18:19], v[18:19], 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 <10 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <10 x double> %a1 to <10 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x double> %a to <10 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <10 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x i64> %phi |
| } |
| |
| define inreg <10 x i64> @bitcast_v10f64_to_v10i64_scalar(<10 x double> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v10f64_to_v10i64_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; SI-NEXT: s_mov_b32 s49, s29 |
| ; SI-NEXT: s_mov_b32 s48, s28 |
| ; SI-NEXT: s_mov_b32 s47, s27 |
| ; SI-NEXT: s_mov_b32 s46, s26 |
| ; SI-NEXT: s_mov_b32 s45, s25 |
| ; SI-NEXT: s_mov_b32 s44, s24 |
| ; SI-NEXT: s_mov_b32 s43, s23 |
| ; SI-NEXT: s_mov_b32 s42, s22 |
| ; SI-NEXT: s_mov_b32 s41, s21 |
| ; SI-NEXT: s_mov_b32 s40, s20 |
| ; SI-NEXT: s_mov_b32 s39, s19 |
| ; SI-NEXT: s_mov_b32 s38, s18 |
| ; SI-NEXT: s_mov_b32 s37, s17 |
| ; SI-NEXT: s_mov_b32 s36, s16 |
| ; SI-NEXT: v_readfirstlane_b32 s55, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s54, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s53, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s52, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s51, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s50, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB39_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB39_4 |
| ; SI-NEXT: .LBB39_2: ; %cmp.true |
| ; SI-NEXT: v_add_f64 v[0:1], s[36:37], 1.0 |
| ; SI-NEXT: v_add_f64 v[2:3], s[38:39], 1.0 |
| ; SI-NEXT: v_add_f64 v[4:5], s[40:41], 1.0 |
| ; SI-NEXT: v_add_f64 v[6:7], s[42:43], 1.0 |
| ; SI-NEXT: v_add_f64 v[8:9], s[44:45], 1.0 |
| ; SI-NEXT: v_add_f64 v[10:11], s[46:47], 1.0 |
| ; SI-NEXT: v_add_f64 v[12:13], s[48:49], 1.0 |
| ; SI-NEXT: v_add_f64 v[14:15], s[50:51], 1.0 |
| ; SI-NEXT: v_add_f64 v[16:17], s[52:53], 1.0 |
| ; SI-NEXT: v_add_f64 v[18:19], s[54:55], 1.0 |
| ; SI-NEXT: s_branch .LBB39_5 |
| ; SI-NEXT: .LBB39_3: |
| ; SI-NEXT: s_branch .LBB39_2 |
| ; SI-NEXT: .LBB39_4: |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_mov_b32_e32 v20, s56 |
| ; SI-NEXT: v_mov_b32_e32 v21, s57 |
| ; SI-NEXT: v_mov_b32_e32 v22, s58 |
| ; SI-NEXT: v_mov_b32_e32 v23, s59 |
| ; SI-NEXT: v_mov_b32_e32 v24, s60 |
| ; SI-NEXT: v_mov_b32_e32 v25, s61 |
| ; SI-NEXT: v_mov_b32_e32 v26, s62 |
| ; SI-NEXT: v_mov_b32_e32 v27, s63 |
| ; SI-NEXT: v_mov_b32_e32 v28, s64 |
| ; SI-NEXT: v_mov_b32_e32 v29, s65 |
| ; SI-NEXT: v_mov_b32_e32 v30, s66 |
| ; SI-NEXT: v_mov_b32_e32 v31, s67 |
| ; SI-NEXT: .LBB39_5: ; %end |
| ; SI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10f64_to_v10i64_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; VI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; VI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; VI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; VI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; VI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; VI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; VI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; VI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; VI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; VI-NEXT: s_mov_b32 s49, s29 |
| ; VI-NEXT: s_mov_b32 s48, s28 |
| ; VI-NEXT: s_mov_b32 s47, s27 |
| ; VI-NEXT: s_mov_b32 s46, s26 |
| ; VI-NEXT: s_mov_b32 s45, s25 |
| ; VI-NEXT: s_mov_b32 s44, s24 |
| ; VI-NEXT: s_mov_b32 s43, s23 |
| ; VI-NEXT: s_mov_b32 s42, s22 |
| ; VI-NEXT: s_mov_b32 s41, s21 |
| ; VI-NEXT: s_mov_b32 s40, s20 |
| ; VI-NEXT: s_mov_b32 s39, s19 |
| ; VI-NEXT: s_mov_b32 s38, s18 |
| ; VI-NEXT: s_mov_b32 s37, s17 |
| ; VI-NEXT: s_mov_b32 s36, s16 |
| ; VI-NEXT: v_readfirstlane_b32 s55, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s54, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s53, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s52, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s51, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s50, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB39_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB39_4 |
| ; VI-NEXT: .LBB39_2: ; %cmp.true |
| ; VI-NEXT: v_add_f64 v[0:1], s[36:37], 1.0 |
| ; VI-NEXT: v_add_f64 v[2:3], s[38:39], 1.0 |
| ; VI-NEXT: v_add_f64 v[4:5], s[40:41], 1.0 |
| ; VI-NEXT: v_add_f64 v[6:7], s[42:43], 1.0 |
| ; VI-NEXT: v_add_f64 v[8:9], s[44:45], 1.0 |
| ; VI-NEXT: v_add_f64 v[10:11], s[46:47], 1.0 |
| ; VI-NEXT: v_add_f64 v[12:13], s[48:49], 1.0 |
| ; VI-NEXT: v_add_f64 v[14:15], s[50:51], 1.0 |
| ; VI-NEXT: v_add_f64 v[16:17], s[52:53], 1.0 |
| ; VI-NEXT: v_add_f64 v[18:19], s[54:55], 1.0 |
| ; VI-NEXT: s_branch .LBB39_5 |
| ; VI-NEXT: .LBB39_3: |
| ; VI-NEXT: s_branch .LBB39_2 |
| ; VI-NEXT: .LBB39_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_mov_b32_e32 v20, s56 |
| ; VI-NEXT: v_mov_b32_e32 v21, s57 |
| ; VI-NEXT: v_mov_b32_e32 v22, s58 |
| ; VI-NEXT: v_mov_b32_e32 v23, s59 |
| ; VI-NEXT: v_mov_b32_e32 v24, s60 |
| ; VI-NEXT: v_mov_b32_e32 v25, s61 |
| ; VI-NEXT: v_mov_b32_e32 v26, s62 |
| ; VI-NEXT: v_mov_b32_e32 v27, s63 |
| ; VI-NEXT: v_mov_b32_e32 v28, s64 |
| ; VI-NEXT: v_mov_b32_e32 v29, s65 |
| ; VI-NEXT: v_mov_b32_e32 v30, s66 |
| ; VI-NEXT: v_mov_b32_e32 v31, s67 |
| ; VI-NEXT: .LBB39_5: ; %end |
| ; VI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; VI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; VI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; VI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; VI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; VI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; VI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; VI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; VI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; VI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; VI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; VI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10f64_to_v10i64_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_mov_b32 s49, s29 |
| ; GFX9-NEXT: s_mov_b32 s48, s28 |
| ; GFX9-NEXT: s_mov_b32 s47, s27 |
| ; GFX9-NEXT: s_mov_b32 s46, s26 |
| ; GFX9-NEXT: s_mov_b32 s45, s25 |
| ; GFX9-NEXT: s_mov_b32 s44, s24 |
| ; GFX9-NEXT: s_mov_b32 s43, s23 |
| ; GFX9-NEXT: s_mov_b32 s42, s22 |
| ; GFX9-NEXT: s_mov_b32 s41, s21 |
| ; GFX9-NEXT: s_mov_b32 s40, s20 |
| ; GFX9-NEXT: s_mov_b32 s39, s19 |
| ; GFX9-NEXT: s_mov_b32 s38, s18 |
| ; GFX9-NEXT: s_mov_b32 s37, s17 |
| ; GFX9-NEXT: s_mov_b32 s36, s16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s55, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s54, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s53, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s52, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s51, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s50, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB39_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB39_4 |
| ; GFX9-NEXT: .LBB39_2: ; %cmp.true |
| ; GFX9-NEXT: v_add_f64 v[0:1], s[36:37], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[2:3], s[38:39], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[4:5], s[40:41], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[6:7], s[42:43], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[8:9], s[44:45], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[10:11], s[46:47], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[12:13], s[48:49], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[14:15], s[50:51], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[16:17], s[52:53], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[18:19], s[54:55], 1.0 |
| ; GFX9-NEXT: s_branch .LBB39_5 |
| ; GFX9-NEXT: .LBB39_3: |
| ; GFX9-NEXT: s_branch .LBB39_2 |
| ; GFX9-NEXT: .LBB39_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB39_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v10f64_to_v10i64_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: s_xor_saveexec_b32 s4, -1 |
| ; GFX11-NEXT: scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill |
| ; GFX11-NEXT: s_mov_b32 exec_lo, s4 |
| ; GFX11-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX11-NEXT: s_mov_b32 s36, s0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s0, v2 |
| ; GFX11-NEXT: s_mov_b32 s47, s23 |
| ; GFX11-NEXT: s_mov_b32 s46, s22 |
| ; GFX11-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX11-NEXT: s_mov_b32 s45, s21 |
| ; GFX11-NEXT: s_mov_b32 s44, s20 |
| ; GFX11-NEXT: s_mov_b32 s43, s19 |
| ; GFX11-NEXT: s_mov_b32 s42, s18 |
| ; GFX11-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX11-NEXT: s_mov_b32 s41, s17 |
| ; GFX11-NEXT: s_mov_b32 s40, s16 |
| ; GFX11-NEXT: s_mov_b32 s38, s2 |
| ; GFX11-NEXT: s_mov_b32 s37, s1 |
| ; GFX11-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX11-NEXT: s_mov_b32 s39, s3 |
| ; GFX11-NEXT: s_cmp_lg_u32 s0, 0 |
| ; GFX11-NEXT: s_mov_b32 s0, 0 |
| ; GFX11-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX11-NEXT: s_mov_b32 s48, s24 |
| ; GFX11-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX11-NEXT: s_mov_b32 s49, s25 |
| ; GFX11-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX11-NEXT: s_mov_b32 s50, s26 |
| ; GFX11-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX11-NEXT: s_mov_b32 s51, s27 |
| ; GFX11-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX11-NEXT: s_mov_b32 s52, s28 |
| ; GFX11-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX11-NEXT: s_mov_b32 s53, s29 |
| ; GFX11-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX11-NEXT: v_readfirstlane_b32 s54, v0 |
| ; GFX11-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX11-NEXT: v_readfirstlane_b32 s55, v1 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB39_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB39_4 |
| ; GFX11-NEXT: .LBB39_2: ; %cmp.true |
| ; GFX11-NEXT: v_add_f64 v[0:1], s[36:37], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[2:3], s[38:39], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[4:5], s[40:41], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[6:7], s[42:43], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[8:9], s[44:45], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[10:11], s[46:47], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[12:13], s[48:49], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[14:15], s[50:51], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[16:17], s[52:53], 1.0 |
| ; GFX11-NEXT: v_add_f64 v[18:19], s[54:55], 1.0 |
| ; GFX11-NEXT: s_branch .LBB39_5 |
| ; GFX11-NEXT: .LBB39_3: |
| ; GFX11-NEXT: s_branch .LBB39_2 |
| ; GFX11-NEXT: .LBB39_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67 |
| ; GFX11-NEXT: .LBB39_5: ; %end |
| ; GFX11-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX11-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX11-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX11-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX11-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX11-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX11-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX11-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX11-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX11-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX11-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX11-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX11-NEXT: s_xor_saveexec_b32 s0, -1 |
| ; GFX11-NEXT: scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload |
| ; GFX11-NEXT: s_mov_b32 exec_lo, s0 |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) |
| ; 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 <10 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <10 x double> %a1 to <10 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x double> %a to <10 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <10 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x i64> %phi |
| } |
| |
| define <40 x i16> @bitcast_v10i64_to_v40i16(<10 x i64> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v10i64_to_v40i16: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr31 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr28 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr26 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr25 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr24 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr23 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; implicit-def: $vgpr22 |
| ; SI-NEXT: ; implicit-def: $vgpr30 |
| ; SI-NEXT: ; implicit-def: $vgpr21 |
| ; SI-NEXT: ; implicit-def: $vgpr29 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr27 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB40_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v28, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v31, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v34, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB40_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB40_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; SI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; SI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; SI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; SI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; SI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; SI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; SI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; SI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; SI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; SI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v28, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v31, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v34, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB40_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v34 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v31 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v28 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v25 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v35 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v23 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v32 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v22 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v30 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v29 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v27 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v34 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v31 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v28 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v26 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v25 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v24 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10i64_to_v40i16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; 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: 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 .LBB40_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB40_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB40_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18 |
| ; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc |
| ; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16 |
| ; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc |
| ; 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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB40_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10i64_to_v40i16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; 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 .LBB40_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB40_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB40_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, 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: 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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB40_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_mov_b32 s4, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v39, v0, s4 |
| ; GFX9-NEXT: v_perm_b32 v1, v38, v1, s4 |
| ; GFX9-NEXT: v_perm_b32 v2, v37, v2, s4 |
| ; GFX9-NEXT: v_perm_b32 v3, v36, v3, s4 |
| ; GFX9-NEXT: v_perm_b32 v4, v35, v4, s4 |
| ; GFX9-NEXT: v_perm_b32 v5, v34, v5, s4 |
| ; GFX9-NEXT: v_perm_b32 v6, v33, v6, s4 |
| ; GFX9-NEXT: v_perm_b32 v7, v32, v7, s4 |
| ; GFX9-NEXT: v_perm_b32 v8, v31, v8, s4 |
| ; GFX9-NEXT: v_perm_b32 v9, v30, v9, s4 |
| ; GFX9-NEXT: v_perm_b32 v10, v29, v10, s4 |
| ; GFX9-NEXT: v_perm_b32 v11, v28, v11, s4 |
| ; GFX9-NEXT: v_perm_b32 v12, v27, v12, s4 |
| ; GFX9-NEXT: v_perm_b32 v13, v26, v13, s4 |
| ; GFX9-NEXT: v_perm_b32 v14, v25, v14, s4 |
| ; GFX9-NEXT: v_perm_b32 v15, v24, v15, s4 |
| ; GFX9-NEXT: v_perm_b32 v16, v23, v16, s4 |
| ; GFX9-NEXT: v_perm_b32 v17, v22, v17, s4 |
| ; GFX9-NEXT: v_perm_b32 v18, v21, v18, s4 |
| ; GFX9-NEXT: v_perm_b32 v19, v20, v19, s4 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v10i64_to_v40i16: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB40_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3 |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3 |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3 |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3 |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3 |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo |
| ; GFX11-TRUE16-NEXT: .LBB40_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v10i64_to_v40i16: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB40_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB40_2: ; %Flow |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB40_4 |
| ; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3 |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3 |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3 |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3 |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3 |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB40_4: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v39, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v37, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v36, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v35, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v34, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v33, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v32, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v31, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v30, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v29, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v28, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v27, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v26, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v25, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v24, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v23, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v22, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v21, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v20, v19, 0x5040100 |
| ; GFX11-FAKE16-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 <10 x i64> %a, splat (i64 3) |
| %a2 = bitcast <10 x i64> %a1 to <40 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x i64> %a to <40 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <40 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x i16> %phi |
| } |
| |
| define inreg <40 x i16> @bitcast_v10i64_to_v40i16_scalar(<10 x i64> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v10i64_to_v40i16_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s8, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s5, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s8, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB41_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_lshr_b32 s72, s5, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s17, 16 |
| ; SI-NEXT: s_lshr_b64 s[10:11], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b64 s[12:13], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[14:15], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[40:41], s[28:29], 16 |
| ; SI-NEXT: s_lshr_b64 s[42:43], s[26:27], 16 |
| ; SI-NEXT: s_lshr_b64 s[44:45], s[24:25], 16 |
| ; SI-NEXT: s_lshr_b64 s[46:47], s[22:23], 16 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[20:21], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[18:19], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[16:17], 16 |
| ; SI-NEXT: s_cbranch_execnz .LBB41_3 |
| ; SI-NEXT: .LBB41_2: ; %cmp.true |
| ; SI-NEXT: s_add_u32 s4, s4, 3 |
| ; SI-NEXT: s_addc_u32 s5, s5, 0 |
| ; SI-NEXT: s_add_u32 s6, s6, 3 |
| ; SI-NEXT: s_addc_u32 s7, s7, 0 |
| ; SI-NEXT: s_add_u32 s8, s8, 3 |
| ; SI-NEXT: s_addc_u32 s9, s9, 0 |
| ; SI-NEXT: s_add_u32 s28, s28, 3 |
| ; SI-NEXT: s_addc_u32 s29, s29, 0 |
| ; SI-NEXT: s_add_u32 s26, s26, 3 |
| ; SI-NEXT: s_addc_u32 s27, s27, 0 |
| ; SI-NEXT: s_add_u32 s24, s24, 3 |
| ; SI-NEXT: s_addc_u32 s25, s25, 0 |
| ; SI-NEXT: s_add_u32 s22, s22, 3 |
| ; SI-NEXT: s_addc_u32 s23, s23, 0 |
| ; SI-NEXT: s_add_u32 s20, s20, 3 |
| ; SI-NEXT: s_addc_u32 s21, s21, 0 |
| ; SI-NEXT: s_add_u32 s18, s18, 3 |
| ; SI-NEXT: s_addc_u32 s19, s19, 0 |
| ; SI-NEXT: s_add_u32 s16, s16, 3 |
| ; SI-NEXT: s_addc_u32 s17, s17, 0 |
| ; SI-NEXT: s_lshr_b32 s72, s5, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s17, 16 |
| ; SI-NEXT: s_lshr_b64 s[10:11], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b64 s[12:13], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[14:15], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[40:41], s[28:29], 16 |
| ; SI-NEXT: s_lshr_b64 s[42:43], s[26:27], 16 |
| ; SI-NEXT: s_lshr_b64 s[44:45], s[24:25], 16 |
| ; SI-NEXT: s_lshr_b64 s[46:47], s[22:23], 16 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[20:21], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[18:19], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[16:17], 16 |
| ; SI-NEXT: .LBB41_3: ; %end |
| ; SI-NEXT: s_lshl_b32 s11, s60, 16 |
| ; SI-NEXT: s_and_b32 s13, s16, 0xffff |
| ; SI-NEXT: s_or_b32 s11, s13, s11 |
| ; SI-NEXT: s_and_b32 s13, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s15, s89, 16 |
| ; SI-NEXT: s_or_b32 s13, s13, s15 |
| ; SI-NEXT: s_lshl_b32 s15, s58, 16 |
| ; SI-NEXT: s_and_b32 s16, s18, 0xffff |
| ; SI-NEXT: s_or_b32 s15, s16, s15 |
| ; SI-NEXT: s_and_b32 s16, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s17, s88, 16 |
| ; SI-NEXT: s_or_b32 s16, s16, s17 |
| ; SI-NEXT: s_lshl_b32 s17, s56, 16 |
| ; SI-NEXT: s_and_b32 s18, s20, 0xffff |
| ; SI-NEXT: s_or_b32 s17, s18, s17 |
| ; SI-NEXT: s_and_b32 s18, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s19, s79, 16 |
| ; SI-NEXT: s_or_b32 s18, s18, s19 |
| ; SI-NEXT: s_and_b32 s19, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s20, s46, 16 |
| ; SI-NEXT: s_or_b32 s19, s19, s20 |
| ; SI-NEXT: s_and_b32 s20, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s21, s78, 16 |
| ; SI-NEXT: s_or_b32 s20, s20, s21 |
| ; SI-NEXT: s_and_b32 s21, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s22, s44, 16 |
| ; SI-NEXT: s_or_b32 s21, s21, s22 |
| ; SI-NEXT: s_and_b32 s22, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s23, s77, 16 |
| ; SI-NEXT: s_or_b32 s22, s22, s23 |
| ; SI-NEXT: s_and_b32 s23, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s42, 16 |
| ; SI-NEXT: s_or_b32 s23, s23, s24 |
| ; SI-NEXT: s_and_b32 s24, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s25, s76, 16 |
| ; SI-NEXT: s_or_b32 s24, s24, s25 |
| ; SI-NEXT: s_and_b32 s25, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s26, s40, 16 |
| ; SI-NEXT: s_and_b32 s8, s8, 0xffff |
| ; SI-NEXT: s_lshl_b32 s14, s14, 16 |
| ; SI-NEXT: s_and_b32 s6, s6, 0xffff |
| ; SI-NEXT: s_lshl_b32 s12, s12, 16 |
| ; SI-NEXT: s_and_b32 s4, s4, 0xffff |
| ; SI-NEXT: s_lshl_b32 s10, s10, 16 |
| ; SI-NEXT: s_or_b32 s25, s25, s26 |
| ; SI-NEXT: s_and_b32 s26, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s27, s75, 16 |
| ; SI-NEXT: s_or_b32 s8, s8, s14 |
| ; SI-NEXT: s_and_b32 s9, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s14, s74, 16 |
| ; SI-NEXT: s_or_b32 s6, s6, s12 |
| ; SI-NEXT: s_and_b32 s7, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s12, s73, 16 |
| ; SI-NEXT: s_or_b32 s4, s4, s10 |
| ; SI-NEXT: s_and_b32 s5, s5, 0xffff |
| ; SI-NEXT: s_lshl_b32 s10, s72, 16 |
| ; SI-NEXT: s_or_b32 s26, s26, s27 |
| ; SI-NEXT: s_or_b32 s9, s9, s14 |
| ; SI-NEXT: s_or_b32 s7, s7, s12 |
| ; SI-NEXT: s_or_b32 s5, s5, s10 |
| ; SI-NEXT: v_mov_b32_e32 v0, s11 |
| ; SI-NEXT: v_mov_b32_e32 v1, s13 |
| ; SI-NEXT: v_mov_b32_e32 v2, s15 |
| ; SI-NEXT: v_mov_b32_e32 v3, s16 |
| ; SI-NEXT: v_mov_b32_e32 v4, s17 |
| ; SI-NEXT: v_mov_b32_e32 v5, s18 |
| ; SI-NEXT: v_mov_b32_e32 v6, s19 |
| ; SI-NEXT: v_mov_b32_e32 v7, s20 |
| ; SI-NEXT: v_mov_b32_e32 v8, s21 |
| ; SI-NEXT: v_mov_b32_e32 v9, s22 |
| ; SI-NEXT: v_mov_b32_e32 v10, s23 |
| ; SI-NEXT: v_mov_b32_e32 v11, s24 |
| ; SI-NEXT: v_mov_b32_e32 v12, s25 |
| ; SI-NEXT: v_mov_b32_e32 v13, s26 |
| ; SI-NEXT: v_mov_b32_e32 v14, s8 |
| ; SI-NEXT: v_mov_b32_e32 v15, s9 |
| ; SI-NEXT: v_mov_b32_e32 v16, s6 |
| ; SI-NEXT: v_mov_b32_e32 v17, s7 |
| ; SI-NEXT: v_mov_b32_e32 v18, s4 |
| ; SI-NEXT: v_mov_b32_e32 v19, s5 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB41_4: |
| ; SI-NEXT: ; implicit-def: $sgpr60 |
| ; SI-NEXT: ; implicit-def: $sgpr89 |
| ; SI-NEXT: ; implicit-def: $sgpr58 |
| ; SI-NEXT: ; implicit-def: $sgpr88 |
| ; SI-NEXT: ; implicit-def: $sgpr56 |
| ; SI-NEXT: ; implicit-def: $sgpr79 |
| ; SI-NEXT: ; implicit-def: $sgpr46 |
| ; SI-NEXT: ; implicit-def: $sgpr78 |
| ; SI-NEXT: ; implicit-def: $sgpr44 |
| ; SI-NEXT: ; implicit-def: $sgpr77 |
| ; SI-NEXT: ; implicit-def: $sgpr42 |
| ; SI-NEXT: ; implicit-def: $sgpr76 |
| ; SI-NEXT: ; implicit-def: $sgpr40 |
| ; SI-NEXT: ; implicit-def: $sgpr75 |
| ; SI-NEXT: ; implicit-def: $sgpr14 |
| ; SI-NEXT: ; implicit-def: $sgpr74 |
| ; SI-NEXT: ; implicit-def: $sgpr12 |
| ; SI-NEXT: ; implicit-def: $sgpr73 |
| ; SI-NEXT: ; implicit-def: $sgpr10 |
| ; SI-NEXT: ; implicit-def: $sgpr72 |
| ; SI-NEXT: s_branch .LBB41_2 |
| ; |
| ; VI-LABEL: bitcast_v10i64_to_v40i16_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB41_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_lshr_b32 s12, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s10, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s46, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: s_cbranch_execnz .LBB41_3 |
| ; VI-NEXT: .LBB41_2: ; %cmp.true |
| ; VI-NEXT: s_add_u32 s7, s7, 3 |
| ; VI-NEXT: s_addc_u32 s6, s6, 0 |
| ; VI-NEXT: s_add_u32 s9, s9, 3 |
| ; VI-NEXT: s_addc_u32 s8, s8, 0 |
| ; VI-NEXT: s_add_u32 s11, s11, 3 |
| ; VI-NEXT: s_addc_u32 s10, s10, 0 |
| ; VI-NEXT: s_add_u32 s28, s28, 3 |
| ; VI-NEXT: s_addc_u32 s29, s29, 0 |
| ; VI-NEXT: s_add_u32 s26, s26, 3 |
| ; VI-NEXT: s_addc_u32 s27, s27, 0 |
| ; VI-NEXT: s_add_u32 s24, s24, 3 |
| ; VI-NEXT: s_addc_u32 s25, s25, 0 |
| ; VI-NEXT: s_add_u32 s22, s22, 3 |
| ; VI-NEXT: s_addc_u32 s23, s23, 0 |
| ; VI-NEXT: s_add_u32 s20, s20, 3 |
| ; VI-NEXT: s_addc_u32 s21, s21, 0 |
| ; VI-NEXT: s_add_u32 s18, s18, 3 |
| ; VI-NEXT: s_addc_u32 s19, s19, 0 |
| ; VI-NEXT: s_add_u32 s16, s16, 3 |
| ; VI-NEXT: s_addc_u32 s17, s17, 0 |
| ; VI-NEXT: s_lshr_b32 s12, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s10, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s46, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: .LBB41_3: ; %end |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s63, 16 |
| ; VI-NEXT: s_or_b32 s4, s4, s5 |
| ; VI-NEXT: s_and_b32 s5, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s16, s62, 16 |
| ; VI-NEXT: s_or_b32 s5, s5, s16 |
| ; VI-NEXT: s_and_b32 s16, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s17, s61, 16 |
| ; VI-NEXT: s_or_b32 s16, s16, s17 |
| ; VI-NEXT: s_and_b32 s17, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s18, s60, 16 |
| ; VI-NEXT: s_or_b32 s17, s17, s18 |
| ; VI-NEXT: s_and_b32 s18, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s19, s59, 16 |
| ; VI-NEXT: s_or_b32 s18, s18, s19 |
| ; VI-NEXT: s_and_b32 s19, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s20, s58, 16 |
| ; VI-NEXT: s_or_b32 s19, s19, s20 |
| ; VI-NEXT: s_and_b32 s20, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s21, s57, 16 |
| ; VI-NEXT: s_or_b32 s20, s20, s21 |
| ; VI-NEXT: s_and_b32 s21, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s22, s56, 16 |
| ; VI-NEXT: s_or_b32 s21, s21, s22 |
| ; VI-NEXT: s_and_b32 s22, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s23, s47, 16 |
| ; VI-NEXT: s_or_b32 s22, s22, s23 |
| ; VI-NEXT: s_and_b32 s23, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s24, s46, 16 |
| ; VI-NEXT: s_or_b32 s23, s23, s24 |
| ; VI-NEXT: s_and_b32 s24, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s25, s45, 16 |
| ; VI-NEXT: s_or_b32 s24, s24, s25 |
| ; VI-NEXT: s_and_b32 s25, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s26, s44, 16 |
| ; VI-NEXT: s_or_b32 s25, s25, s26 |
| ; VI-NEXT: s_and_b32 s26, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s27, s43, 16 |
| ; VI-NEXT: s_or_b32 s26, s26, s27 |
| ; VI-NEXT: s_and_b32 s27, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s28, s42, 16 |
| ; VI-NEXT: s_or_b32 s27, s27, s28 |
| ; VI-NEXT: s_and_b32 s11, 0xffff, s11 |
| ; VI-NEXT: s_lshl_b32 s28, s41, 16 |
| ; VI-NEXT: s_or_b32 s11, s11, s28 |
| ; VI-NEXT: s_and_b32 s10, 0xffff, s10 |
| ; VI-NEXT: s_lshl_b32 s28, s40, 16 |
| ; VI-NEXT: s_and_b32 s9, 0xffff, s9 |
| ; VI-NEXT: s_lshl_b32 s15, s15, 16 |
| ; VI-NEXT: s_and_b32 s8, 0xffff, s8 |
| ; VI-NEXT: s_lshl_b32 s14, s14, 16 |
| ; VI-NEXT: s_and_b32 s7, 0xffff, s7 |
| ; VI-NEXT: s_lshl_b32 s13, s13, 16 |
| ; VI-NEXT: s_and_b32 s6, 0xffff, s6 |
| ; VI-NEXT: s_lshl_b32 s12, s12, 16 |
| ; VI-NEXT: s_or_b32 s10, s10, s28 |
| ; VI-NEXT: s_or_b32 s9, s9, s15 |
| ; VI-NEXT: s_or_b32 s8, s8, s14 |
| ; VI-NEXT: s_or_b32 s7, s7, s13 |
| ; VI-NEXT: s_or_b32 s6, s6, s12 |
| ; VI-NEXT: v_mov_b32_e32 v0, s4 |
| ; VI-NEXT: v_mov_b32_e32 v1, s5 |
| ; VI-NEXT: v_mov_b32_e32 v2, s16 |
| ; VI-NEXT: v_mov_b32_e32 v3, s17 |
| ; VI-NEXT: v_mov_b32_e32 v4, s18 |
| ; VI-NEXT: v_mov_b32_e32 v5, s19 |
| ; VI-NEXT: v_mov_b32_e32 v6, s20 |
| ; VI-NEXT: v_mov_b32_e32 v7, s21 |
| ; VI-NEXT: v_mov_b32_e32 v8, s22 |
| ; VI-NEXT: v_mov_b32_e32 v9, s23 |
| ; VI-NEXT: v_mov_b32_e32 v10, s24 |
| ; VI-NEXT: v_mov_b32_e32 v11, s25 |
| ; VI-NEXT: v_mov_b32_e32 v12, s26 |
| ; VI-NEXT: v_mov_b32_e32 v13, s27 |
| ; VI-NEXT: v_mov_b32_e32 v14, s11 |
| ; VI-NEXT: v_mov_b32_e32 v15, s10 |
| ; VI-NEXT: v_mov_b32_e32 v16, s9 |
| ; VI-NEXT: v_mov_b32_e32 v17, s8 |
| ; VI-NEXT: v_mov_b32_e32 v18, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB41_4: |
| ; VI-NEXT: ; implicit-def: $sgpr63 |
| ; VI-NEXT: ; implicit-def: $sgpr62 |
| ; VI-NEXT: ; implicit-def: $sgpr61 |
| ; VI-NEXT: ; implicit-def: $sgpr60 |
| ; VI-NEXT: ; implicit-def: $sgpr59 |
| ; VI-NEXT: ; implicit-def: $sgpr58 |
| ; VI-NEXT: ; implicit-def: $sgpr57 |
| ; VI-NEXT: ; implicit-def: $sgpr56 |
| ; VI-NEXT: ; implicit-def: $sgpr47 |
| ; VI-NEXT: ; implicit-def: $sgpr46 |
| ; VI-NEXT: ; implicit-def: $sgpr45 |
| ; VI-NEXT: ; implicit-def: $sgpr44 |
| ; VI-NEXT: ; implicit-def: $sgpr43 |
| ; VI-NEXT: ; implicit-def: $sgpr42 |
| ; VI-NEXT: ; implicit-def: $sgpr41 |
| ; VI-NEXT: ; implicit-def: $sgpr40 |
| ; VI-NEXT: ; implicit-def: $sgpr15 |
| ; VI-NEXT: ; implicit-def: $sgpr14 |
| ; VI-NEXT: ; implicit-def: $sgpr13 |
| ; VI-NEXT: ; implicit-def: $sgpr12 |
| ; VI-NEXT: s_branch .LBB41_2 |
| ; |
| ; GFX9-LABEL: bitcast_v10i64_to_v40i16_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s10, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s11, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB41_4 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_lshr_b32 s12, s6, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s7, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s8, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s9, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s10, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s11, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s43, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s58, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s60, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s62, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s16, 16 |
| ; GFX9-NEXT: s_cbranch_execnz .LBB41_3 |
| ; GFX9-NEXT: .LBB41_2: ; %cmp.true |
| ; GFX9-NEXT: s_add_u32 s7, s7, 3 |
| ; GFX9-NEXT: s_addc_u32 s6, s6, 0 |
| ; GFX9-NEXT: s_add_u32 s9, s9, 3 |
| ; GFX9-NEXT: s_addc_u32 s8, s8, 0 |
| ; GFX9-NEXT: s_add_u32 s11, s11, 3 |
| ; GFX9-NEXT: s_addc_u32 s10, s10, 0 |
| ; GFX9-NEXT: s_add_u32 s28, s28, 3 |
| ; GFX9-NEXT: s_addc_u32 s29, s29, 0 |
| ; GFX9-NEXT: s_add_u32 s26, s26, 3 |
| ; GFX9-NEXT: s_addc_u32 s27, s27, 0 |
| ; GFX9-NEXT: s_add_u32 s24, s24, 3 |
| ; GFX9-NEXT: s_addc_u32 s25, s25, 0 |
| ; GFX9-NEXT: s_add_u32 s22, s22, 3 |
| ; GFX9-NEXT: s_addc_u32 s23, s23, 0 |
| ; GFX9-NEXT: s_add_u32 s20, s20, 3 |
| ; GFX9-NEXT: s_addc_u32 s21, s21, 0 |
| ; GFX9-NEXT: s_add_u32 s18, s18, 3 |
| ; GFX9-NEXT: s_addc_u32 s19, s19, 0 |
| ; GFX9-NEXT: s_add_u32 s16, s16, 3 |
| ; GFX9-NEXT: s_addc_u32 s17, s17, 0 |
| ; GFX9-NEXT: s_lshr_b32 s12, s6, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s7, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s8, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s9, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s10, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s11, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s43, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s58, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s60, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s62, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s16, 16 |
| ; GFX9-NEXT: .LBB41_3: ; %end |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s63 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s5, s17, s62 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s61 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s60 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s59 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s58 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s57 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s56 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s47 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s46 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s45 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s44 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s43 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s42 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s11, s11, s41 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s10, s10, s40 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s9, s9, s15 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s14 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s13 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s12 |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s5 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s6 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; GFX9-NEXT: .LBB41_4: |
| ; GFX9-NEXT: ; implicit-def: $sgpr63 |
| ; GFX9-NEXT: ; implicit-def: $sgpr62 |
| ; GFX9-NEXT: ; implicit-def: $sgpr61 |
| ; GFX9-NEXT: ; implicit-def: $sgpr60 |
| ; GFX9-NEXT: ; implicit-def: $sgpr59 |
| ; GFX9-NEXT: ; implicit-def: $sgpr58 |
| ; GFX9-NEXT: ; implicit-def: $sgpr57 |
| ; GFX9-NEXT: ; implicit-def: $sgpr56 |
| ; GFX9-NEXT: ; implicit-def: $sgpr47 |
| ; GFX9-NEXT: ; implicit-def: $sgpr46 |
| ; GFX9-NEXT: ; implicit-def: $sgpr45 |
| ; GFX9-NEXT: ; implicit-def: $sgpr44 |
| ; GFX9-NEXT: ; implicit-def: $sgpr43 |
| ; GFX9-NEXT: ; implicit-def: $sgpr42 |
| ; GFX9-NEXT: ; implicit-def: $sgpr41 |
| ; GFX9-NEXT: ; implicit-def: $sgpr40 |
| ; GFX9-NEXT: ; implicit-def: $sgpr15 |
| ; GFX9-NEXT: ; implicit-def: $sgpr14 |
| ; GFX9-NEXT: ; implicit-def: $sgpr13 |
| ; GFX9-NEXT: ; implicit-def: $sgpr12 |
| ; GFX9-NEXT: s_branch .LBB41_2 |
| ; |
| ; GFX11-LABEL: bitcast_v10i64_to_v40i16_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-NEXT: s_mov_b32 s58, 0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB41_4 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-NEXT: s_lshr_b32 s6, s4, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s5, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s40, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s41, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s45, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s46, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s56, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s58 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB41_3 |
| ; GFX11-NEXT: .LBB41_2: ; %cmp.true |
| ; GFX11-NEXT: s_add_u32 s5, s5, 3 |
| ; GFX11-NEXT: s_addc_u32 s4, s4, 0 |
| ; GFX11-NEXT: s_add_u32 s28, s28, 3 |
| ; GFX11-NEXT: s_addc_u32 s29, s29, 0 |
| ; GFX11-NEXT: s_add_u32 s26, s26, 3 |
| ; GFX11-NEXT: s_addc_u32 s27, s27, 0 |
| ; GFX11-NEXT: s_add_u32 s24, s24, 3 |
| ; GFX11-NEXT: s_addc_u32 s25, s25, 0 |
| ; GFX11-NEXT: s_add_u32 s22, s22, 3 |
| ; GFX11-NEXT: s_addc_u32 s23, s23, 0 |
| ; GFX11-NEXT: s_add_u32 s20, s20, 3 |
| ; GFX11-NEXT: s_addc_u32 s21, s21, 0 |
| ; GFX11-NEXT: s_add_u32 s18, s18, 3 |
| ; GFX11-NEXT: s_addc_u32 s19, s19, 0 |
| ; GFX11-NEXT: s_add_u32 s16, s16, 3 |
| ; GFX11-NEXT: s_addc_u32 s17, s17, 0 |
| ; GFX11-NEXT: s_add_u32 s2, s2, 3 |
| ; GFX11-NEXT: s_addc_u32 s3, s3, 0 |
| ; GFX11-NEXT: s_add_u32 s0, s0, 3 |
| ; GFX11-NEXT: s_addc_u32 s1, s1, 0 |
| ; GFX11-NEXT: s_lshr_b32 s6, s4, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s5, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s40, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s41, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s45, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s46, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s56, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: .LBB41_3: ; %end |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s57 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s56 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s47 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s46 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s45 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s44 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s43 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s42 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s41 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s40 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s15, s22, s15 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s14, s23, s14 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s13, s24, s13 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s12, s25, s12 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s11, s26, s11 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s10, s27, s10 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s9, s28, s9 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s8, s29, s8 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s7 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s6 |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB41_4: |
| ; GFX11-NEXT: ; implicit-def: $sgpr57 |
| ; GFX11-NEXT: ; implicit-def: $sgpr56 |
| ; GFX11-NEXT: ; implicit-def: $sgpr47 |
| ; GFX11-NEXT: ; implicit-def: $sgpr46 |
| ; GFX11-NEXT: ; implicit-def: $sgpr45 |
| ; GFX11-NEXT: ; implicit-def: $sgpr44 |
| ; GFX11-NEXT: ; implicit-def: $sgpr43 |
| ; GFX11-NEXT: ; implicit-def: $sgpr42 |
| ; GFX11-NEXT: ; implicit-def: $sgpr41 |
| ; GFX11-NEXT: ; implicit-def: $sgpr40 |
| ; GFX11-NEXT: ; implicit-def: $sgpr15 |
| ; GFX11-NEXT: ; implicit-def: $sgpr14 |
| ; GFX11-NEXT: ; implicit-def: $sgpr13 |
| ; GFX11-NEXT: ; implicit-def: $sgpr12 |
| ; GFX11-NEXT: ; implicit-def: $sgpr11 |
| ; GFX11-NEXT: ; implicit-def: $sgpr10 |
| ; GFX11-NEXT: ; implicit-def: $sgpr9 |
| ; GFX11-NEXT: ; implicit-def: $sgpr8 |
| ; GFX11-NEXT: ; implicit-def: $sgpr7 |
| ; GFX11-NEXT: ; implicit-def: $sgpr6 |
| ; GFX11-NEXT: s_branch .LBB41_2 |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <10 x i64> %a, splat (i64 3) |
| %a2 = bitcast <10 x i64> %a1 to <40 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x i64> %a to <40 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <40 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x i16> %phi |
| } |
| |
| define <10 x i64> @bitcast_v40i16_to_v10i64(<40 x i16> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v40i16_to_v10i64: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: v_mov_b32_e32 v32, v19 |
| ; SI-NEXT: v_mov_b32_e32 v33, v18 |
| ; SI-NEXT: v_mov_b32_e32 v34, v17 |
| ; SI-NEXT: v_mov_b32_e32 v35, v16 |
| ; SI-NEXT: v_mov_b32_e32 v36, v15 |
| ; SI-NEXT: v_mov_b32_e32 v37, v14 |
| ; SI-NEXT: v_mov_b32_e32 v38, v13 |
| ; SI-NEXT: v_mov_b32_e32 v39, v12 |
| ; SI-NEXT: v_mov_b32_e32 v48, v11 |
| ; SI-NEXT: v_mov_b32_e32 v49, v10 |
| ; SI-NEXT: v_mov_b32_e32 v50, v9 |
| ; SI-NEXT: v_mov_b32_e32 v51, v8 |
| ; SI-NEXT: v_mov_b32_e32 v52, v7 |
| ; SI-NEXT: v_mov_b32_e32 v53, v6 |
| ; SI-NEXT: v_mov_b32_e32 v54, v5 |
| ; SI-NEXT: v_mov_b32_e32 v55, v4 |
| ; SI-NEXT: v_mov_b32_e32 v40, v3 |
| ; SI-NEXT: v_mov_b32_e32 v41, v2 |
| ; SI-NEXT: v_mov_b32_e32 v42, v1 |
| ; SI-NEXT: v_mov_b32_e32 v43, v0 |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v33 |
| ; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v34 |
| ; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v35 |
| ; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v36 |
| ; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v37 |
| ; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v38 |
| ; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v39 |
| ; SI-NEXT: v_lshrrev_b32_e32 v8, 16, v48 |
| ; SI-NEXT: v_lshrrev_b32_e32 v9, 16, v49 |
| ; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v50 |
| ; SI-NEXT: v_lshrrev_b32_e32 v11, 16, v51 |
| ; SI-NEXT: v_lshrrev_b32_e32 v12, 16, v52 |
| ; SI-NEXT: v_lshrrev_b32_e32 v13, 16, v53 |
| ; SI-NEXT: v_lshrrev_b32_e32 v14, 16, v54 |
| ; SI-NEXT: v_lshrrev_b32_e32 v15, 16, v55 |
| ; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v40 |
| ; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v41 |
| ; SI-NEXT: v_lshrrev_b32_e32 v18, 16, v42 |
| ; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v43 |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_waitcnt expcnt(4) |
| ; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v56, 16, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v47, 16, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v12 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB42_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v39 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v43 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v42 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v41 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v40 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v55 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v54 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v53 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v52 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v51 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v50 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v49 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v48 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v59 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v58 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v57 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v56 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v47 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v46 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v45 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v44 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v63 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v62 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v61 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v60 |
| ; SI-NEXT: ; implicit-def: $vgpr43 |
| ; SI-NEXT: ; implicit-def: $vgpr42 |
| ; SI-NEXT: ; implicit-def: $vgpr41 |
| ; SI-NEXT: ; implicit-def: $vgpr40 |
| ; SI-NEXT: ; implicit-def: $vgpr55 |
| ; SI-NEXT: ; implicit-def: $vgpr54 |
| ; SI-NEXT: ; implicit-def: $vgpr53 |
| ; SI-NEXT: ; implicit-def: $vgpr52 |
| ; SI-NEXT: ; implicit-def: $vgpr51 |
| ; SI-NEXT: ; implicit-def: $vgpr50 |
| ; SI-NEXT: ; implicit-def: $vgpr49 |
| ; SI-NEXT: ; implicit-def: $vgpr48 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr59 |
| ; SI-NEXT: ; implicit-def: $vgpr58 |
| ; SI-NEXT: ; implicit-def: $vgpr57 |
| ; SI-NEXT: ; implicit-def: $vgpr56 |
| ; SI-NEXT: ; implicit-def: $vgpr47 |
| ; SI-NEXT: ; implicit-def: $vgpr46 |
| ; SI-NEXT: ; implicit-def: $vgpr45 |
| ; SI-NEXT: ; implicit-def: $vgpr44 |
| ; SI-NEXT: ; implicit-def: $vgpr63 |
| ; SI-NEXT: ; implicit-def: $vgpr62 |
| ; SI-NEXT: ; implicit-def: $vgpr61 |
| ; SI-NEXT: ; implicit-def: $vgpr60 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v38 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v37 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v36 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v35 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v34 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v33 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v19 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v32 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: .LBB42_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB42_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v39 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v43 |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v42 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v41 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v40 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v55 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v54 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v53 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v52 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v51 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v50 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v49 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v48 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_or_b32_e32 v0, v59, v0 |
| ; SI-NEXT: s_mov_b32 s6, 0x30000 |
| ; SI-NEXT: v_or_b32_e32 v1, v58, v1 |
| ; SI-NEXT: v_or_b32_e32 v2, v57, v2 |
| ; SI-NEXT: v_or_b32_e32 v3, v56, v3 |
| ; SI-NEXT: v_or_b32_e32 v4, v47, v4 |
| ; SI-NEXT: v_or_b32_e32 v5, v46, v5 |
| ; SI-NEXT: v_or_b32_e32 v6, v45, v6 |
| ; SI-NEXT: v_or_b32_e32 v7, v44, v7 |
| ; SI-NEXT: v_or_b32_e32 v8, v63, v8 |
| ; SI-NEXT: v_or_b32_e32 v9, v62, v9 |
| ; SI-NEXT: v_or_b32_e32 v10, v61, v10 |
| ; SI-NEXT: v_or_b32_e32 v11, v60, v11 |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0 |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, s6, v3 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, s6, v5 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, s6, v7 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v38 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_or_b32_e32 v13, v14, v13 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v37 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_or_b32_e32 v14, v15, v14 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v36 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v35 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v34 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_or_b32_e32 v17, v18, v17 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v33 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_or_b32_e32 v18, v19, v18 |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v32 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_or_b32_e32 v19, v20, v19 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, s6, v11 |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, s6, v13 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v14 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, s6, v15 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, s6, v16 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, s6, v17 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v18 |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 0x30000, v19 |
| ; SI-NEXT: .LBB42_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40i16_to_v10i64: |
| ; 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:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: v_mov_b32_e32 v32, v19 |
| ; VI-NEXT: v_mov_b32_e32 v33, v18 |
| ; VI-NEXT: v_mov_b32_e32 v34, v17 |
| ; VI-NEXT: v_mov_b32_e32 v35, v16 |
| ; VI-NEXT: v_mov_b32_e32 v36, v15 |
| ; VI-NEXT: v_mov_b32_e32 v37, v14 |
| ; VI-NEXT: v_mov_b32_e32 v38, v13 |
| ; VI-NEXT: v_mov_b32_e32 v39, v12 |
| ; VI-NEXT: v_mov_b32_e32 v48, v11 |
| ; VI-NEXT: v_mov_b32_e32 v49, v10 |
| ; VI-NEXT: v_mov_b32_e32 v50, v9 |
| ; VI-NEXT: v_mov_b32_e32 v51, v8 |
| ; VI-NEXT: v_mov_b32_e32 v52, v7 |
| ; VI-NEXT: v_mov_b32_e32 v53, v6 |
| ; VI-NEXT: v_mov_b32_e32 v54, v5 |
| ; VI-NEXT: v_mov_b32_e32 v55, v4 |
| ; VI-NEXT: v_mov_b32_e32 v40, v3 |
| ; VI-NEXT: v_mov_b32_e32 v41, v2 |
| ; VI-NEXT: v_mov_b32_e32 v42, v1 |
| ; VI-NEXT: v_mov_b32_e32 v43, v0 |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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_mov_b32_e32 v19, 16 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v0, v19, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v1, v19, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v2, v19, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v3, v19, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v4, v19, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v5, v19, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v6, v19, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v7, v19, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v8, v19, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v9, v19, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v10, v19, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v11, v19, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v12, v19, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v13, v19, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v14, v19, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v15, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v16, v19, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v17, v19, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v18, v19, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v19, v19, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v40, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v55, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v54, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v53, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v52, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v51, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v50, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v49, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v48, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v39, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v38, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v37, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v36, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v35, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v34, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v33, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v32, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; 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: .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_mov_b32_e32 v19, 3 |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v43 |
| ; VI-NEXT: v_add_u16_sdwa v1, v43, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v42 |
| ; VI-NEXT: v_add_u16_sdwa v2, v42, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v1, v1, v2 |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v41 |
| ; VI-NEXT: v_add_u16_sdwa v3, v41, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v40 |
| ; VI-NEXT: v_add_u16_sdwa v4, v40, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v55 |
| ; VI-NEXT: v_add_u16_sdwa v5, v55, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v54 |
| ; VI-NEXT: v_add_u16_sdwa v6, v54, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v53 |
| ; VI-NEXT: v_add_u16_sdwa v7, v53, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v52 |
| ; VI-NEXT: v_add_u16_sdwa v8, v52, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v51 |
| ; VI-NEXT: v_add_u16_sdwa v9, v51, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v50 |
| ; VI-NEXT: v_add_u16_sdwa v10, v50, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v49 |
| ; VI-NEXT: v_add_u16_sdwa v11, v49, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v48 |
| ; VI-NEXT: v_add_u16_sdwa v12, v48, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v39 |
| ; VI-NEXT: v_add_u16_sdwa v13, v39, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v38 |
| ; VI-NEXT: v_add_u16_sdwa v14, v38, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v37 |
| ; VI-NEXT: v_add_u16_sdwa v15, v37, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; VI-NEXT: v_add_u16_e32 v15, 3, v36 |
| ; VI-NEXT: v_add_u16_sdwa v16, v36, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v35 |
| ; VI-NEXT: v_add_u16_sdwa v17, v35, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; VI-NEXT: v_add_u16_e32 v17, 3, v34 |
| ; VI-NEXT: v_add_u16_sdwa v18, v34, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; VI-NEXT: v_add_u16_e32 v18, 3, v33 |
| ; VI-NEXT: v_add_u16_sdwa v20, v33, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; VI-NEXT: v_add_u16_e32 v20, 3, v32 |
| ; VI-NEXT: v_add_u16_sdwa v19, v32, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v19, v20, v19 |
| ; VI-NEXT: .LBB42_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40i16_to_v10i64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_mov_b32_e32 v32, v19 |
| ; 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: v_mov_b32_e32 v33, v18 |
| ; GFX9-NEXT: v_mov_b32_e32 v43, v0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, v17 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, v16 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, v15 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, v14 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, v13 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, v12 |
| ; GFX9-NEXT: v_mov_b32_e32 v48, v11 |
| ; GFX9-NEXT: v_mov_b32_e32 v49, v10 |
| ; GFX9-NEXT: v_mov_b32_e32 v50, v9 |
| ; GFX9-NEXT: v_mov_b32_e32 v51, v8 |
| ; GFX9-NEXT: v_mov_b32_e32 v52, v7 |
| ; GFX9-NEXT: v_mov_b32_e32 v53, v6 |
| ; GFX9-NEXT: v_mov_b32_e32 v54, v5 |
| ; GFX9-NEXT: v_mov_b32_e32 v55, v4 |
| ; GFX9-NEXT: v_mov_b32_e32 v40, v3 |
| ; GFX9-NEXT: v_mov_b32_e32 v41, v2 |
| ; GFX9-NEXT: v_mov_b32_e32 v42, v1 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v48 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v49 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v50 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v51 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v52 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v53 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v54 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v55 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43 |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v35 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v34 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v33 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v32 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; 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: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; 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: buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_pk_add_u16 v0, v0, 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 v2, v2, 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 v4, v4, 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 v6, v6, 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 v8, v8, 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 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; GFX9-NEXT: v_pk_add_u16 v12, v12, 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 v14, v14, 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 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB42_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; 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-TRUE16-LABEL: bitcast_v40i16_to_v10i64: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB42_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: .LBB42_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v40i16_to_v10i64: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v37, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v39, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v48, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v36, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v35, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v34, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v33, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v32, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v31, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v30, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v29, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v28, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v27, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v26, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v25, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v24, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v23, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v22, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v21, v19, 0x5040100 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB42_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: .LBB42_2: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-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 <40 x i16> %a, splat (i16 3) |
| %a2 = bitcast <40 x i16> %a1 to <10 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x i16> %a to <10 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <10 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x i64> %phi |
| } |
| |
| define inreg <10 x i64> @bitcast_v40i16_to_v10i64_scalar(<40 x i16> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v40i16_to_v10i64_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v20, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v20, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v20, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v20, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v20, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v20, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v20, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v20, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v20, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v20, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v20, s54, 10 |
| ; SI-NEXT: v_writelane_b32 v20, s55, 11 |
| ; SI-NEXT: v_writelane_b32 v20, s64, 12 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s12, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s15, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s74, v1 |
| ; SI-NEXT: v_readfirstlane_b32 s77, v0 |
| ; SI-NEXT: v_writelane_b32 v20, s65, 13 |
| ; SI-NEXT: s_lshr_b32 s11, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s13, s28, 16 |
| ; SI-NEXT: s_lshr_b32 s72, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s26, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; SI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; SI-NEXT: s_lshr_b32 s92, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s93, s18, 16 |
| ; SI-NEXT: s_lshr_b32 s94, s17, 16 |
| ; SI-NEXT: s_lshr_b32 s95, s16, 16 |
| ; SI-NEXT: s_lshr_b32 s6, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s8, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s10, s12, 16 |
| ; SI-NEXT: s_lshr_b32 s14, s15, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s74, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s77, 16 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v20, s66, 14 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_writelane_b32 v20, s67, 15 |
| ; SI-NEXT: s_cbranch_scc0 .LBB43_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s95, 16 |
| ; SI-NEXT: s_or_b32 s36, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s94, 16 |
| ; SI-NEXT: s_or_b32 s37, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s93, 16 |
| ; SI-NEXT: s_or_b32 s38, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s92, 16 |
| ; SI-NEXT: s_or_b32 s39, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; SI-NEXT: s_or_b32 s40, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; SI-NEXT: s_or_b32 s41, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; SI-NEXT: s_or_b32 s42, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; SI-NEXT: s_or_b32 s43, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; SI-NEXT: s_or_b32 s44, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; SI-NEXT: s_or_b32 s45, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; SI-NEXT: s_or_b32 s46, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; SI-NEXT: s_or_b32 s47, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; SI-NEXT: s_or_b32 s48, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; SI-NEXT: s_or_b32 s49, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s77, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; SI-NEXT: s_or_b32 s50, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s74, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; SI-NEXT: s_or_b32 s51, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s15, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; SI-NEXT: s_or_b32 s52, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; SI-NEXT: s_or_b32 s53, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; SI-NEXT: s_or_b32 s54, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; SI-NEXT: s_or_b32 s55, s4, s5 |
| ; SI-NEXT: s_cbranch_execnz .LBB43_3 |
| ; SI-NEXT: .LBB43_2: ; %cmp.true |
| ; SI-NEXT: s_add_i32 s16, s16, 3 |
| ; SI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s95, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s17, s17, 3 |
| ; SI-NEXT: s_add_i32 s36, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s94, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s18, s18, 3 |
| ; SI-NEXT: s_add_i32 s37, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s93, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s19, s19, 3 |
| ; SI-NEXT: s_add_i32 s38, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s92, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s20, s20, 3 |
| ; SI-NEXT: s_add_i32 s39, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s21, s21, 3 |
| ; SI-NEXT: s_add_i32 s40, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s22, s22, 3 |
| ; SI-NEXT: s_add_i32 s41, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s23, s23, 3 |
| ; SI-NEXT: s_add_i32 s42, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s24, s24, 3 |
| ; SI-NEXT: s_add_i32 s43, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s25, s25, 3 |
| ; SI-NEXT: s_add_i32 s44, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s26, s26, 3 |
| ; SI-NEXT: s_add_i32 s45, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s27, s27, 3 |
| ; SI-NEXT: s_add_i32 s46, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s28, s28, 3 |
| ; SI-NEXT: s_add_i32 s47, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s29, s29, 3 |
| ; SI-NEXT: s_add_i32 s48, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s77, s77, 3 |
| ; SI-NEXT: s_add_i32 s49, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s77, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s74, s74, 3 |
| ; SI-NEXT: s_add_i32 s50, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s74, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s15, s15, 3 |
| ; SI-NEXT: s_add_i32 s51, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s15, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s12, s12, 3 |
| ; SI-NEXT: s_add_i32 s52, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s9, s9, 3 |
| ; SI-NEXT: s_add_i32 s53, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s7, s7, 3 |
| ; SI-NEXT: s_add_i32 s54, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s55, s4, 0x30000 |
| ; SI-NEXT: .LBB43_3: ; %end |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_readlane_b32 s67, v20, 15 |
| ; SI-NEXT: v_readlane_b32 s66, v20, 14 |
| ; SI-NEXT: v_readlane_b32 s65, v20, 13 |
| ; SI-NEXT: v_readlane_b32 s64, v20, 12 |
| ; SI-NEXT: v_readlane_b32 s55, v20, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v20, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v20, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v20, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v20, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v20, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v20, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v20, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v20, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v20, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v20, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v20, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB43_4: |
| ; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; SI-NEXT: s_branch .LBB43_2 |
| ; |
| ; VI-LABEL: bitcast_v40i16_to_v10i64_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v20, s30, 0 |
| ; VI-NEXT: v_writelane_b32 v20, s31, 1 |
| ; VI-NEXT: v_writelane_b32 v20, s34, 2 |
| ; VI-NEXT: v_writelane_b32 v20, s35, 3 |
| ; VI-NEXT: v_writelane_b32 v20, s36, 4 |
| ; VI-NEXT: v_writelane_b32 v20, s37, 5 |
| ; VI-NEXT: v_writelane_b32 v20, s38, 6 |
| ; VI-NEXT: v_writelane_b32 v20, s39, 7 |
| ; VI-NEXT: v_writelane_b32 v20, s48, 8 |
| ; VI-NEXT: v_writelane_b32 v20, s49, 9 |
| ; VI-NEXT: v_writelane_b32 v20, s50, 10 |
| ; VI-NEXT: v_writelane_b32 v20, s51, 11 |
| ; VI-NEXT: v_writelane_b32 v20, s52, 12 |
| ; VI-NEXT: v_writelane_b32 v20, s53, 13 |
| ; VI-NEXT: v_writelane_b32 v20, s54, 14 |
| ; VI-NEXT: v_writelane_b32 v20, s55, 15 |
| ; VI-NEXT: v_writelane_b32 v20, s64, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s12, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s15, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s74, v1 |
| ; VI-NEXT: v_readfirstlane_b32 s77, v0 |
| ; VI-NEXT: v_writelane_b32 v20, s65, 17 |
| ; VI-NEXT: s_lshr_b32 s11, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s72, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s75, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s78, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s30, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s31, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s34, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s35, s16, 16 |
| ; VI-NEXT: s_lshr_b32 s6, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s8, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s10, s12, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s15, 16 |
| ; VI-NEXT: s_lshr_b32 s73, s74, 16 |
| ; VI-NEXT: s_lshr_b32 s76, s77, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v20, s66, 18 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_writelane_b32 v20, s67, 19 |
| ; VI-NEXT: s_cbranch_scc0 .LBB43_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s35, 16 |
| ; VI-NEXT: s_or_b32 s36, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s5, s34, 16 |
| ; VI-NEXT: s_or_b32 s37, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s5, s31, 16 |
| ; VI-NEXT: s_or_b32 s38, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s5, s30, 16 |
| ; VI-NEXT: s_or_b32 s39, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; VI-NEXT: s_or_b32 s40, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; VI-NEXT: s_or_b32 s41, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; VI-NEXT: s_or_b32 s42, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; VI-NEXT: s_or_b32 s43, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; VI-NEXT: s_or_b32 s44, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; VI-NEXT: s_or_b32 s45, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; VI-NEXT: s_or_b32 s46, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; VI-NEXT: s_or_b32 s47, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; VI-NEXT: s_or_b32 s48, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; VI-NEXT: s_or_b32 s49, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s77 |
| ; VI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; VI-NEXT: s_or_b32 s50, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s74 |
| ; VI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; VI-NEXT: s_or_b32 s51, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s15 |
| ; VI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; VI-NEXT: s_or_b32 s52, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s12 |
| ; VI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; VI-NEXT: s_or_b32 s53, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s9 |
| ; VI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; VI-NEXT: s_or_b32 s54, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s7 |
| ; VI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; VI-NEXT: s_or_b32 s55, s4, s5 |
| ; VI-NEXT: s_cbranch_execnz .LBB43_3 |
| ; VI-NEXT: .LBB43_2: ; %cmp.true |
| ; VI-NEXT: s_add_i32 s16, s16, 3 |
| ; VI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s35, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s17, s17, 3 |
| ; VI-NEXT: s_add_i32 s36, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s34, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s18, s18, 3 |
| ; VI-NEXT: s_add_i32 s37, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s31, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s19, s19, 3 |
| ; VI-NEXT: s_add_i32 s38, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s30, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s20, s20, 3 |
| ; VI-NEXT: s_add_i32 s39, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s21, s21, 3 |
| ; VI-NEXT: s_add_i32 s40, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s22, s22, 3 |
| ; VI-NEXT: s_add_i32 s41, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s23, s23, 3 |
| ; VI-NEXT: s_add_i32 s42, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s24, s24, 3 |
| ; VI-NEXT: s_add_i32 s43, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s25, s25, 3 |
| ; VI-NEXT: s_add_i32 s44, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s26, s26, 3 |
| ; VI-NEXT: s_add_i32 s45, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s27, s27, 3 |
| ; VI-NEXT: s_add_i32 s46, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s28, s28, 3 |
| ; VI-NEXT: s_add_i32 s47, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s29, s29, 3 |
| ; VI-NEXT: s_add_i32 s48, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s77, s77, 3 |
| ; VI-NEXT: s_add_i32 s49, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s77, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s74, s74, 3 |
| ; VI-NEXT: s_add_i32 s50, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s74, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s15, s15, 3 |
| ; VI-NEXT: s_add_i32 s51, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s15, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s12, s12, 3 |
| ; VI-NEXT: s_add_i32 s52, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s9, s9, 3 |
| ; VI-NEXT: s_add_i32 s53, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s9, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s7, s7, 3 |
| ; VI-NEXT: s_add_i32 s54, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s7, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s55, s4, 0x30000 |
| ; VI-NEXT: .LBB43_3: ; %end |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_readlane_b32 s67, v20, 19 |
| ; VI-NEXT: v_readlane_b32 s66, v20, 18 |
| ; VI-NEXT: v_readlane_b32 s65, v20, 17 |
| ; VI-NEXT: v_readlane_b32 s64, v20, 16 |
| ; VI-NEXT: v_readlane_b32 s55, v20, 15 |
| ; VI-NEXT: v_readlane_b32 s54, v20, 14 |
| ; VI-NEXT: v_readlane_b32 s53, v20, 13 |
| ; VI-NEXT: v_readlane_b32 s52, v20, 12 |
| ; VI-NEXT: v_readlane_b32 s51, v20, 11 |
| ; VI-NEXT: v_readlane_b32 s50, v20, 10 |
| ; VI-NEXT: v_readlane_b32 s49, v20, 9 |
| ; VI-NEXT: v_readlane_b32 s48, v20, 8 |
| ; VI-NEXT: v_readlane_b32 s39, v20, 7 |
| ; VI-NEXT: v_readlane_b32 s38, v20, 6 |
| ; VI-NEXT: v_readlane_b32 s37, v20, 5 |
| ; VI-NEXT: v_readlane_b32 s36, v20, 4 |
| ; VI-NEXT: v_readlane_b32 s35, v20, 3 |
| ; VI-NEXT: v_readlane_b32 s34, v20, 2 |
| ; VI-NEXT: v_readlane_b32 s31, v20, 1 |
| ; VI-NEXT: v_readlane_b32 s30, v20, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB43_4: |
| ; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; VI-NEXT: s_branch .LBB43_2 |
| ; |
| ; GFX9-LABEL: bitcast_v40i16_to_v10i64_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_readfirstlane_b32 s56, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s58, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s60, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s62, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s72, v1 |
| ; GFX9-NEXT: v_readfirstlane_b32 s74, v0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: s_lshr_b32 s4, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s5, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s6, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s7, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s8, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s9, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s10, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s11, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s12, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s16, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s56, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s58, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s60, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s62, 16 |
| ; GFX9-NEXT: s_lshr_b32 s73, s72, 16 |
| ; GFX9-NEXT: s_lshr_b32 s75, s74, 16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s42, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_cmp_lg_u32 s42, 0 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s50, s74, s75 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s51, s72, s73 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s52, s62, s63 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s53, s60, s61 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s54, s58, s59 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s55, s56, s57 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB43_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB43_4 |
| ; GFX9-NEXT: .LBB43_2: ; %cmp.true |
| ; GFX9-NEXT: v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_branch .LBB43_5 |
| ; GFX9-NEXT: .LBB43_3: |
| ; GFX9-NEXT: s_branch .LBB43_2 |
| ; GFX9-NEXT: .LBB43_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB43_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v40i16_to_v10i64_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s45, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s46, v0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s56, v2 |
| ; GFX11-NEXT: s_lshr_b32 s41, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s6, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s5, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s4, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: s_lshr_b32 s58, s45, 16 |
| ; GFX11-NEXT: s_lshr_b32 s59, s46, 16 |
| ; GFX11-NEXT: s_mov_b32 s40, 0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s56, 0 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s57 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s47 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s44 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s18, s46, s59 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s19, s45, s58 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB43_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB43_4 |
| ; GFX11-NEXT: .LBB43_2: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v18, s18, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v19, s19, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB43_3: |
| ; GFX11-NEXT: s_branch .LBB43_2 |
| ; GFX11-NEXT: .LBB43_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31 |
| ; 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 <40 x i16> %a, splat (i16 3) |
| %a2 = bitcast <40 x i16> %a1 to <10 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x i16> %a to <10 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <10 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x i64> %phi |
| } |
| |
| define <40 x half> @bitcast_v10i64_to_v40f16(<10 x i64> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v10i64_to_v40f16: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr31 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr28 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr26 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr25 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr24 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr23 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; implicit-def: $vgpr22 |
| ; SI-NEXT: ; implicit-def: $vgpr30 |
| ; SI-NEXT: ; implicit-def: $vgpr21 |
| ; SI-NEXT: ; implicit-def: $vgpr29 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr27 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB44_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v28, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v31, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v34, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB44_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB44_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; SI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; SI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; SI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; SI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; SI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; SI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; SI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; SI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; SI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; SI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v28, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v31, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v34, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB44_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v34 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v31 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v28 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v25 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v35 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v23 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v32 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v22 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v30 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v29 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v27 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v34 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v31 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v28 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v26 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v25 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v24 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10i64_to_v40f16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; 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: 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 .LBB44_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB44_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB44_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18 |
| ; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc |
| ; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16 |
| ; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc |
| ; 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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB44_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10i64_to_v40f16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; 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 .LBB44_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB44_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB44_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc |
| ; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16 |
| ; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, 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: 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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB44_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_mov_b32 s4, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v39, v0, s4 |
| ; GFX9-NEXT: v_perm_b32 v1, v38, v1, s4 |
| ; GFX9-NEXT: v_perm_b32 v2, v37, v2, s4 |
| ; GFX9-NEXT: v_perm_b32 v3, v36, v3, s4 |
| ; GFX9-NEXT: v_perm_b32 v4, v35, v4, s4 |
| ; GFX9-NEXT: v_perm_b32 v5, v34, v5, s4 |
| ; GFX9-NEXT: v_perm_b32 v6, v33, v6, s4 |
| ; GFX9-NEXT: v_perm_b32 v7, v32, v7, s4 |
| ; GFX9-NEXT: v_perm_b32 v8, v31, v8, s4 |
| ; GFX9-NEXT: v_perm_b32 v9, v30, v9, s4 |
| ; GFX9-NEXT: v_perm_b32 v10, v29, v10, s4 |
| ; GFX9-NEXT: v_perm_b32 v11, v28, v11, s4 |
| ; GFX9-NEXT: v_perm_b32 v12, v27, v12, s4 |
| ; GFX9-NEXT: v_perm_b32 v13, v26, v13, s4 |
| ; GFX9-NEXT: v_perm_b32 v14, v25, v14, s4 |
| ; GFX9-NEXT: v_perm_b32 v15, v24, v15, s4 |
| ; GFX9-NEXT: v_perm_b32 v16, v23, v16, s4 |
| ; GFX9-NEXT: v_perm_b32 v17, v22, v17, s4 |
| ; GFX9-NEXT: v_perm_b32 v18, v21, v18, s4 |
| ; GFX9-NEXT: v_perm_b32 v19, v20, v19, s4 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v10i64_to_v40f16: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB44_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3 |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3 |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3 |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3 |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo |
| ; GFX11-TRUE16-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3 |
| ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo |
| ; GFX11-TRUE16-NEXT: .LBB44_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v10i64_to_v40f16: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB44_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB44_2: ; %Flow |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB44_4 |
| ; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3 |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3 |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3 |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3 |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3 |
| ; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB44_4: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v39, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v37, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v36, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v35, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v34, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v33, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v32, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v31, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v30, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v29, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v28, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v27, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v26, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v25, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v24, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v23, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v22, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v21, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v20, v19, 0x5040100 |
| ; GFX11-FAKE16-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 <10 x i64> %a, splat (i64 3) |
| %a2 = bitcast <10 x i64> %a1 to <40 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x i64> %a to <40 x half> |
| br label %end |
| |
| end: |
| %phi = phi <40 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x half> %phi |
| } |
| |
| define inreg <40 x half> @bitcast_v10i64_to_v40f16_scalar(<10 x i64> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v10i64_to_v40f16_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s8, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s5, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s8, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB45_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_lshr_b32 s72, s5, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s17, 16 |
| ; SI-NEXT: s_lshr_b64 s[10:11], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b64 s[12:13], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[14:15], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[40:41], s[28:29], 16 |
| ; SI-NEXT: s_lshr_b64 s[42:43], s[26:27], 16 |
| ; SI-NEXT: s_lshr_b64 s[44:45], s[24:25], 16 |
| ; SI-NEXT: s_lshr_b64 s[46:47], s[22:23], 16 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[20:21], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[18:19], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[16:17], 16 |
| ; SI-NEXT: s_cbranch_execnz .LBB45_3 |
| ; SI-NEXT: .LBB45_2: ; %cmp.true |
| ; SI-NEXT: s_add_u32 s4, s4, 3 |
| ; SI-NEXT: s_addc_u32 s5, s5, 0 |
| ; SI-NEXT: s_add_u32 s6, s6, 3 |
| ; SI-NEXT: s_addc_u32 s7, s7, 0 |
| ; SI-NEXT: s_add_u32 s8, s8, 3 |
| ; SI-NEXT: s_addc_u32 s9, s9, 0 |
| ; SI-NEXT: s_add_u32 s28, s28, 3 |
| ; SI-NEXT: s_addc_u32 s29, s29, 0 |
| ; SI-NEXT: s_add_u32 s26, s26, 3 |
| ; SI-NEXT: s_addc_u32 s27, s27, 0 |
| ; SI-NEXT: s_add_u32 s24, s24, 3 |
| ; SI-NEXT: s_addc_u32 s25, s25, 0 |
| ; SI-NEXT: s_add_u32 s22, s22, 3 |
| ; SI-NEXT: s_addc_u32 s23, s23, 0 |
| ; SI-NEXT: s_add_u32 s20, s20, 3 |
| ; SI-NEXT: s_addc_u32 s21, s21, 0 |
| ; SI-NEXT: s_add_u32 s18, s18, 3 |
| ; SI-NEXT: s_addc_u32 s19, s19, 0 |
| ; SI-NEXT: s_add_u32 s16, s16, 3 |
| ; SI-NEXT: s_addc_u32 s17, s17, 0 |
| ; SI-NEXT: s_lshr_b32 s72, s5, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s17, 16 |
| ; SI-NEXT: s_lshr_b64 s[10:11], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b64 s[12:13], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[14:15], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[40:41], s[28:29], 16 |
| ; SI-NEXT: s_lshr_b64 s[42:43], s[26:27], 16 |
| ; SI-NEXT: s_lshr_b64 s[44:45], s[24:25], 16 |
| ; SI-NEXT: s_lshr_b64 s[46:47], s[22:23], 16 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[20:21], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[18:19], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[16:17], 16 |
| ; SI-NEXT: .LBB45_3: ; %end |
| ; SI-NEXT: s_lshl_b32 s11, s60, 16 |
| ; SI-NEXT: s_and_b32 s13, s16, 0xffff |
| ; SI-NEXT: s_or_b32 s11, s13, s11 |
| ; SI-NEXT: s_and_b32 s13, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s15, s89, 16 |
| ; SI-NEXT: s_or_b32 s13, s13, s15 |
| ; SI-NEXT: s_lshl_b32 s15, s58, 16 |
| ; SI-NEXT: s_and_b32 s16, s18, 0xffff |
| ; SI-NEXT: s_or_b32 s15, s16, s15 |
| ; SI-NEXT: s_and_b32 s16, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s17, s88, 16 |
| ; SI-NEXT: s_or_b32 s16, s16, s17 |
| ; SI-NEXT: s_lshl_b32 s17, s56, 16 |
| ; SI-NEXT: s_and_b32 s18, s20, 0xffff |
| ; SI-NEXT: s_or_b32 s17, s18, s17 |
| ; SI-NEXT: s_and_b32 s18, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s19, s79, 16 |
| ; SI-NEXT: s_or_b32 s18, s18, s19 |
| ; SI-NEXT: s_and_b32 s19, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s20, s46, 16 |
| ; SI-NEXT: s_or_b32 s19, s19, s20 |
| ; SI-NEXT: s_and_b32 s20, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s21, s78, 16 |
| ; SI-NEXT: s_or_b32 s20, s20, s21 |
| ; SI-NEXT: s_and_b32 s21, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s22, s44, 16 |
| ; SI-NEXT: s_or_b32 s21, s21, s22 |
| ; SI-NEXT: s_and_b32 s22, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s23, s77, 16 |
| ; SI-NEXT: s_or_b32 s22, s22, s23 |
| ; SI-NEXT: s_and_b32 s23, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s42, 16 |
| ; SI-NEXT: s_or_b32 s23, s23, s24 |
| ; SI-NEXT: s_and_b32 s24, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s25, s76, 16 |
| ; SI-NEXT: s_or_b32 s24, s24, s25 |
| ; SI-NEXT: s_and_b32 s25, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s26, s40, 16 |
| ; SI-NEXT: s_and_b32 s8, s8, 0xffff |
| ; SI-NEXT: s_lshl_b32 s14, s14, 16 |
| ; SI-NEXT: s_and_b32 s6, s6, 0xffff |
| ; SI-NEXT: s_lshl_b32 s12, s12, 16 |
| ; SI-NEXT: s_and_b32 s4, s4, 0xffff |
| ; SI-NEXT: s_lshl_b32 s10, s10, 16 |
| ; SI-NEXT: s_or_b32 s25, s25, s26 |
| ; SI-NEXT: s_and_b32 s26, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s27, s75, 16 |
| ; SI-NEXT: s_or_b32 s8, s8, s14 |
| ; SI-NEXT: s_and_b32 s9, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s14, s74, 16 |
| ; SI-NEXT: s_or_b32 s6, s6, s12 |
| ; SI-NEXT: s_and_b32 s7, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s12, s73, 16 |
| ; SI-NEXT: s_or_b32 s4, s4, s10 |
| ; SI-NEXT: s_and_b32 s5, s5, 0xffff |
| ; SI-NEXT: s_lshl_b32 s10, s72, 16 |
| ; SI-NEXT: s_or_b32 s26, s26, s27 |
| ; SI-NEXT: s_or_b32 s9, s9, s14 |
| ; SI-NEXT: s_or_b32 s7, s7, s12 |
| ; SI-NEXT: s_or_b32 s5, s5, s10 |
| ; SI-NEXT: v_mov_b32_e32 v0, s11 |
| ; SI-NEXT: v_mov_b32_e32 v1, s13 |
| ; SI-NEXT: v_mov_b32_e32 v2, s15 |
| ; SI-NEXT: v_mov_b32_e32 v3, s16 |
| ; SI-NEXT: v_mov_b32_e32 v4, s17 |
| ; SI-NEXT: v_mov_b32_e32 v5, s18 |
| ; SI-NEXT: v_mov_b32_e32 v6, s19 |
| ; SI-NEXT: v_mov_b32_e32 v7, s20 |
| ; SI-NEXT: v_mov_b32_e32 v8, s21 |
| ; SI-NEXT: v_mov_b32_e32 v9, s22 |
| ; SI-NEXT: v_mov_b32_e32 v10, s23 |
| ; SI-NEXT: v_mov_b32_e32 v11, s24 |
| ; SI-NEXT: v_mov_b32_e32 v12, s25 |
| ; SI-NEXT: v_mov_b32_e32 v13, s26 |
| ; SI-NEXT: v_mov_b32_e32 v14, s8 |
| ; SI-NEXT: v_mov_b32_e32 v15, s9 |
| ; SI-NEXT: v_mov_b32_e32 v16, s6 |
| ; SI-NEXT: v_mov_b32_e32 v17, s7 |
| ; SI-NEXT: v_mov_b32_e32 v18, s4 |
| ; SI-NEXT: v_mov_b32_e32 v19, s5 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB45_4: |
| ; SI-NEXT: ; implicit-def: $sgpr60 |
| ; SI-NEXT: ; implicit-def: $sgpr89 |
| ; SI-NEXT: ; implicit-def: $sgpr58 |
| ; SI-NEXT: ; implicit-def: $sgpr88 |
| ; SI-NEXT: ; implicit-def: $sgpr56 |
| ; SI-NEXT: ; implicit-def: $sgpr79 |
| ; SI-NEXT: ; implicit-def: $sgpr46 |
| ; SI-NEXT: ; implicit-def: $sgpr78 |
| ; SI-NEXT: ; implicit-def: $sgpr44 |
| ; SI-NEXT: ; implicit-def: $sgpr77 |
| ; SI-NEXT: ; implicit-def: $sgpr42 |
| ; SI-NEXT: ; implicit-def: $sgpr76 |
| ; SI-NEXT: ; implicit-def: $sgpr40 |
| ; SI-NEXT: ; implicit-def: $sgpr75 |
| ; SI-NEXT: ; implicit-def: $sgpr14 |
| ; SI-NEXT: ; implicit-def: $sgpr74 |
| ; SI-NEXT: ; implicit-def: $sgpr12 |
| ; SI-NEXT: ; implicit-def: $sgpr73 |
| ; SI-NEXT: ; implicit-def: $sgpr10 |
| ; SI-NEXT: ; implicit-def: $sgpr72 |
| ; SI-NEXT: s_branch .LBB45_2 |
| ; |
| ; VI-LABEL: bitcast_v10i64_to_v40f16_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s10, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB45_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_lshr_b32 s12, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s10, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s46, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: s_cbranch_execnz .LBB45_3 |
| ; VI-NEXT: .LBB45_2: ; %cmp.true |
| ; VI-NEXT: s_add_u32 s7, s7, 3 |
| ; VI-NEXT: s_addc_u32 s6, s6, 0 |
| ; VI-NEXT: s_add_u32 s9, s9, 3 |
| ; VI-NEXT: s_addc_u32 s8, s8, 0 |
| ; VI-NEXT: s_add_u32 s11, s11, 3 |
| ; VI-NEXT: s_addc_u32 s10, s10, 0 |
| ; VI-NEXT: s_add_u32 s28, s28, 3 |
| ; VI-NEXT: s_addc_u32 s29, s29, 0 |
| ; VI-NEXT: s_add_u32 s26, s26, 3 |
| ; VI-NEXT: s_addc_u32 s27, s27, 0 |
| ; VI-NEXT: s_add_u32 s24, s24, 3 |
| ; VI-NEXT: s_addc_u32 s25, s25, 0 |
| ; VI-NEXT: s_add_u32 s22, s22, 3 |
| ; VI-NEXT: s_addc_u32 s23, s23, 0 |
| ; VI-NEXT: s_add_u32 s20, s20, 3 |
| ; VI-NEXT: s_addc_u32 s21, s21, 0 |
| ; VI-NEXT: s_add_u32 s18, s18, 3 |
| ; VI-NEXT: s_addc_u32 s19, s19, 0 |
| ; VI-NEXT: s_add_u32 s16, s16, 3 |
| ; VI-NEXT: s_addc_u32 s17, s17, 0 |
| ; VI-NEXT: s_lshr_b32 s12, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s10, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s46, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: .LBB45_3: ; %end |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s63, 16 |
| ; VI-NEXT: s_or_b32 s4, s4, s5 |
| ; VI-NEXT: s_and_b32 s5, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s16, s62, 16 |
| ; VI-NEXT: s_or_b32 s5, s5, s16 |
| ; VI-NEXT: s_and_b32 s16, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s17, s61, 16 |
| ; VI-NEXT: s_or_b32 s16, s16, s17 |
| ; VI-NEXT: s_and_b32 s17, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s18, s60, 16 |
| ; VI-NEXT: s_or_b32 s17, s17, s18 |
| ; VI-NEXT: s_and_b32 s18, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s19, s59, 16 |
| ; VI-NEXT: s_or_b32 s18, s18, s19 |
| ; VI-NEXT: s_and_b32 s19, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s20, s58, 16 |
| ; VI-NEXT: s_or_b32 s19, s19, s20 |
| ; VI-NEXT: s_and_b32 s20, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s21, s57, 16 |
| ; VI-NEXT: s_or_b32 s20, s20, s21 |
| ; VI-NEXT: s_and_b32 s21, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s22, s56, 16 |
| ; VI-NEXT: s_or_b32 s21, s21, s22 |
| ; VI-NEXT: s_and_b32 s22, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s23, s47, 16 |
| ; VI-NEXT: s_or_b32 s22, s22, s23 |
| ; VI-NEXT: s_and_b32 s23, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s24, s46, 16 |
| ; VI-NEXT: s_or_b32 s23, s23, s24 |
| ; VI-NEXT: s_and_b32 s24, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s25, s45, 16 |
| ; VI-NEXT: s_or_b32 s24, s24, s25 |
| ; VI-NEXT: s_and_b32 s25, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s26, s44, 16 |
| ; VI-NEXT: s_or_b32 s25, s25, s26 |
| ; VI-NEXT: s_and_b32 s26, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s27, s43, 16 |
| ; VI-NEXT: s_or_b32 s26, s26, s27 |
| ; VI-NEXT: s_and_b32 s27, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s28, s42, 16 |
| ; VI-NEXT: s_or_b32 s27, s27, s28 |
| ; VI-NEXT: s_and_b32 s11, 0xffff, s11 |
| ; VI-NEXT: s_lshl_b32 s28, s41, 16 |
| ; VI-NEXT: s_or_b32 s11, s11, s28 |
| ; VI-NEXT: s_and_b32 s10, 0xffff, s10 |
| ; VI-NEXT: s_lshl_b32 s28, s40, 16 |
| ; VI-NEXT: s_and_b32 s9, 0xffff, s9 |
| ; VI-NEXT: s_lshl_b32 s15, s15, 16 |
| ; VI-NEXT: s_and_b32 s8, 0xffff, s8 |
| ; VI-NEXT: s_lshl_b32 s14, s14, 16 |
| ; VI-NEXT: s_and_b32 s7, 0xffff, s7 |
| ; VI-NEXT: s_lshl_b32 s13, s13, 16 |
| ; VI-NEXT: s_and_b32 s6, 0xffff, s6 |
| ; VI-NEXT: s_lshl_b32 s12, s12, 16 |
| ; VI-NEXT: s_or_b32 s10, s10, s28 |
| ; VI-NEXT: s_or_b32 s9, s9, s15 |
| ; VI-NEXT: s_or_b32 s8, s8, s14 |
| ; VI-NEXT: s_or_b32 s7, s7, s13 |
| ; VI-NEXT: s_or_b32 s6, s6, s12 |
| ; VI-NEXT: v_mov_b32_e32 v0, s4 |
| ; VI-NEXT: v_mov_b32_e32 v1, s5 |
| ; VI-NEXT: v_mov_b32_e32 v2, s16 |
| ; VI-NEXT: v_mov_b32_e32 v3, s17 |
| ; VI-NEXT: v_mov_b32_e32 v4, s18 |
| ; VI-NEXT: v_mov_b32_e32 v5, s19 |
| ; VI-NEXT: v_mov_b32_e32 v6, s20 |
| ; VI-NEXT: v_mov_b32_e32 v7, s21 |
| ; VI-NEXT: v_mov_b32_e32 v8, s22 |
| ; VI-NEXT: v_mov_b32_e32 v9, s23 |
| ; VI-NEXT: v_mov_b32_e32 v10, s24 |
| ; VI-NEXT: v_mov_b32_e32 v11, s25 |
| ; VI-NEXT: v_mov_b32_e32 v12, s26 |
| ; VI-NEXT: v_mov_b32_e32 v13, s27 |
| ; VI-NEXT: v_mov_b32_e32 v14, s11 |
| ; VI-NEXT: v_mov_b32_e32 v15, s10 |
| ; VI-NEXT: v_mov_b32_e32 v16, s9 |
| ; VI-NEXT: v_mov_b32_e32 v17, s8 |
| ; VI-NEXT: v_mov_b32_e32 v18, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB45_4: |
| ; VI-NEXT: ; implicit-def: $sgpr63 |
| ; VI-NEXT: ; implicit-def: $sgpr62 |
| ; VI-NEXT: ; implicit-def: $sgpr61 |
| ; VI-NEXT: ; implicit-def: $sgpr60 |
| ; VI-NEXT: ; implicit-def: $sgpr59 |
| ; VI-NEXT: ; implicit-def: $sgpr58 |
| ; VI-NEXT: ; implicit-def: $sgpr57 |
| ; VI-NEXT: ; implicit-def: $sgpr56 |
| ; VI-NEXT: ; implicit-def: $sgpr47 |
| ; VI-NEXT: ; implicit-def: $sgpr46 |
| ; VI-NEXT: ; implicit-def: $sgpr45 |
| ; VI-NEXT: ; implicit-def: $sgpr44 |
| ; VI-NEXT: ; implicit-def: $sgpr43 |
| ; VI-NEXT: ; implicit-def: $sgpr42 |
| ; VI-NEXT: ; implicit-def: $sgpr41 |
| ; VI-NEXT: ; implicit-def: $sgpr40 |
| ; VI-NEXT: ; implicit-def: $sgpr15 |
| ; VI-NEXT: ; implicit-def: $sgpr14 |
| ; VI-NEXT: ; implicit-def: $sgpr13 |
| ; VI-NEXT: ; implicit-def: $sgpr12 |
| ; VI-NEXT: s_branch .LBB45_2 |
| ; |
| ; GFX9-LABEL: bitcast_v10i64_to_v40f16_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s10, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s11, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB45_4 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_lshr_b32 s12, s6, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s7, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s8, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s9, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s10, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s11, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s43, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s58, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s60, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s62, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s16, 16 |
| ; GFX9-NEXT: s_cbranch_execnz .LBB45_3 |
| ; GFX9-NEXT: .LBB45_2: ; %cmp.true |
| ; GFX9-NEXT: s_add_u32 s7, s7, 3 |
| ; GFX9-NEXT: s_addc_u32 s6, s6, 0 |
| ; GFX9-NEXT: s_add_u32 s9, s9, 3 |
| ; GFX9-NEXT: s_addc_u32 s8, s8, 0 |
| ; GFX9-NEXT: s_add_u32 s11, s11, 3 |
| ; GFX9-NEXT: s_addc_u32 s10, s10, 0 |
| ; GFX9-NEXT: s_add_u32 s28, s28, 3 |
| ; GFX9-NEXT: s_addc_u32 s29, s29, 0 |
| ; GFX9-NEXT: s_add_u32 s26, s26, 3 |
| ; GFX9-NEXT: s_addc_u32 s27, s27, 0 |
| ; GFX9-NEXT: s_add_u32 s24, s24, 3 |
| ; GFX9-NEXT: s_addc_u32 s25, s25, 0 |
| ; GFX9-NEXT: s_add_u32 s22, s22, 3 |
| ; GFX9-NEXT: s_addc_u32 s23, s23, 0 |
| ; GFX9-NEXT: s_add_u32 s20, s20, 3 |
| ; GFX9-NEXT: s_addc_u32 s21, s21, 0 |
| ; GFX9-NEXT: s_add_u32 s18, s18, 3 |
| ; GFX9-NEXT: s_addc_u32 s19, s19, 0 |
| ; GFX9-NEXT: s_add_u32 s16, s16, 3 |
| ; GFX9-NEXT: s_addc_u32 s17, s17, 0 |
| ; GFX9-NEXT: s_lshr_b32 s12, s6, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s7, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s8, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s9, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s10, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s11, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s43, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s58, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s60, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s62, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s16, 16 |
| ; GFX9-NEXT: .LBB45_3: ; %end |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s63 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s5, s17, s62 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s61 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s60 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s59 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s58 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s57 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s56 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s47 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s46 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s45 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s44 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s43 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s42 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s11, s11, s41 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s10, s10, s40 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s9, s9, s15 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s14 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s13 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s12 |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s5 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s6 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; GFX9-NEXT: .LBB45_4: |
| ; GFX9-NEXT: ; implicit-def: $sgpr63 |
| ; GFX9-NEXT: ; implicit-def: $sgpr62 |
| ; GFX9-NEXT: ; implicit-def: $sgpr61 |
| ; GFX9-NEXT: ; implicit-def: $sgpr60 |
| ; GFX9-NEXT: ; implicit-def: $sgpr59 |
| ; GFX9-NEXT: ; implicit-def: $sgpr58 |
| ; GFX9-NEXT: ; implicit-def: $sgpr57 |
| ; GFX9-NEXT: ; implicit-def: $sgpr56 |
| ; GFX9-NEXT: ; implicit-def: $sgpr47 |
| ; GFX9-NEXT: ; implicit-def: $sgpr46 |
| ; GFX9-NEXT: ; implicit-def: $sgpr45 |
| ; GFX9-NEXT: ; implicit-def: $sgpr44 |
| ; GFX9-NEXT: ; implicit-def: $sgpr43 |
| ; GFX9-NEXT: ; implicit-def: $sgpr42 |
| ; GFX9-NEXT: ; implicit-def: $sgpr41 |
| ; GFX9-NEXT: ; implicit-def: $sgpr40 |
| ; GFX9-NEXT: ; implicit-def: $sgpr15 |
| ; GFX9-NEXT: ; implicit-def: $sgpr14 |
| ; GFX9-NEXT: ; implicit-def: $sgpr13 |
| ; GFX9-NEXT: ; implicit-def: $sgpr12 |
| ; GFX9-NEXT: s_branch .LBB45_2 |
| ; |
| ; GFX11-LABEL: bitcast_v10i64_to_v40f16_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-NEXT: v_readfirstlane_b32 s4, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s5, v0 |
| ; GFX11-NEXT: s_mov_b32 s58, 0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB45_4 |
| ; GFX11-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-NEXT: s_lshr_b32 s6, s4, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s5, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s40, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s41, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s45, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s46, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s56, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s58 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB45_3 |
| ; GFX11-NEXT: .LBB45_2: ; %cmp.true |
| ; GFX11-NEXT: s_add_u32 s5, s5, 3 |
| ; GFX11-NEXT: s_addc_u32 s4, s4, 0 |
| ; GFX11-NEXT: s_add_u32 s28, s28, 3 |
| ; GFX11-NEXT: s_addc_u32 s29, s29, 0 |
| ; GFX11-NEXT: s_add_u32 s26, s26, 3 |
| ; GFX11-NEXT: s_addc_u32 s27, s27, 0 |
| ; GFX11-NEXT: s_add_u32 s24, s24, 3 |
| ; GFX11-NEXT: s_addc_u32 s25, s25, 0 |
| ; GFX11-NEXT: s_add_u32 s22, s22, 3 |
| ; GFX11-NEXT: s_addc_u32 s23, s23, 0 |
| ; GFX11-NEXT: s_add_u32 s20, s20, 3 |
| ; GFX11-NEXT: s_addc_u32 s21, s21, 0 |
| ; GFX11-NEXT: s_add_u32 s18, s18, 3 |
| ; GFX11-NEXT: s_addc_u32 s19, s19, 0 |
| ; GFX11-NEXT: s_add_u32 s16, s16, 3 |
| ; GFX11-NEXT: s_addc_u32 s17, s17, 0 |
| ; GFX11-NEXT: s_add_u32 s2, s2, 3 |
| ; GFX11-NEXT: s_addc_u32 s3, s3, 0 |
| ; GFX11-NEXT: s_add_u32 s0, s0, 3 |
| ; GFX11-NEXT: s_addc_u32 s1, s1, 0 |
| ; GFX11-NEXT: s_lshr_b32 s6, s4, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s5, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s40, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s41, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s45, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s46, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s56, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: .LBB45_3: ; %end |
| ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s57 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s56 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s47 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s46 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s45 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s44 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s43 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s42 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s41 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s40 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s15, s22, s15 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s14, s23, s14 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s13, s24, s13 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s12, s25, s12 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s11, s26, s11 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s10, s27, s10 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s9, s28, s9 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s8, s29, s8 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s7 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s6 |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4 |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB45_4: |
| ; GFX11-NEXT: ; implicit-def: $sgpr57 |
| ; GFX11-NEXT: ; implicit-def: $sgpr56 |
| ; GFX11-NEXT: ; implicit-def: $sgpr47 |
| ; GFX11-NEXT: ; implicit-def: $sgpr46 |
| ; GFX11-NEXT: ; implicit-def: $sgpr45 |
| ; GFX11-NEXT: ; implicit-def: $sgpr44 |
| ; GFX11-NEXT: ; implicit-def: $sgpr43 |
| ; GFX11-NEXT: ; implicit-def: $sgpr42 |
| ; GFX11-NEXT: ; implicit-def: $sgpr41 |
| ; GFX11-NEXT: ; implicit-def: $sgpr40 |
| ; GFX11-NEXT: ; implicit-def: $sgpr15 |
| ; GFX11-NEXT: ; implicit-def: $sgpr14 |
| ; GFX11-NEXT: ; implicit-def: $sgpr13 |
| ; GFX11-NEXT: ; implicit-def: $sgpr12 |
| ; GFX11-NEXT: ; implicit-def: $sgpr11 |
| ; GFX11-NEXT: ; implicit-def: $sgpr10 |
| ; GFX11-NEXT: ; implicit-def: $sgpr9 |
| ; GFX11-NEXT: ; implicit-def: $sgpr8 |
| ; GFX11-NEXT: ; implicit-def: $sgpr7 |
| ; GFX11-NEXT: ; implicit-def: $sgpr6 |
| ; GFX11-NEXT: s_branch .LBB45_2 |
| %cmp = icmp eq i32 %b, 0 |
| br i1 %cmp, label %cmp.true, label %cmp.false |
| |
| cmp.true: |
| %a1 = add <10 x i64> %a, splat (i64 3) |
| %a2 = bitcast <10 x i64> %a1 to <40 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x i64> %a to <40 x half> |
| br label %end |
| |
| end: |
| %phi = phi <40 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x half> %phi |
| } |
| |
| define <10 x i64> @bitcast_v40f16_to_v10i64(<40 x half> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v40f16_to_v10i64: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_mov_b32_e32 v32, v19 |
| ; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: v_mov_b32_e32 v33, v18 |
| ; SI-NEXT: v_mov_b32_e32 v43, v0 |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; SI-NEXT: v_mov_b32_e32 v34, v17 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; SI-NEXT: v_mov_b32_e32 v35, v16 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; SI-NEXT: v_mov_b32_e32 v36, v15 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v35 |
| ; SI-NEXT: v_mov_b32_e32 v37, v14 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v36 |
| ; SI-NEXT: v_mov_b32_e32 v38, v13 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v37 |
| ; SI-NEXT: v_mov_b32_e32 v39, v12 |
| ; SI-NEXT: v_mov_b32_e32 v48, v11 |
| ; SI-NEXT: v_mov_b32_e32 v49, v10 |
| ; SI-NEXT: v_mov_b32_e32 v50, v9 |
| ; SI-NEXT: v_mov_b32_e32 v51, v8 |
| ; SI-NEXT: v_mov_b32_e32 v52, v7 |
| ; SI-NEXT: v_mov_b32_e32 v53, v6 |
| ; SI-NEXT: v_mov_b32_e32 v54, v5 |
| ; SI-NEXT: v_mov_b32_e32 v55, v4 |
| ; SI-NEXT: v_mov_b32_e32 v40, v3 |
| ; SI-NEXT: v_mov_b32_e32 v41, v2 |
| ; SI-NEXT: v_mov_b32_e32 v42, v1 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v38 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v39 |
| ; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v48 |
| ; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v49 |
| ; SI-NEXT: v_lshrrev_b32_e32 v62, 16, v50 |
| ; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v51 |
| ; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v52 |
| ; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v53 |
| ; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v54 |
| ; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v55 |
| ; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v40 |
| ; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v41 |
| ; SI-NEXT: v_lshrrev_b32_e32 v58, 16, v42 |
| ; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v43 |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB46_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v43 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v59 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v42 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v58 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; SI-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v41 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v57 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v40 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v56 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v55 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v47 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v54 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v46 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v53 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v45 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v52 |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v44 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v51 |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v63 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v50 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v62 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v49 |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v61 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v60 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v39 |
| ; SI-NEXT: ; implicit-def: $vgpr43 |
| ; SI-NEXT: ; implicit-def: $vgpr42 |
| ; SI-NEXT: ; implicit-def: $vgpr41 |
| ; SI-NEXT: ; implicit-def: $vgpr40 |
| ; SI-NEXT: ; implicit-def: $vgpr55 |
| ; SI-NEXT: ; implicit-def: $vgpr54 |
| ; SI-NEXT: ; implicit-def: $vgpr53 |
| ; SI-NEXT: ; implicit-def: $vgpr52 |
| ; SI-NEXT: ; implicit-def: $vgpr51 |
| ; SI-NEXT: ; implicit-def: $vgpr50 |
| ; SI-NEXT: ; implicit-def: $vgpr49 |
| ; SI-NEXT: ; implicit-def: $vgpr48 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr59 |
| ; SI-NEXT: ; implicit-def: $vgpr58 |
| ; SI-NEXT: ; implicit-def: $vgpr57 |
| ; SI-NEXT: ; implicit-def: $vgpr56 |
| ; SI-NEXT: ; implicit-def: $vgpr47 |
| ; SI-NEXT: ; implicit-def: $vgpr46 |
| ; SI-NEXT: ; implicit-def: $vgpr45 |
| ; SI-NEXT: ; implicit-def: $vgpr44 |
| ; SI-NEXT: ; implicit-def: $vgpr63 |
| ; SI-NEXT: ; implicit-def: $vgpr62 |
| ; SI-NEXT: ; implicit-def: $vgpr61 |
| ; SI-NEXT: ; implicit-def: $vgpr60 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v38 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v37 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v36 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v35 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v34 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v33 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v19 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v32 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: .LBB46_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB46_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v0, v59 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, v58 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, v43 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v42 |
| ; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; SI-NEXT: v_or_b32_e32 v1, v3, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, v57 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v41 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, v40 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, v55 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v56 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, v54 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, v53 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, v47 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, v44 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, v52 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; SI-NEXT: v_or_b32_e32 v4, v5, v4 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, v46 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, v51 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, v61 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, v45 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, v49 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; SI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, v63 |
| ; SI-NEXT: v_or_b32_e32 v7, v9, v7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, v62 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; SI-NEXT: v_or_b32_e32 v8, v10, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, v50 |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, v38 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, v37 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v11 |
| ; SI-NEXT: v_or_b32_e32 v10, v12, v10 |
| ; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, v60 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, v35 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, v34 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; SI-NEXT: v_or_b32_e32 v11, v13, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, v39 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v21, v32 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, v12 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; SI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v14 |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: v_or_b32_e32 v13, v15, v13 |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v20, v20 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; SI-NEXT: v_or_b32_e32 v14, v16, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, v36 |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v17 |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: v_or_b32_e32 v16, v18, v16 |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_or_b32_e32 v17, v19, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, v33 |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_or_b32_e32 v18, v19, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v19, v21, v19 |
| ; SI-NEXT: .LBB46_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40f16_to_v10i64: |
| ; 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:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: v_mov_b32_e32 v32, v19 |
| ; VI-NEXT: v_mov_b32_e32 v33, v18 |
| ; VI-NEXT: v_mov_b32_e32 v34, v17 |
| ; VI-NEXT: v_mov_b32_e32 v35, v16 |
| ; VI-NEXT: v_mov_b32_e32 v36, v15 |
| ; VI-NEXT: v_mov_b32_e32 v37, v14 |
| ; VI-NEXT: v_mov_b32_e32 v38, v13 |
| ; VI-NEXT: v_mov_b32_e32 v39, v12 |
| ; VI-NEXT: v_mov_b32_e32 v48, v11 |
| ; VI-NEXT: v_mov_b32_e32 v49, v10 |
| ; VI-NEXT: v_mov_b32_e32 v50, v9 |
| ; VI-NEXT: v_mov_b32_e32 v51, v8 |
| ; VI-NEXT: v_mov_b32_e32 v52, v7 |
| ; VI-NEXT: v_mov_b32_e32 v53, v6 |
| ; VI-NEXT: v_mov_b32_e32 v54, v5 |
| ; VI-NEXT: v_mov_b32_e32 v55, v4 |
| ; VI-NEXT: v_mov_b32_e32 v40, v3 |
| ; VI-NEXT: v_mov_b32_e32 v41, v2 |
| ; VI-NEXT: v_mov_b32_e32 v42, v1 |
| ; VI-NEXT: v_mov_b32_e32 v43, v0 |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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 .LBB46_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_mov_b32_e32 v19, 16 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v0, v19, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v1, v19, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v2, v19, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v3, v19, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v4, v19, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v5, v19, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v6, v19, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v7, v19, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v8, v19, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v9, v19, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v10, v19, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v11, v19, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v12, v19, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v13, v19, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v14, v19, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v15, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v16, v19, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v17, v19, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v18, v19, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v19, v19, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v40, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v55, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v54, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v53, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v52, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v51, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v50, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v49, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v48, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v39, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v38, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v37, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v36, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v35, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v34, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v33, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v32, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; 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: .LBB46_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB46_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v19, 0x200 |
| ; VI-NEXT: v_add_f16_sdwa v0, v43, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v1, 0x200, v43 |
| ; VI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; VI-NEXT: v_add_f16_sdwa v1, v42, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v2, 0x200, v42 |
| ; VI-NEXT: v_or_b32_e32 v1, v2, v1 |
| ; VI-NEXT: v_add_f16_sdwa v2, v41, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, 0x200, v41 |
| ; VI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; VI-NEXT: v_add_f16_sdwa v3, v40, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v4, 0x200, v40 |
| ; VI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; VI-NEXT: v_add_f16_sdwa v4, v55, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v5, 0x200, v55 |
| ; VI-NEXT: v_or_b32_e32 v4, v5, v4 |
| ; VI-NEXT: v_add_f16_sdwa v5, v54, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v6, 0x200, v54 |
| ; VI-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; VI-NEXT: v_add_f16_sdwa v6, v53, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v7, 0x200, v53 |
| ; VI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; VI-NEXT: v_add_f16_sdwa v7, v52, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v8, 0x200, v52 |
| ; VI-NEXT: v_or_b32_e32 v7, v8, v7 |
| ; VI-NEXT: v_add_f16_sdwa v8, v51, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v9, 0x200, v51 |
| ; VI-NEXT: v_or_b32_e32 v8, v9, v8 |
| ; VI-NEXT: v_add_f16_sdwa v9, v50, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v10, 0x200, v50 |
| ; VI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; VI-NEXT: v_add_f16_sdwa v10, v49, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v11, 0x200, v49 |
| ; VI-NEXT: v_or_b32_e32 v10, v11, v10 |
| ; VI-NEXT: v_add_f16_sdwa v11, v48, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v12, 0x200, v48 |
| ; VI-NEXT: v_or_b32_e32 v11, v12, v11 |
| ; VI-NEXT: v_add_f16_sdwa v12, v39, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v13, 0x200, v39 |
| ; VI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; VI-NEXT: v_add_f16_sdwa v13, v38, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v14, 0x200, v38 |
| ; VI-NEXT: v_or_b32_e32 v13, v14, v13 |
| ; VI-NEXT: v_add_f16_sdwa v14, v37, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v15, 0x200, v37 |
| ; VI-NEXT: v_or_b32_e32 v14, v15, v14 |
| ; VI-NEXT: v_add_f16_sdwa v15, v36, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v16, 0x200, v36 |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: v_add_f16_sdwa v16, v35, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v17, 0x200, v35 |
| ; VI-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; VI-NEXT: v_add_f16_sdwa v17, v34, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v18, 0x200, v34 |
| ; VI-NEXT: v_or_b32_e32 v17, v18, v17 |
| ; VI-NEXT: v_add_f16_sdwa v18, v33, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v20, 0x200, v33 |
| ; VI-NEXT: v_or_b32_e32 v18, v20, v18 |
| ; VI-NEXT: v_add_f16_sdwa v19, v32, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v20, 0x200, v32 |
| ; VI-NEXT: v_or_b32_e32 v19, v20, v19 |
| ; VI-NEXT: .LBB46_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40f16_to_v10i64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_mov_b32_e32 v32, v19 |
| ; 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: v_mov_b32_e32 v33, v18 |
| ; GFX9-NEXT: v_mov_b32_e32 v43, v0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, v17 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, v16 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, v15 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, v14 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, v13 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, v12 |
| ; GFX9-NEXT: v_mov_b32_e32 v48, v11 |
| ; GFX9-NEXT: v_mov_b32_e32 v49, v10 |
| ; GFX9-NEXT: v_mov_b32_e32 v50, v9 |
| ; GFX9-NEXT: v_mov_b32_e32 v51, v8 |
| ; GFX9-NEXT: v_mov_b32_e32 v52, v7 |
| ; GFX9-NEXT: v_mov_b32_e32 v53, v6 |
| ; GFX9-NEXT: v_mov_b32_e32 v54, v5 |
| ; GFX9-NEXT: v_mov_b32_e32 v55, v4 |
| ; GFX9-NEXT: v_mov_b32_e32 v40, v3 |
| ; GFX9-NEXT: v_mov_b32_e32 v41, v2 |
| ; GFX9-NEXT: v_mov_b32_e32 v42, v1 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v48 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v49 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v50 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v51 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v52 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v53 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v54 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v55 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43 |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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 .LBB46_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v35 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v34 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v33 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v32 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; 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: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: .LBB46_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB46_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: s_movk_i32 s7, 0x200 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_pk_add_f16 v0, v0, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, v1, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, v2, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, v3, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, v4, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, v5, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; GFX9-NEXT: v_pk_add_f16 v12, v12, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, v13, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, v14, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v15, v15, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v16, v16, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v17, v17, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v18, v18, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v19, v19, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB46_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; 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-TRUE16-LABEL: bitcast_v40f16_to_v10i64: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB46_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: .LBB46_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v40f16_to_v10i64: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v37, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v39, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v48, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v36, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v35, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v34, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v33, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v32, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v31, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v30, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v29, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v28, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v27, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v26, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v25, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v24, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v23, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v22, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v21, v19, 0x5040100 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB46_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: .LBB46_2: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-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 <40 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <40 x half> %a1 to <10 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x half> %a to <10 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <10 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x i64> %phi |
| } |
| |
| define inreg <10 x i64> @bitcast_v40f16_to_v10i64_scalar(<40 x half> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v40f16_to_v10i64_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; SI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; SI-NEXT: v_writelane_b32 v32, s64, 12 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s10, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s12, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s14, v1 |
| ; SI-NEXT: v_readfirstlane_b32 s73, v0 |
| ; SI-NEXT: v_writelane_b32 v32, s65, 13 |
| ; SI-NEXT: s_lshr_b32 s72, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s28, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s26, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; SI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; SI-NEXT: s_lshr_b32 s92, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s93, s18, 16 |
| ; SI-NEXT: s_lshr_b32 s94, s17, 16 |
| ; SI-NEXT: s_lshr_b32 s95, s16, 16 |
| ; SI-NEXT: s_lshr_b32 s7, s6, 16 |
| ; SI-NEXT: s_lshr_b32 s9, s8, 16 |
| ; SI-NEXT: s_lshr_b32 s11, s10, 16 |
| ; SI-NEXT: s_lshr_b32 s13, s12, 16 |
| ; SI-NEXT: s_lshr_b32 s15, s14, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s73, 16 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v32, s66, 14 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s67, 15 |
| ; SI-NEXT: s_cbranch_scc0 .LBB47_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s95, 16 |
| ; SI-NEXT: s_or_b32 s36, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s94, 16 |
| ; SI-NEXT: s_or_b32 s37, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s93, 16 |
| ; SI-NEXT: s_or_b32 s38, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s92, 16 |
| ; SI-NEXT: s_or_b32 s39, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; SI-NEXT: s_or_b32 s40, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; SI-NEXT: s_or_b32 s41, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; SI-NEXT: s_or_b32 s42, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; SI-NEXT: s_or_b32 s43, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; SI-NEXT: s_or_b32 s44, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; SI-NEXT: s_or_b32 s45, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s77, 16 |
| ; SI-NEXT: s_or_b32 s46, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; SI-NEXT: s_or_b32 s47, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; SI-NEXT: s_or_b32 s48, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; SI-NEXT: s_or_b32 s49, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s73, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s74, 16 |
| ; SI-NEXT: s_or_b32 s50, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s14, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s15, 16 |
| ; SI-NEXT: s_or_b32 s51, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; SI-NEXT: s_or_b32 s52, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s10, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; SI-NEXT: s_or_b32 s53, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s8, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s9, 16 |
| ; SI-NEXT: s_or_b32 s54, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s6, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s7, 16 |
| ; SI-NEXT: s_or_b32 s55, s4, s5 |
| ; SI-NEXT: s_cbranch_execnz .LBB47_4 |
| ; SI-NEXT: .LBB47_2: ; %cmp.true |
| ; SI-NEXT: v_cvt_f32_f16_e32 v0, s95 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, s16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, s94 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s17 |
| ; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, s18 |
| ; SI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, s93 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_or_b32_e32 v1, v3, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s92 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_or_b32_e32 v2, v4, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, s19 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s91 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, s20 |
| ; SI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s90 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, s21 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_or_b32_e32 v4, v6, v4 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, s89 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_or_b32_e32 v5, v7, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, s22 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, s88 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, s23 |
| ; SI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, s79 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, s24 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_or_b32_e32 v7, v9, v7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, s78 |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; SI-NEXT: v_or_b32_e32 v8, v10, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, s25 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, s77 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, s26 |
| ; SI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, s76 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, s27 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_or_b32_e32 v10, v12, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, s75 |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; SI-NEXT: v_or_b32_e32 v11, v13, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, s28 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, s72 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, s29 |
| ; SI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, s74 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, s73 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_or_b32_e32 v13, v15, v13 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, s15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; SI-NEXT: v_or_b32_e32 v14, v16, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, s14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, s13 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, s12 |
| ; SI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, s11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, s10 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_or_b32_e32 v16, v18, v16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, s9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_or_b32_e32 v17, v19, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, s8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v20, s7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v21, s6 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_or_b32_e32 v18, v19, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v19, v21, v19 |
| ; SI-NEXT: s_branch .LBB47_5 |
| ; SI-NEXT: .LBB47_3: |
| ; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; SI-NEXT: s_branch .LBB47_2 |
| ; SI-NEXT: .LBB47_4: |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_mov_b32_e32 v20, s56 |
| ; SI-NEXT: v_mov_b32_e32 v21, s57 |
| ; SI-NEXT: v_mov_b32_e32 v22, s58 |
| ; SI-NEXT: v_mov_b32_e32 v23, s59 |
| ; SI-NEXT: v_mov_b32_e32 v24, s60 |
| ; SI-NEXT: v_mov_b32_e32 v25, s61 |
| ; SI-NEXT: v_mov_b32_e32 v26, s62 |
| ; SI-NEXT: v_mov_b32_e32 v27, s63 |
| ; SI-NEXT: v_mov_b32_e32 v28, s64 |
| ; SI-NEXT: v_mov_b32_e32 v29, s65 |
| ; SI-NEXT: v_mov_b32_e32 v30, s66 |
| ; SI-NEXT: v_mov_b32_e32 v31, s67 |
| ; SI-NEXT: .LBB47_5: ; %end |
| ; SI-NEXT: v_readlane_b32 s67, v32, 15 |
| ; SI-NEXT: v_readlane_b32 s66, v32, 14 |
| ; SI-NEXT: v_readlane_b32 s65, v32, 13 |
| ; SI-NEXT: v_readlane_b32 s64, v32, 12 |
| ; SI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40f16_to_v10i64_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v32, s30, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s31, 1 |
| ; VI-NEXT: v_writelane_b32 v32, s34, 2 |
| ; VI-NEXT: v_writelane_b32 v32, s35, 3 |
| ; VI-NEXT: v_writelane_b32 v32, s36, 4 |
| ; VI-NEXT: v_writelane_b32 v32, s37, 5 |
| ; VI-NEXT: v_writelane_b32 v32, s38, 6 |
| ; VI-NEXT: v_writelane_b32 v32, s39, 7 |
| ; VI-NEXT: v_writelane_b32 v32, s48, 8 |
| ; VI-NEXT: v_writelane_b32 v32, s49, 9 |
| ; VI-NEXT: v_writelane_b32 v32, s50, 10 |
| ; VI-NEXT: v_writelane_b32 v32, s51, 11 |
| ; VI-NEXT: v_writelane_b32 v32, s52, 12 |
| ; VI-NEXT: v_writelane_b32 v32, s53, 13 |
| ; VI-NEXT: v_writelane_b32 v32, s54, 14 |
| ; VI-NEXT: v_writelane_b32 v32, s55, 15 |
| ; VI-NEXT: v_writelane_b32 v32, s64, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s14, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s73, v1 |
| ; VI-NEXT: v_readfirstlane_b32 s76, v0 |
| ; VI-NEXT: v_writelane_b32 v32, s65, 17 |
| ; VI-NEXT: s_lshr_b32 s10, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s72, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s74, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s77, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s30, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s31, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s34, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s35, s16, 16 |
| ; VI-NEXT: s_lshr_b32 s7, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s9, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s12, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s14, 16 |
| ; VI-NEXT: s_lshr_b32 s75, s73, 16 |
| ; VI-NEXT: s_lshr_b32 s78, s76, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v32, s66, 18 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s67, 19 |
| ; VI-NEXT: s_cbranch_scc0 .LBB47_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s35, 16 |
| ; VI-NEXT: s_or_b32 s36, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s5, s34, 16 |
| ; VI-NEXT: s_or_b32 s37, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s5, s31, 16 |
| ; VI-NEXT: s_or_b32 s38, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s5, s30, 16 |
| ; VI-NEXT: s_or_b32 s39, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; VI-NEXT: s_or_b32 s40, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; VI-NEXT: s_or_b32 s41, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; VI-NEXT: s_or_b32 s42, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; VI-NEXT: s_or_b32 s43, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; VI-NEXT: s_or_b32 s44, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s5, s77, 16 |
| ; VI-NEXT: s_or_b32 s45, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s5, s74, 16 |
| ; VI-NEXT: s_or_b32 s46, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; VI-NEXT: s_or_b32 s47, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; VI-NEXT: s_or_b32 s48, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; VI-NEXT: s_or_b32 s49, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s76 |
| ; VI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; VI-NEXT: s_or_b32 s50, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s73 |
| ; VI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; VI-NEXT: s_or_b32 s51, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s14 |
| ; VI-NEXT: s_lshl_b32 s5, s15, 16 |
| ; VI-NEXT: s_or_b32 s52, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s11 |
| ; VI-NEXT: s_lshl_b32 s5, s12, 16 |
| ; VI-NEXT: s_or_b32 s53, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s8 |
| ; VI-NEXT: s_lshl_b32 s5, s9, 16 |
| ; VI-NEXT: s_or_b32 s54, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s6 |
| ; VI-NEXT: s_lshl_b32 s5, s7, 16 |
| ; VI-NEXT: s_or_b32 s55, s4, s5 |
| ; VI-NEXT: s_cbranch_execnz .LBB47_4 |
| ; VI-NEXT: .LBB47_2: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v19, 0x200 |
| ; VI-NEXT: v_mov_b32_e32 v0, s35 |
| ; VI-NEXT: v_mov_b32_e32 v2, s34 |
| ; VI-NEXT: v_add_f16_sdwa v0, v0, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v1, s16, v19 |
| ; VI-NEXT: v_add_f16_sdwa v2, v2, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, s17, v19 |
| ; VI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; VI-NEXT: v_or_b32_e32 v1, v3, v2 |
| ; VI-NEXT: v_mov_b32_e32 v2, s31 |
| ; VI-NEXT: v_add_f16_sdwa v2, v2, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, s18, v19 |
| ; VI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; VI-NEXT: v_mov_b32_e32 v3, s30 |
| ; VI-NEXT: v_add_f16_sdwa v3, v3, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v4, s19, v19 |
| ; VI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; VI-NEXT: v_mov_b32_e32 v4, s91 |
| ; VI-NEXT: v_add_f16_sdwa v4, v4, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v5, s20, v19 |
| ; VI-NEXT: v_or_b32_e32 v4, v5, v4 |
| ; VI-NEXT: v_mov_b32_e32 v5, s90 |
| ; VI-NEXT: v_add_f16_sdwa v5, v5, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v6, s21, v19 |
| ; VI-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; VI-NEXT: v_mov_b32_e32 v6, s89 |
| ; VI-NEXT: v_add_f16_sdwa v6, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v7, s22, v19 |
| ; VI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; VI-NEXT: v_mov_b32_e32 v7, s88 |
| ; VI-NEXT: v_add_f16_sdwa v7, v7, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v8, s23, v19 |
| ; VI-NEXT: v_or_b32_e32 v7, v8, v7 |
| ; VI-NEXT: v_mov_b32_e32 v8, s79 |
| ; VI-NEXT: v_add_f16_sdwa v8, v8, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v9, s24, v19 |
| ; VI-NEXT: v_or_b32_e32 v8, v9, v8 |
| ; VI-NEXT: v_mov_b32_e32 v9, s77 |
| ; VI-NEXT: v_add_f16_sdwa v9, v9, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v10, s25, v19 |
| ; VI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; VI-NEXT: v_mov_b32_e32 v10, s74 |
| ; VI-NEXT: v_add_f16_sdwa v10, v10, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v11, s26, v19 |
| ; VI-NEXT: v_or_b32_e32 v10, v11, v10 |
| ; VI-NEXT: v_mov_b32_e32 v11, s72 |
| ; VI-NEXT: v_add_f16_sdwa v11, v11, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v12, s27, v19 |
| ; VI-NEXT: v_or_b32_e32 v11, v12, v11 |
| ; VI-NEXT: v_mov_b32_e32 v12, s13 |
| ; VI-NEXT: v_add_f16_sdwa v12, v12, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v13, s28, v19 |
| ; VI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; VI-NEXT: v_mov_b32_e32 v13, s10 |
| ; VI-NEXT: v_add_f16_sdwa v13, v13, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v14, s29, v19 |
| ; VI-NEXT: v_or_b32_e32 v13, v14, v13 |
| ; VI-NEXT: v_mov_b32_e32 v14, s78 |
| ; VI-NEXT: v_add_f16_sdwa v14, v14, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v15, s76, v19 |
| ; VI-NEXT: v_or_b32_e32 v14, v15, v14 |
| ; VI-NEXT: v_mov_b32_e32 v15, s75 |
| ; VI-NEXT: v_add_f16_sdwa v15, v15, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v16, s73, v19 |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: v_mov_b32_e32 v16, s15 |
| ; VI-NEXT: v_add_f16_sdwa v16, v16, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v17, s14, v19 |
| ; VI-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; VI-NEXT: v_mov_b32_e32 v17, s12 |
| ; VI-NEXT: v_add_f16_sdwa v17, v17, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v18, s11, v19 |
| ; VI-NEXT: v_or_b32_e32 v17, v18, v17 |
| ; VI-NEXT: v_mov_b32_e32 v18, s9 |
| ; VI-NEXT: v_add_f16_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v20, s8, v19 |
| ; VI-NEXT: v_or_b32_e32 v18, v20, v18 |
| ; VI-NEXT: v_mov_b32_e32 v20, s7 |
| ; VI-NEXT: v_add_f16_sdwa v20, v20, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v19, s6, v19 |
| ; VI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; VI-NEXT: s_branch .LBB47_5 |
| ; VI-NEXT: .LBB47_3: |
| ; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; VI-NEXT: s_branch .LBB47_2 |
| ; VI-NEXT: .LBB47_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_mov_b32_e32 v20, s56 |
| ; VI-NEXT: v_mov_b32_e32 v21, s57 |
| ; VI-NEXT: v_mov_b32_e32 v22, s58 |
| ; VI-NEXT: v_mov_b32_e32 v23, s59 |
| ; VI-NEXT: v_mov_b32_e32 v24, s60 |
| ; VI-NEXT: v_mov_b32_e32 v25, s61 |
| ; VI-NEXT: v_mov_b32_e32 v26, s62 |
| ; VI-NEXT: v_mov_b32_e32 v27, s63 |
| ; VI-NEXT: v_mov_b32_e32 v28, s64 |
| ; VI-NEXT: v_mov_b32_e32 v29, s65 |
| ; VI-NEXT: v_mov_b32_e32 v30, s66 |
| ; VI-NEXT: v_mov_b32_e32 v31, s67 |
| ; VI-NEXT: .LBB47_5: ; %end |
| ; VI-NEXT: v_readlane_b32 s67, v32, 19 |
| ; VI-NEXT: v_readlane_b32 s66, v32, 18 |
| ; VI-NEXT: v_readlane_b32 s65, v32, 17 |
| ; VI-NEXT: v_readlane_b32 s64, v32, 16 |
| ; VI-NEXT: v_readlane_b32 s55, v32, 15 |
| ; VI-NEXT: v_readlane_b32 s54, v32, 14 |
| ; VI-NEXT: v_readlane_b32 s53, v32, 13 |
| ; VI-NEXT: v_readlane_b32 s52, v32, 12 |
| ; VI-NEXT: v_readlane_b32 s51, v32, 11 |
| ; VI-NEXT: v_readlane_b32 s50, v32, 10 |
| ; VI-NEXT: v_readlane_b32 s49, v32, 9 |
| ; VI-NEXT: v_readlane_b32 s48, v32, 8 |
| ; VI-NEXT: v_readlane_b32 s39, v32, 7 |
| ; VI-NEXT: v_readlane_b32 s38, v32, 6 |
| ; VI-NEXT: v_readlane_b32 s37, v32, 5 |
| ; VI-NEXT: v_readlane_b32 s36, v32, 4 |
| ; VI-NEXT: v_readlane_b32 s35, v32, 3 |
| ; VI-NEXT: v_readlane_b32 s34, v32, 2 |
| ; VI-NEXT: v_readlane_b32 s31, v32, 1 |
| ; VI-NEXT: v_readlane_b32 s30, v32, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40f16_to_v10i64_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_readfirstlane_b32 s56, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s58, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s60, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s62, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s72, v1 |
| ; GFX9-NEXT: v_readfirstlane_b32 s74, v0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: s_lshr_b32 s4, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s5, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s6, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s7, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s8, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s9, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s10, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s11, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s12, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s16, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s56, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s58, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s60, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s62, 16 |
| ; GFX9-NEXT: s_lshr_b32 s73, s72, 16 |
| ; GFX9-NEXT: s_lshr_b32 s75, s74, 16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s42, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_cmp_lg_u32 s42, 0 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s50, s74, s75 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s51, s72, s73 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s52, s62, s63 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s53, s60, s61 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s54, s58, s59 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s55, s56, s57 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB47_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB47_4 |
| ; GFX9-NEXT: .LBB47_2: ; %cmp.true |
| ; GFX9-NEXT: v_mov_b32_e32 v19, 0x200 |
| ; GFX9-NEXT: v_pk_add_f16 v0, s36, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, s37, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, s38, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, s39, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, s40, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, s41, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, s42, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, s43, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, s44, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, s45, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, s46, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, s47, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v12, s48, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, s49, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, s50, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v15, s51, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v16, s52, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v17, s53, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v18, s54, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v19, s55, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_branch .LBB47_5 |
| ; GFX9-NEXT: .LBB47_3: |
| ; GFX9-NEXT: s_branch .LBB47_2 |
| ; GFX9-NEXT: .LBB47_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB47_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v40f16_to_v10i64_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s45, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s46, v0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s56, v2 |
| ; GFX11-NEXT: s_lshr_b32 s41, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s6, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s5, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s4, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: s_lshr_b32 s58, s45, 16 |
| ; GFX11-NEXT: s_lshr_b32 s59, s46, 16 |
| ; GFX11-NEXT: s_mov_b32 s40, 0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s56, 0 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s57 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s47 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s44 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s18, s46, s59 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s19, s45, s58 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB47_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB47_4 |
| ; GFX11-NEXT: .LBB47_2: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v18, 0x200, s18 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v19, 0x200, s19 op_sel_hi:[0,1] |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB47_3: |
| ; GFX11-NEXT: s_branch .LBB47_2 |
| ; GFX11-NEXT: .LBB47_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31 |
| ; 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 <40 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <40 x half> %a1 to <10 x i64> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x half> %a to <10 x i64> |
| br label %end |
| |
| end: |
| %phi = phi <10 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x i64> %phi |
| } |
| |
| define <40 x i16> @bitcast_v10f64_to_v40i16(<10 x double> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v10f64_to_v40i16: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr29 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr26 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr25 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr24 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr23 |
| ; SI-NEXT: ; implicit-def: $vgpr31 |
| ; SI-NEXT: ; implicit-def: $vgpr22 |
| ; SI-NEXT: ; implicit-def: $vgpr30 |
| ; SI-NEXT: ; implicit-def: $vgpr21 |
| ; SI-NEXT: ; implicit-def: $vgpr28 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr27 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB48_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v32, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v35, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB48_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB48_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; SI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; SI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; SI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; SI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; SI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; SI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; SI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; SI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0 |
| ; SI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v32, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v35, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB48_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v35 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v32 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v29 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v25 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v34 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v23 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v31 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v22 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v30 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v28 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v27 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v35 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v32 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v29 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v26 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v25 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v24 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10f64_to_v40i16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; 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: 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 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; 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_add_f64 v[18:19], v[18:19], 1.0 |
| ; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; 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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB48_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10f64_to_v40i16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; 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 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; 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_add_f64 v[18:19], v[18:19], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; 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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB48_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_mov_b32 s4, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v39, v0, s4 |
| ; GFX9-NEXT: v_perm_b32 v1, v38, v1, s4 |
| ; GFX9-NEXT: v_perm_b32 v2, v37, v2, s4 |
| ; GFX9-NEXT: v_perm_b32 v3, v36, v3, s4 |
| ; GFX9-NEXT: v_perm_b32 v4, v35, v4, s4 |
| ; GFX9-NEXT: v_perm_b32 v5, v34, v5, s4 |
| ; GFX9-NEXT: v_perm_b32 v6, v33, v6, s4 |
| ; GFX9-NEXT: v_perm_b32 v7, v32, v7, s4 |
| ; GFX9-NEXT: v_perm_b32 v8, v31, v8, s4 |
| ; GFX9-NEXT: v_perm_b32 v9, v30, v9, s4 |
| ; GFX9-NEXT: v_perm_b32 v10, v29, v10, s4 |
| ; GFX9-NEXT: v_perm_b32 v11, v28, v11, s4 |
| ; GFX9-NEXT: v_perm_b32 v12, v27, v12, s4 |
| ; GFX9-NEXT: v_perm_b32 v13, v26, v13, s4 |
| ; GFX9-NEXT: v_perm_b32 v14, v25, v14, s4 |
| ; GFX9-NEXT: v_perm_b32 v15, v24, v15, s4 |
| ; GFX9-NEXT: v_perm_b32 v16, v23, v16, s4 |
| ; GFX9-NEXT: v_perm_b32 v17, v22, v17, s4 |
| ; GFX9-NEXT: v_perm_b32 v18, v21, v18, s4 |
| ; GFX9-NEXT: v_perm_b32 v19, v20, v19, s4 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v10f64_to_v40i16: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB48_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[18:19], v[18:19], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX11-TRUE16-NEXT: .LBB48_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v10f64_to_v40i16: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB48_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB48_2: ; %Flow |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB48_4 |
| ; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[18:19], v[18:19], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB48_4: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v39, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v37, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v36, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v35, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v34, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v33, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v32, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v31, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v30, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v29, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v28, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v27, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v26, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v25, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v24, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v23, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v22, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v21, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v20, v19, 0x5040100 |
| ; GFX11-FAKE16-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 <10 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <10 x double> %a1 to <40 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x double> %a to <40 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <40 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x i16> %phi |
| } |
| |
| define inreg <40 x i16> @bitcast_v10f64_to_v40i16_scalar(<10 x double> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v10f64_to_v40i16_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s5, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB49_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_lshr_b32 s89, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s5, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s72, s17, 16 |
| ; SI-NEXT: s_lshr_b64 s[10:11], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[12:13], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[14:15], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b64 s[40:41], s[28:29], 16 |
| ; SI-NEXT: s_lshr_b64 s[42:43], s[26:27], 16 |
| ; SI-NEXT: s_lshr_b64 s[44:45], s[24:25], 16 |
| ; SI-NEXT: s_lshr_b64 s[46:47], s[22:23], 16 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[20:21], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[18:19], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[16:17], 16 |
| ; SI-NEXT: s_cbranch_execnz .LBB49_4 |
| ; SI-NEXT: .LBB49_2: ; %cmp.true |
| ; SI-NEXT: v_add_f64 v[18:19], s[8:9], 1.0 |
| ; SI-NEXT: v_add_f64 v[16:17], s[6:7], 1.0 |
| ; SI-NEXT: v_add_f64 v[14:15], s[4:5], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[20:21], v[18:19], 16 |
| ; SI-NEXT: v_add_f64 v[12:13], s[28:29], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[21:22], v[16:17], 16 |
| ; SI-NEXT: v_add_f64 v[10:11], s[26:27], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[22:23], v[14:15], 16 |
| ; SI-NEXT: v_add_f64 v[8:9], s[24:25], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[23:24], v[12:13], 16 |
| ; SI-NEXT: v_add_f64 v[6:7], s[22:23], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[24:25], v[10:11], 16 |
| ; SI-NEXT: v_add_f64 v[4:5], s[20:21], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[25:26], v[8:9], 16 |
| ; SI-NEXT: v_add_f64 v[2:3], s[18:19], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[26:27], v[6:7], 16 |
| ; SI-NEXT: v_add_f64 v[0:1], s[16:17], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[27:28], v[4:5], 16 |
| ; SI-NEXT: v_lshr_b64 v[28:29], v[2:3], 16 |
| ; SI-NEXT: v_lshr_b64 v[29:30], v[0:1], 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v1 |
| ; SI-NEXT: s_branch .LBB49_5 |
| ; SI-NEXT: .LBB49_3: |
| ; SI-NEXT: ; implicit-def: $sgpr60 |
| ; SI-NEXT: ; implicit-def: $sgpr72 |
| ; SI-NEXT: ; implicit-def: $sgpr58 |
| ; SI-NEXT: ; implicit-def: $sgpr73 |
| ; SI-NEXT: ; implicit-def: $sgpr56 |
| ; SI-NEXT: ; implicit-def: $sgpr74 |
| ; SI-NEXT: ; implicit-def: $sgpr46 |
| ; SI-NEXT: ; implicit-def: $sgpr75 |
| ; SI-NEXT: ; implicit-def: $sgpr44 |
| ; SI-NEXT: ; implicit-def: $sgpr76 |
| ; SI-NEXT: ; implicit-def: $sgpr42 |
| ; SI-NEXT: ; implicit-def: $sgpr77 |
| ; SI-NEXT: ; implicit-def: $sgpr40 |
| ; SI-NEXT: ; implicit-def: $sgpr78 |
| ; SI-NEXT: ; implicit-def: $sgpr14 |
| ; SI-NEXT: ; implicit-def: $sgpr79 |
| ; SI-NEXT: ; implicit-def: $sgpr12 |
| ; SI-NEXT: ; implicit-def: $sgpr88 |
| ; SI-NEXT: ; implicit-def: $sgpr10 |
| ; SI-NEXT: ; implicit-def: $sgpr89 |
| ; SI-NEXT: s_branch .LBB49_2 |
| ; SI-NEXT: .LBB49_4: |
| ; SI-NEXT: v_mov_b32_e32 v1, s17 |
| ; SI-NEXT: v_mov_b32_e32 v0, s16 |
| ; SI-NEXT: v_mov_b32_e32 v3, s19 |
| ; SI-NEXT: v_mov_b32_e32 v2, s18 |
| ; SI-NEXT: v_mov_b32_e32 v5, s21 |
| ; SI-NEXT: v_mov_b32_e32 v4, s20 |
| ; SI-NEXT: v_mov_b32_e32 v19, s9 |
| ; SI-NEXT: v_mov_b32_e32 v17, s7 |
| ; SI-NEXT: v_mov_b32_e32 v15, s5 |
| ; SI-NEXT: v_mov_b32_e32 v13, s29 |
| ; SI-NEXT: v_mov_b32_e32 v11, s27 |
| ; SI-NEXT: v_mov_b32_e32 v9, s25 |
| ; SI-NEXT: v_mov_b32_e32 v7, s23 |
| ; SI-NEXT: v_mov_b32_e32 v6, s22 |
| ; SI-NEXT: v_mov_b32_e32 v8, s24 |
| ; SI-NEXT: v_mov_b32_e32 v10, s26 |
| ; SI-NEXT: v_mov_b32_e32 v12, s28 |
| ; SI-NEXT: v_mov_b32_e32 v14, s4 |
| ; SI-NEXT: v_mov_b32_e32 v16, s6 |
| ; SI-NEXT: v_mov_b32_e32 v18, s8 |
| ; SI-NEXT: v_mov_b32_e32 v31, s89 |
| ; SI-NEXT: v_mov_b32_e32 v32, s88 |
| ; SI-NEXT: v_mov_b32_e32 v33, s79 |
| ; SI-NEXT: v_mov_b32_e32 v34, s78 |
| ; SI-NEXT: v_mov_b32_e32 v35, s77 |
| ; SI-NEXT: v_mov_b32_e32 v36, s76 |
| ; SI-NEXT: v_mov_b32_e32 v37, s75 |
| ; SI-NEXT: v_mov_b32_e32 v38, s74 |
| ; SI-NEXT: v_mov_b32_e32 v39, s73 |
| ; SI-NEXT: v_mov_b32_e32 v48, s72 |
| ; SI-NEXT: v_mov_b32_e32 v29, s60 |
| ; SI-NEXT: v_mov_b32_e32 v28, s58 |
| ; SI-NEXT: v_mov_b32_e32 v27, s56 |
| ; SI-NEXT: v_mov_b32_e32 v26, s46 |
| ; SI-NEXT: v_mov_b32_e32 v25, s44 |
| ; SI-NEXT: v_mov_b32_e32 v24, s42 |
| ; SI-NEXT: v_mov_b32_e32 v23, s40 |
| ; SI-NEXT: v_mov_b32_e32 v22, s14 |
| ; SI-NEXT: v_mov_b32_e32 v21, s12 |
| ; SI-NEXT: v_mov_b32_e32 v20, s10 |
| ; SI-NEXT: .LBB49_5: ; %end |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v29 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v48 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v28 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v27 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v25 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v35 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v23 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v34 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v22 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v32 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v31 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v29 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v28 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v27 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v26 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v25 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v24 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10f64_to_v40i16_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s5, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB49_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_lshr_b32 s12, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s46, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s5, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s4, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: s_cbranch_execnz .LBB49_4 |
| ; VI-NEXT: .LBB49_2: ; %cmp.true |
| ; VI-NEXT: v_add_f64 v[18:19], s[8:9], 1.0 |
| ; VI-NEXT: v_add_f64 v[16:17], s[6:7], 1.0 |
| ; VI-NEXT: v_add_f64 v[14:15], s[4:5], 1.0 |
| ; VI-NEXT: v_add_f64 v[12:13], s[28:29], 1.0 |
| ; VI-NEXT: v_add_f64 v[10:11], s[26:27], 1.0 |
| ; VI-NEXT: v_add_f64 v[8:9], s[24:25], 1.0 |
| ; VI-NEXT: v_add_f64 v[6:7], s[22:23], 1.0 |
| ; VI-NEXT: v_add_f64 v[4:5], s[20:21], 1.0 |
| ; VI-NEXT: v_add_f64 v[2:3], s[18:19], 1.0 |
| ; VI-NEXT: v_add_f64 v[0:1], s[16:17], 1.0 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; VI-NEXT: s_branch .LBB49_5 |
| ; VI-NEXT: .LBB49_3: |
| ; VI-NEXT: ; implicit-def: $sgpr63 |
| ; VI-NEXT: ; implicit-def: $sgpr45 |
| ; VI-NEXT: ; implicit-def: $sgpr62 |
| ; VI-NEXT: ; implicit-def: $sgpr44 |
| ; VI-NEXT: ; implicit-def: $sgpr61 |
| ; VI-NEXT: ; implicit-def: $sgpr43 |
| ; VI-NEXT: ; implicit-def: $sgpr60 |
| ; VI-NEXT: ; implicit-def: $sgpr42 |
| ; VI-NEXT: ; implicit-def: $sgpr59 |
| ; VI-NEXT: ; implicit-def: $sgpr41 |
| ; VI-NEXT: ; implicit-def: $sgpr58 |
| ; VI-NEXT: ; implicit-def: $sgpr40 |
| ; VI-NEXT: ; implicit-def: $sgpr57 |
| ; VI-NEXT: ; implicit-def: $sgpr15 |
| ; VI-NEXT: ; implicit-def: $sgpr56 |
| ; VI-NEXT: ; implicit-def: $sgpr14 |
| ; VI-NEXT: ; implicit-def: $sgpr47 |
| ; VI-NEXT: ; implicit-def: $sgpr13 |
| ; VI-NEXT: ; implicit-def: $sgpr46 |
| ; VI-NEXT: ; implicit-def: $sgpr12 |
| ; VI-NEXT: s_branch .LBB49_2 |
| ; VI-NEXT: .LBB49_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s16 |
| ; VI-NEXT: v_mov_b32_e32 v2, s18 |
| ; VI-NEXT: v_mov_b32_e32 v4, s20 |
| ; VI-NEXT: v_mov_b32_e32 v6, s22 |
| ; VI-NEXT: v_mov_b32_e32 v8, s24 |
| ; VI-NEXT: v_mov_b32_e32 v10, s26 |
| ; VI-NEXT: v_mov_b32_e32 v12, s28 |
| ; VI-NEXT: v_mov_b32_e32 v14, s4 |
| ; VI-NEXT: v_mov_b32_e32 v16, s6 |
| ; VI-NEXT: v_mov_b32_e32 v18, s8 |
| ; VI-NEXT: v_mov_b32_e32 v1, s17 |
| ; VI-NEXT: v_mov_b32_e32 v3, s19 |
| ; VI-NEXT: v_mov_b32_e32 v5, s21 |
| ; VI-NEXT: v_mov_b32_e32 v7, s23 |
| ; VI-NEXT: v_mov_b32_e32 v9, s25 |
| ; VI-NEXT: v_mov_b32_e32 v11, s27 |
| ; VI-NEXT: v_mov_b32_e32 v13, s29 |
| ; VI-NEXT: v_mov_b32_e32 v15, s5 |
| ; VI-NEXT: v_mov_b32_e32 v17, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s9 |
| ; VI-NEXT: v_mov_b32_e32 v37, s63 |
| ; VI-NEXT: v_mov_b32_e32 v35, s62 |
| ; VI-NEXT: v_mov_b32_e32 v33, s61 |
| ; VI-NEXT: v_mov_b32_e32 v32, s60 |
| ; VI-NEXT: v_mov_b32_e32 v31, s59 |
| ; VI-NEXT: v_mov_b32_e32 v27, s58 |
| ; VI-NEXT: v_mov_b32_e32 v25, s57 |
| ; VI-NEXT: v_mov_b32_e32 v23, s56 |
| ; VI-NEXT: v_mov_b32_e32 v21, s47 |
| ; VI-NEXT: v_mov_b32_e32 v20, s46 |
| ; VI-NEXT: v_mov_b32_e32 v39, s45 |
| ; VI-NEXT: v_mov_b32_e32 v38, s44 |
| ; VI-NEXT: v_mov_b32_e32 v36, s43 |
| ; VI-NEXT: v_mov_b32_e32 v34, s42 |
| ; VI-NEXT: v_mov_b32_e32 v30, s41 |
| ; VI-NEXT: v_mov_b32_e32 v29, s40 |
| ; VI-NEXT: v_mov_b32_e32 v28, s15 |
| ; VI-NEXT: v_mov_b32_e32 v26, s14 |
| ; VI-NEXT: v_mov_b32_e32 v24, s13 |
| ; VI-NEXT: v_mov_b32_e32 v22, s12 |
| ; VI-NEXT: .LBB49_5: ; %end |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v39 |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v38 |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v36 |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v29 |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v28 |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v26 |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v24 |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v22 |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10f64_to_v40i16_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s5, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB49_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_lshr_b32 s12, s9, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s8, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s7, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s6, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s5, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s4, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s58, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s60, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s43, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s62, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s16, 16 |
| ; GFX9-NEXT: s_cbranch_execnz .LBB49_4 |
| ; GFX9-NEXT: .LBB49_2: ; %cmp.true |
| ; GFX9-NEXT: v_add_f64 v[18:19], s[8:9], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[16:17], s[6:7], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[14:15], s[4:5], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[12:13], s[28:29], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[10:11], s[26:27], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[8:9], s[24:25], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[6:7], s[22:23], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[4:5], s[20:21], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[2:3], s[18:19], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[0:1], s[16:17], 1.0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; GFX9-NEXT: s_branch .LBB49_5 |
| ; GFX9-NEXT: .LBB49_3: |
| ; GFX9-NEXT: ; implicit-def: $sgpr63 |
| ; GFX9-NEXT: ; implicit-def: $sgpr45 |
| ; GFX9-NEXT: ; implicit-def: $sgpr62 |
| ; GFX9-NEXT: ; implicit-def: $sgpr44 |
| ; GFX9-NEXT: ; implicit-def: $sgpr61 |
| ; GFX9-NEXT: ; implicit-def: $sgpr43 |
| ; GFX9-NEXT: ; implicit-def: $sgpr60 |
| ; GFX9-NEXT: ; implicit-def: $sgpr42 |
| ; GFX9-NEXT: ; implicit-def: $sgpr59 |
| ; GFX9-NEXT: ; implicit-def: $sgpr41 |
| ; GFX9-NEXT: ; implicit-def: $sgpr58 |
| ; GFX9-NEXT: ; implicit-def: $sgpr40 |
| ; GFX9-NEXT: ; implicit-def: $sgpr57 |
| ; GFX9-NEXT: ; implicit-def: $sgpr15 |
| ; GFX9-NEXT: ; implicit-def: $sgpr56 |
| ; GFX9-NEXT: ; implicit-def: $sgpr14 |
| ; GFX9-NEXT: ; implicit-def: $sgpr47 |
| ; GFX9-NEXT: ; implicit-def: $sgpr13 |
| ; GFX9-NEXT: ; implicit-def: $sgpr46 |
| ; GFX9-NEXT: ; implicit-def: $sgpr12 |
| ; GFX9-NEXT: s_branch .LBB49_2 |
| ; GFX9-NEXT: .LBB49_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s28 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s4 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s6 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s29 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s5 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v33, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v32, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s15 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s14 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s13 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s12 |
| ; GFX9-NEXT: .LBB49_5: ; %end |
| ; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX9-NEXT: v_lshl_or_b32 v0, v37, 16, v0 |
| ; GFX9-NEXT: v_lshl_or_b32 v1, v39, 16, v1 |
| ; GFX9-NEXT: v_lshl_or_b32 v2, v35, 16, v2 |
| ; GFX9-NEXT: v_lshl_or_b32 v3, v38, 16, v3 |
| ; GFX9-NEXT: v_lshl_or_b32 v4, v33, 16, v4 |
| ; GFX9-NEXT: v_lshl_or_b32 v5, v36, 16, v5 |
| ; GFX9-NEXT: v_lshl_or_b32 v6, v32, 16, v6 |
| ; GFX9-NEXT: v_lshl_or_b32 v7, v34, 16, v7 |
| ; GFX9-NEXT: v_lshl_or_b32 v8, v31, 16, v8 |
| ; GFX9-NEXT: v_lshl_or_b32 v9, v30, 16, v9 |
| ; GFX9-NEXT: v_lshl_or_b32 v10, v27, 16, v10 |
| ; GFX9-NEXT: v_lshl_or_b32 v11, v29, 16, v11 |
| ; GFX9-NEXT: v_lshl_or_b32 v12, v25, 16, v12 |
| ; GFX9-NEXT: v_lshl_or_b32 v13, v28, 16, v13 |
| ; GFX9-NEXT: v_lshl_or_b32 v14, v23, 16, v14 |
| ; GFX9-NEXT: v_lshl_or_b32 v15, v26, 16, v15 |
| ; GFX9-NEXT: v_lshl_or_b32 v16, v21, 16, v16 |
| ; GFX9-NEXT: v_lshl_or_b32 v17, v24, 16, v17 |
| ; GFX9-NEXT: v_lshl_or_b32 v18, v20, 16, v18 |
| ; GFX9-NEXT: v_lshl_or_b32 v19, v22, 16, v19 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v10f64_to_v40i16_scalar: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v1 |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 |
| ; GFX11-TRUE16-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB49_3 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s5, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s4, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s29, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s27, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s26, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s25, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s24, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s22, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s21, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s20, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s19, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s47, s18, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s17, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s56, s16, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s3, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s57, s2, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s1, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s58, s0, 16 |
| ; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB49_4 |
| ; GFX11-TRUE16-NEXT: .LBB49_2: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[18:19], s[4:5], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[16:17], s[28:29], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[14:15], s[26:27], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[12:13], s[24:25], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[10:11], s[22:23], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[8:9], s[20:21], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[6:7], s[18:19], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[4:5], s[16:17], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[2:3], s[2:3], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[0:1], s[0:1], 1.0 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 16, v19 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 16, v18 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v17 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 16, v16 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v15 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 16, v14 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v13 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v12 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v11 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v10 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v9 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v8 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v7 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v6 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v5 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v2 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; GFX11-TRUE16-NEXT: s_branch .LBB49_5 |
| ; GFX11-TRUE16-NEXT: .LBB49_3: |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr58 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr40 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr57 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr15 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr56 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr14 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr47 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr13 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr46 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr12 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr45 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr11 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr44 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr10 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr43 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr9 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr8 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr41 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr7 |
| ; GFX11-TRUE16-NEXT: s_branch .LBB49_2 |
| ; GFX11-TRUE16-NEXT: .LBB49_4: |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v37, s58 :: v_dual_mov_b32 v20, s41 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v35, s57 :: v_dual_mov_b32 v38, s15 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v33, s56 :: v_dual_mov_b32 v36, s14 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, s47 :: v_dual_mov_b32 v34, s13 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, s46 :: v_dual_mov_b32 v32, s12 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, s45 :: v_dual_mov_b32 v30, s11 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, s44 :: v_dual_mov_b32 v28, s10 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, s43 :: v_dual_mov_b32 v26, s9 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, s42 :: v_dual_mov_b32 v24, s8 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v39, s40 :: v_dual_mov_b32 v22, s7 |
| ; GFX11-TRUE16-NEXT: .LBB49_5: ; %end |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v37.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v39.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v35.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v38.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v33.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v36.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v31.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v34.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v29.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v32.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v27.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v30.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v25.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v28.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v23.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v26.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v21.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v24.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v20.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v22.l |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v10f64_to_v40i16_scalar: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s5, v1 |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s4, v0 |
| ; GFX11-FAKE16-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-FAKE16-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB49_3 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s5, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s4, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s29, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s27, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s26, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s25, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s24, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s23, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s22, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s21, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s20, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s19, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s47, s18, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s17, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s56, s16, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s3, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s57, s2, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s40, s1, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s58, s0, 16 |
| ; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB49_4 |
| ; GFX11-FAKE16-NEXT: .LBB49_2: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[15:16], s[4:5], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[17:18], s[28:29], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[19:20], s[26:27], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[10:11], s[24:25], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[12:13], s[22:23], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[5:6], s[20:21], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[7:8], s[18:19], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[21:22], s[16:17], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[0:1], s[2:3], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[2:3], s[0:1], 1.0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v14, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v9, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v22 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v4, 16, v21 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: s_branch .LBB49_5 |
| ; GFX11-FAKE16-NEXT: .LBB49_3: |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr58 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr40 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr57 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr15 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr56 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr14 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr47 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr13 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr46 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr12 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr45 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr11 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr44 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr10 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr43 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr9 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr8 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr41 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr7 |
| ; GFX11-FAKE16-NEXT: s_branch .LBB49_2 |
| ; GFX11-FAKE16-NEXT: .LBB49_4: |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v21, s16 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v7, s18 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, s20 :: v_dual_mov_b32 v12, s22 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v19, s26 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v22, s17 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v15, s4 :: v_dual_mov_b32 v8, s19 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v6, s21 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v1, s3 :: v_dual_mov_b32 v20, s27 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v13, s23 :: v_dual_mov_b32 v18, s29 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s25 :: v_dual_mov_b32 v16, s5 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, s58 :: v_dual_mov_b32 v36, s57 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s56 :: v_dual_mov_b32 v9, s46 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v32, s47 :: v_dual_mov_b32 v29, s45 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v28, s44 :: v_dual_mov_b32 v23, s41 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s43 :: v_dual_mov_b32 v39, s40 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s42 :: v_dual_mov_b32 v35, s14 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v38, s15 :: v_dual_mov_b32 v33, s12 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v34, s13 :: v_dual_mov_b32 v31, s11 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v30, s10 :: v_dual_mov_b32 v27, s9 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v26, s8 :: v_dual_mov_b32 v25, s7 |
| ; GFX11-FAKE16-NEXT: .LBB49_5: ; %end |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v0 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v4, 16, v21 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v22 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v5 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v35, 16, v21 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v10 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v1 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v36, 16, v48 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xffff, v6 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v34, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v9, 16, v22 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v11 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v29, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v28, 16, v21 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v14, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v20 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v15 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v16 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v39, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v38, 16, v49 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v33, 16, v36 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v31, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v30, 16, v22 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v27, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v24, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v26, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v23, 16, v20 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v25, 16, v21 |
| ; GFX11-FAKE16-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 <10 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <10 x double> %a1 to <40 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x double> %a to <40 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <40 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x i16> %phi |
| } |
| |
| define <10 x double> @bitcast_v40i16_to_v10f64(<40 x i16> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v40i16_to_v10f64: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: v_mov_b32_e32 v32, v19 |
| ; SI-NEXT: v_mov_b32_e32 v33, v18 |
| ; SI-NEXT: v_mov_b32_e32 v34, v17 |
| ; SI-NEXT: v_mov_b32_e32 v35, v16 |
| ; SI-NEXT: v_mov_b32_e32 v36, v15 |
| ; SI-NEXT: v_mov_b32_e32 v37, v14 |
| ; SI-NEXT: v_mov_b32_e32 v38, v13 |
| ; SI-NEXT: v_mov_b32_e32 v39, v12 |
| ; SI-NEXT: v_mov_b32_e32 v48, v11 |
| ; SI-NEXT: v_mov_b32_e32 v49, v10 |
| ; SI-NEXT: v_mov_b32_e32 v50, v9 |
| ; SI-NEXT: v_mov_b32_e32 v51, v8 |
| ; SI-NEXT: v_mov_b32_e32 v52, v7 |
| ; SI-NEXT: v_mov_b32_e32 v53, v6 |
| ; SI-NEXT: v_mov_b32_e32 v54, v5 |
| ; SI-NEXT: v_mov_b32_e32 v55, v4 |
| ; SI-NEXT: v_mov_b32_e32 v40, v3 |
| ; SI-NEXT: v_mov_b32_e32 v41, v2 |
| ; SI-NEXT: v_mov_b32_e32 v42, v1 |
| ; SI-NEXT: v_mov_b32_e32 v43, v0 |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v33 |
| ; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v34 |
| ; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v35 |
| ; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v36 |
| ; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v37 |
| ; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v38 |
| ; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v39 |
| ; SI-NEXT: v_lshrrev_b32_e32 v8, 16, v48 |
| ; SI-NEXT: v_lshrrev_b32_e32 v9, 16, v49 |
| ; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v50 |
| ; SI-NEXT: v_lshrrev_b32_e32 v11, 16, v51 |
| ; SI-NEXT: v_lshrrev_b32_e32 v12, 16, v52 |
| ; SI-NEXT: v_lshrrev_b32_e32 v13, 16, v53 |
| ; SI-NEXT: v_lshrrev_b32_e32 v14, 16, v54 |
| ; SI-NEXT: v_lshrrev_b32_e32 v15, 16, v55 |
| ; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v40 |
| ; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v41 |
| ; SI-NEXT: v_lshrrev_b32_e32 v18, 16, v42 |
| ; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v43 |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: s_waitcnt expcnt(4) |
| ; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v56, 16, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v47, 16, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v12 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB50_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v39 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v43 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v42 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v41 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v40 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v55 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v54 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v53 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v52 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v51 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v50 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v49 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v48 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v59 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v58 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v57 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v56 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v47 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v46 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v45 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v44 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v63 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v62 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v61 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v60 |
| ; SI-NEXT: ; implicit-def: $vgpr43 |
| ; SI-NEXT: ; implicit-def: $vgpr42 |
| ; SI-NEXT: ; implicit-def: $vgpr41 |
| ; SI-NEXT: ; implicit-def: $vgpr40 |
| ; SI-NEXT: ; implicit-def: $vgpr55 |
| ; SI-NEXT: ; implicit-def: $vgpr54 |
| ; SI-NEXT: ; implicit-def: $vgpr53 |
| ; SI-NEXT: ; implicit-def: $vgpr52 |
| ; SI-NEXT: ; implicit-def: $vgpr51 |
| ; SI-NEXT: ; implicit-def: $vgpr50 |
| ; SI-NEXT: ; implicit-def: $vgpr49 |
| ; SI-NEXT: ; implicit-def: $vgpr48 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr59 |
| ; SI-NEXT: ; implicit-def: $vgpr58 |
| ; SI-NEXT: ; implicit-def: $vgpr57 |
| ; SI-NEXT: ; implicit-def: $vgpr56 |
| ; SI-NEXT: ; implicit-def: $vgpr47 |
| ; SI-NEXT: ; implicit-def: $vgpr46 |
| ; SI-NEXT: ; implicit-def: $vgpr45 |
| ; SI-NEXT: ; implicit-def: $vgpr44 |
| ; SI-NEXT: ; implicit-def: $vgpr63 |
| ; SI-NEXT: ; implicit-def: $vgpr62 |
| ; SI-NEXT: ; implicit-def: $vgpr61 |
| ; SI-NEXT: ; implicit-def: $vgpr60 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v38 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v37 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v36 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v35 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v34 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v33 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v19 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v32 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: .LBB50_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB50_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v39 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v43 |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v42 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v41 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v40 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v55 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v54 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v53 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v52 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v51 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v50 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v49 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v48 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_or_b32_e32 v0, v59, v0 |
| ; SI-NEXT: s_mov_b32 s6, 0x30000 |
| ; SI-NEXT: v_or_b32_e32 v1, v58, v1 |
| ; SI-NEXT: v_or_b32_e32 v2, v57, v2 |
| ; SI-NEXT: v_or_b32_e32 v3, v56, v3 |
| ; SI-NEXT: v_or_b32_e32 v4, v47, v4 |
| ; SI-NEXT: v_or_b32_e32 v5, v46, v5 |
| ; SI-NEXT: v_or_b32_e32 v6, v45, v6 |
| ; SI-NEXT: v_or_b32_e32 v7, v44, v7 |
| ; SI-NEXT: v_or_b32_e32 v8, v63, v8 |
| ; SI-NEXT: v_or_b32_e32 v9, v62, v9 |
| ; SI-NEXT: v_or_b32_e32 v10, v61, v10 |
| ; SI-NEXT: v_or_b32_e32 v11, v60, v11 |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0 |
| ; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v2 |
| ; SI-NEXT: v_add_i32_e32 v3, vcc, s6, v3 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, s6, v4 |
| ; SI-NEXT: v_add_i32_e32 v5, vcc, s6, v5 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, s6, v6 |
| ; SI-NEXT: v_add_i32_e32 v7, vcc, s6, v7 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8 |
| ; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v38 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_or_b32_e32 v13, v14, v13 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v37 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_or_b32_e32 v14, v15, v14 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v36 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v35 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v34 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_or_b32_e32 v17, v18, v17 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v33 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_or_b32_e32 v18, v19, v18 |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v32 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_or_b32_e32 v19, v20, v19 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v10 |
| ; SI-NEXT: v_add_i32_e32 v11, vcc, s6, v11 |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, s6, v12 |
| ; SI-NEXT: v_add_i32_e32 v13, vcc, s6, v13 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v14 |
| ; SI-NEXT: v_add_i32_e32 v15, vcc, s6, v15 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, s6, v16 |
| ; SI-NEXT: v_add_i32_e32 v17, vcc, s6, v17 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v18 |
| ; SI-NEXT: v_add_i32_e32 v19, vcc, 0x30000, v19 |
| ; SI-NEXT: .LBB50_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40i16_to_v10f64: |
| ; 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:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: v_mov_b32_e32 v32, v19 |
| ; VI-NEXT: v_mov_b32_e32 v33, v18 |
| ; VI-NEXT: v_mov_b32_e32 v34, v17 |
| ; VI-NEXT: v_mov_b32_e32 v35, v16 |
| ; VI-NEXT: v_mov_b32_e32 v36, v15 |
| ; VI-NEXT: v_mov_b32_e32 v37, v14 |
| ; VI-NEXT: v_mov_b32_e32 v38, v13 |
| ; VI-NEXT: v_mov_b32_e32 v39, v12 |
| ; VI-NEXT: v_mov_b32_e32 v48, v11 |
| ; VI-NEXT: v_mov_b32_e32 v49, v10 |
| ; VI-NEXT: v_mov_b32_e32 v50, v9 |
| ; VI-NEXT: v_mov_b32_e32 v51, v8 |
| ; VI-NEXT: v_mov_b32_e32 v52, v7 |
| ; VI-NEXT: v_mov_b32_e32 v53, v6 |
| ; VI-NEXT: v_mov_b32_e32 v54, v5 |
| ; VI-NEXT: v_mov_b32_e32 v55, v4 |
| ; VI-NEXT: v_mov_b32_e32 v40, v3 |
| ; VI-NEXT: v_mov_b32_e32 v41, v2 |
| ; VI-NEXT: v_mov_b32_e32 v42, v1 |
| ; VI-NEXT: v_mov_b32_e32 v43, v0 |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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 .LBB50_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: v_mov_b32_e32 v19, 16 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v0, v19, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v1, v19, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v2, v19, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v3, v19, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v4, v19, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v5, v19, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v6, v19, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v7, v19, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v8, v19, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v9, v19, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v10, v19, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v11, v19, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v12, v19, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v13, v19, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v14, v19, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v15, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v16, v19, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v17, v19, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v18, v19, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v19, v19, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v40, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v55, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v54, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v53, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v52, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v51, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v50, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v49, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v48, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v39, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v38, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v37, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v36, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v35, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v34, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v33, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v32, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; 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: .LBB50_2: ; %Flow |
| ; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; VI-NEXT: s_cbranch_execz .LBB50_4 |
| ; VI-NEXT: ; %bb.3: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v19, 3 |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v43 |
| ; VI-NEXT: v_add_u16_sdwa v1, v43, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v42 |
| ; VI-NEXT: v_add_u16_sdwa v2, v42, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v1, v1, v2 |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v41 |
| ; VI-NEXT: v_add_u16_sdwa v3, v41, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v40 |
| ; VI-NEXT: v_add_u16_sdwa v4, v40, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v55 |
| ; VI-NEXT: v_add_u16_sdwa v5, v55, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v54 |
| ; VI-NEXT: v_add_u16_sdwa v6, v54, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v53 |
| ; VI-NEXT: v_add_u16_sdwa v7, v53, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v52 |
| ; VI-NEXT: v_add_u16_sdwa v8, v52, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v51 |
| ; VI-NEXT: v_add_u16_sdwa v9, v51, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v50 |
| ; VI-NEXT: v_add_u16_sdwa v10, v50, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v49 |
| ; VI-NEXT: v_add_u16_sdwa v11, v49, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v48 |
| ; VI-NEXT: v_add_u16_sdwa v12, v48, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v39 |
| ; VI-NEXT: v_add_u16_sdwa v13, v39, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v38 |
| ; VI-NEXT: v_add_u16_sdwa v14, v38, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v37 |
| ; VI-NEXT: v_add_u16_sdwa v15, v37, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; VI-NEXT: v_add_u16_e32 v15, 3, v36 |
| ; VI-NEXT: v_add_u16_sdwa v16, v36, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v35 |
| ; VI-NEXT: v_add_u16_sdwa v17, v35, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; VI-NEXT: v_add_u16_e32 v17, 3, v34 |
| ; VI-NEXT: v_add_u16_sdwa v18, v34, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; VI-NEXT: v_add_u16_e32 v18, 3, v33 |
| ; VI-NEXT: v_add_u16_sdwa v20, v33, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; VI-NEXT: v_add_u16_e32 v20, 3, v32 |
| ; VI-NEXT: v_add_u16_sdwa v19, v32, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_e32 v19, v20, v19 |
| ; VI-NEXT: .LBB50_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40i16_to_v10f64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_mov_b32_e32 v32, v19 |
| ; 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: v_mov_b32_e32 v33, v18 |
| ; GFX9-NEXT: v_mov_b32_e32 v43, v0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, v17 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, v16 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, v15 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, v14 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, v13 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, v12 |
| ; GFX9-NEXT: v_mov_b32_e32 v48, v11 |
| ; GFX9-NEXT: v_mov_b32_e32 v49, v10 |
| ; GFX9-NEXT: v_mov_b32_e32 v50, v9 |
| ; GFX9-NEXT: v_mov_b32_e32 v51, v8 |
| ; GFX9-NEXT: v_mov_b32_e32 v52, v7 |
| ; GFX9-NEXT: v_mov_b32_e32 v53, v6 |
| ; GFX9-NEXT: v_mov_b32_e32 v54, v5 |
| ; GFX9-NEXT: v_mov_b32_e32 v55, v4 |
| ; GFX9-NEXT: v_mov_b32_e32 v40, v3 |
| ; GFX9-NEXT: v_mov_b32_e32 v41, v2 |
| ; GFX9-NEXT: v_mov_b32_e32 v42, v1 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v48 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v49 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v50 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v51 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v52 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v53 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v54 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v55 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43 |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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 .LBB50_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v35 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v34 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v33 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v32 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; 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: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; GFX9-NEXT: .LBB50_2: ; %Flow |
| ; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; GFX9-NEXT: s_cbranch_execz .LBB50_4 |
| ; GFX9-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_pk_add_u16 v0, v0, 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 v2, v2, 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 v4, v4, 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 v6, v6, 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 v8, v8, 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 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; GFX9-NEXT: v_pk_add_u16 v12, v12, 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 v14, v14, 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 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB50_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; 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-TRUE16-LABEL: bitcast_v40i16_to_v10f64: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: .LBB50_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v40i16_to_v10f64: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v37, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v39, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v48, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v36, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v35, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v34, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v33, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v32, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v31, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v30, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v29, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v28, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v27, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v26, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v25, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v24, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v23, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v22, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v21, v19, 0x5040100 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB50_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: .LBB50_2: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-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 <40 x i16> %a, splat (i16 3) |
| %a2 = bitcast <40 x i16> %a1 to <10 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x i16> %a to <10 x double> |
| br label %end |
| |
| end: |
| %phi = phi <10 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x double> %phi |
| } |
| |
| define inreg <10 x double> @bitcast_v40i16_to_v10f64_scalar(<40 x i16> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v40i16_to_v10f64_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v20, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v20, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v20, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v20, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v20, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v20, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v20, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v20, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v20, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v20, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v20, s54, 10 |
| ; SI-NEXT: v_writelane_b32 v20, s55, 11 |
| ; SI-NEXT: v_writelane_b32 v20, s64, 12 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s12, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s15, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s74, v1 |
| ; SI-NEXT: v_readfirstlane_b32 s77, v0 |
| ; SI-NEXT: v_writelane_b32 v20, s65, 13 |
| ; SI-NEXT: s_lshr_b32 s11, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s13, s28, 16 |
| ; SI-NEXT: s_lshr_b32 s72, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s26, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; SI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; SI-NEXT: s_lshr_b32 s92, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s93, s18, 16 |
| ; SI-NEXT: s_lshr_b32 s94, s17, 16 |
| ; SI-NEXT: s_lshr_b32 s95, s16, 16 |
| ; SI-NEXT: s_lshr_b32 s6, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s8, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s10, s12, 16 |
| ; SI-NEXT: s_lshr_b32 s14, s15, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s74, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s77, 16 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v20, s66, 14 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_writelane_b32 v20, s67, 15 |
| ; SI-NEXT: s_cbranch_scc0 .LBB51_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s95, 16 |
| ; SI-NEXT: s_or_b32 s36, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s94, 16 |
| ; SI-NEXT: s_or_b32 s37, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s93, 16 |
| ; SI-NEXT: s_or_b32 s38, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s92, 16 |
| ; SI-NEXT: s_or_b32 s39, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; SI-NEXT: s_or_b32 s40, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; SI-NEXT: s_or_b32 s41, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; SI-NEXT: s_or_b32 s42, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; SI-NEXT: s_or_b32 s43, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; SI-NEXT: s_or_b32 s44, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; SI-NEXT: s_or_b32 s45, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; SI-NEXT: s_or_b32 s46, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; SI-NEXT: s_or_b32 s47, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; SI-NEXT: s_or_b32 s48, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; SI-NEXT: s_or_b32 s49, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s77, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; SI-NEXT: s_or_b32 s50, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s74, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; SI-NEXT: s_or_b32 s51, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s15, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; SI-NEXT: s_or_b32 s52, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; SI-NEXT: s_or_b32 s53, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; SI-NEXT: s_or_b32 s54, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; SI-NEXT: s_or_b32 s55, s4, s5 |
| ; SI-NEXT: s_cbranch_execnz .LBB51_3 |
| ; SI-NEXT: .LBB51_2: ; %cmp.true |
| ; SI-NEXT: s_add_i32 s16, s16, 3 |
| ; SI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s95, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s17, s17, 3 |
| ; SI-NEXT: s_add_i32 s36, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s94, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s18, s18, 3 |
| ; SI-NEXT: s_add_i32 s37, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s93, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s19, s19, 3 |
| ; SI-NEXT: s_add_i32 s38, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s92, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s20, s20, 3 |
| ; SI-NEXT: s_add_i32 s39, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s21, s21, 3 |
| ; SI-NEXT: s_add_i32 s40, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s22, s22, 3 |
| ; SI-NEXT: s_add_i32 s41, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s23, s23, 3 |
| ; SI-NEXT: s_add_i32 s42, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s24, s24, 3 |
| ; SI-NEXT: s_add_i32 s43, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s25, s25, 3 |
| ; SI-NEXT: s_add_i32 s44, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s26, s26, 3 |
| ; SI-NEXT: s_add_i32 s45, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s27, s27, 3 |
| ; SI-NEXT: s_add_i32 s46, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s28, s28, 3 |
| ; SI-NEXT: s_add_i32 s47, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s29, s29, 3 |
| ; SI-NEXT: s_add_i32 s48, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s77, s77, 3 |
| ; SI-NEXT: s_add_i32 s49, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s77, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s74, s74, 3 |
| ; SI-NEXT: s_add_i32 s50, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s74, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s15, s15, 3 |
| ; SI-NEXT: s_add_i32 s51, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s15, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s12, s12, 3 |
| ; SI-NEXT: s_add_i32 s52, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s9, s9, 3 |
| ; SI-NEXT: s_add_i32 s53, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s7, s7, 3 |
| ; SI-NEXT: s_add_i32 s54, s4, 0x30000 |
| ; SI-NEXT: s_and_b32 s4, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_add_i32 s55, s4, 0x30000 |
| ; SI-NEXT: .LBB51_3: ; %end |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_readlane_b32 s67, v20, 15 |
| ; SI-NEXT: v_readlane_b32 s66, v20, 14 |
| ; SI-NEXT: v_readlane_b32 s65, v20, 13 |
| ; SI-NEXT: v_readlane_b32 s64, v20, 12 |
| ; SI-NEXT: v_readlane_b32 s55, v20, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v20, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v20, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v20, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v20, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v20, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v20, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v20, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v20, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v20, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v20, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v20, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB51_4: |
| ; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; SI-NEXT: s_branch .LBB51_2 |
| ; |
| ; VI-LABEL: bitcast_v40i16_to_v10f64_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v20, s30, 0 |
| ; VI-NEXT: v_writelane_b32 v20, s31, 1 |
| ; VI-NEXT: v_writelane_b32 v20, s34, 2 |
| ; VI-NEXT: v_writelane_b32 v20, s35, 3 |
| ; VI-NEXT: v_writelane_b32 v20, s36, 4 |
| ; VI-NEXT: v_writelane_b32 v20, s37, 5 |
| ; VI-NEXT: v_writelane_b32 v20, s38, 6 |
| ; VI-NEXT: v_writelane_b32 v20, s39, 7 |
| ; VI-NEXT: v_writelane_b32 v20, s48, 8 |
| ; VI-NEXT: v_writelane_b32 v20, s49, 9 |
| ; VI-NEXT: v_writelane_b32 v20, s50, 10 |
| ; VI-NEXT: v_writelane_b32 v20, s51, 11 |
| ; VI-NEXT: v_writelane_b32 v20, s52, 12 |
| ; VI-NEXT: v_writelane_b32 v20, s53, 13 |
| ; VI-NEXT: v_writelane_b32 v20, s54, 14 |
| ; VI-NEXT: v_writelane_b32 v20, s55, 15 |
| ; VI-NEXT: v_writelane_b32 v20, s64, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s12, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s15, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s74, v1 |
| ; VI-NEXT: v_readfirstlane_b32 s77, v0 |
| ; VI-NEXT: v_writelane_b32 v20, s65, 17 |
| ; VI-NEXT: s_lshr_b32 s11, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s72, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s75, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s78, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s30, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s31, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s34, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s35, s16, 16 |
| ; VI-NEXT: s_lshr_b32 s6, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s8, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s10, s12, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s15, 16 |
| ; VI-NEXT: s_lshr_b32 s73, s74, 16 |
| ; VI-NEXT: s_lshr_b32 s76, s77, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v20, s66, 18 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_writelane_b32 v20, s67, 19 |
| ; VI-NEXT: s_cbranch_scc0 .LBB51_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s35, 16 |
| ; VI-NEXT: s_or_b32 s36, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s5, s34, 16 |
| ; VI-NEXT: s_or_b32 s37, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s5, s31, 16 |
| ; VI-NEXT: s_or_b32 s38, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s5, s30, 16 |
| ; VI-NEXT: s_or_b32 s39, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; VI-NEXT: s_or_b32 s40, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; VI-NEXT: s_or_b32 s41, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; VI-NEXT: s_or_b32 s42, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; VI-NEXT: s_or_b32 s43, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; VI-NEXT: s_or_b32 s44, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; VI-NEXT: s_or_b32 s45, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; VI-NEXT: s_or_b32 s46, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; VI-NEXT: s_or_b32 s47, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; VI-NEXT: s_or_b32 s48, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; VI-NEXT: s_or_b32 s49, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s77 |
| ; VI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; VI-NEXT: s_or_b32 s50, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s74 |
| ; VI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; VI-NEXT: s_or_b32 s51, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s15 |
| ; VI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; VI-NEXT: s_or_b32 s52, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s12 |
| ; VI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; VI-NEXT: s_or_b32 s53, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s9 |
| ; VI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; VI-NEXT: s_or_b32 s54, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s7 |
| ; VI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; VI-NEXT: s_or_b32 s55, s4, s5 |
| ; VI-NEXT: s_cbranch_execnz .LBB51_3 |
| ; VI-NEXT: .LBB51_2: ; %cmp.true |
| ; VI-NEXT: s_add_i32 s16, s16, 3 |
| ; VI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s35, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s17, s17, 3 |
| ; VI-NEXT: s_add_i32 s36, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s34, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s18, s18, 3 |
| ; VI-NEXT: s_add_i32 s37, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s31, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s19, s19, 3 |
| ; VI-NEXT: s_add_i32 s38, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s30, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s20, s20, 3 |
| ; VI-NEXT: s_add_i32 s39, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s21, s21, 3 |
| ; VI-NEXT: s_add_i32 s40, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s22, s22, 3 |
| ; VI-NEXT: s_add_i32 s41, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s23, s23, 3 |
| ; VI-NEXT: s_add_i32 s42, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s24, s24, 3 |
| ; VI-NEXT: s_add_i32 s43, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s25, s25, 3 |
| ; VI-NEXT: s_add_i32 s44, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s26, s26, 3 |
| ; VI-NEXT: s_add_i32 s45, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s27, s27, 3 |
| ; VI-NEXT: s_add_i32 s46, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s28, s28, 3 |
| ; VI-NEXT: s_add_i32 s47, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s29, s29, 3 |
| ; VI-NEXT: s_add_i32 s48, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s77, s77, 3 |
| ; VI-NEXT: s_add_i32 s49, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s77, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s74, s74, 3 |
| ; VI-NEXT: s_add_i32 s50, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s74, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s73, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s15, s15, 3 |
| ; VI-NEXT: s_add_i32 s51, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s15, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s14, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s12, s12, 3 |
| ; VI-NEXT: s_add_i32 s52, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s9, s9, 3 |
| ; VI-NEXT: s_add_i32 s53, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s9, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s8, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s7, s7, 3 |
| ; VI-NEXT: s_add_i32 s54, s4, 0x30000 |
| ; VI-NEXT: s_and_b32 s4, s7, 0xffff |
| ; VI-NEXT: s_lshl_b32 s5, s6, 16 |
| ; VI-NEXT: s_or_b32 s4, s5, s4 |
| ; VI-NEXT: s_add_i32 s55, s4, 0x30000 |
| ; VI-NEXT: .LBB51_3: ; %end |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_readlane_b32 s67, v20, 19 |
| ; VI-NEXT: v_readlane_b32 s66, v20, 18 |
| ; VI-NEXT: v_readlane_b32 s65, v20, 17 |
| ; VI-NEXT: v_readlane_b32 s64, v20, 16 |
| ; VI-NEXT: v_readlane_b32 s55, v20, 15 |
| ; VI-NEXT: v_readlane_b32 s54, v20, 14 |
| ; VI-NEXT: v_readlane_b32 s53, v20, 13 |
| ; VI-NEXT: v_readlane_b32 s52, v20, 12 |
| ; VI-NEXT: v_readlane_b32 s51, v20, 11 |
| ; VI-NEXT: v_readlane_b32 s50, v20, 10 |
| ; VI-NEXT: v_readlane_b32 s49, v20, 9 |
| ; VI-NEXT: v_readlane_b32 s48, v20, 8 |
| ; VI-NEXT: v_readlane_b32 s39, v20, 7 |
| ; VI-NEXT: v_readlane_b32 s38, v20, 6 |
| ; VI-NEXT: v_readlane_b32 s37, v20, 5 |
| ; VI-NEXT: v_readlane_b32 s36, v20, 4 |
| ; VI-NEXT: v_readlane_b32 s35, v20, 3 |
| ; VI-NEXT: v_readlane_b32 s34, v20, 2 |
| ; VI-NEXT: v_readlane_b32 s31, v20, 1 |
| ; VI-NEXT: v_readlane_b32 s30, v20, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB51_4: |
| ; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; VI-NEXT: s_branch .LBB51_2 |
| ; |
| ; GFX9-LABEL: bitcast_v40i16_to_v10f64_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_readfirstlane_b32 s56, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s58, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s60, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s62, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s72, v1 |
| ; GFX9-NEXT: v_readfirstlane_b32 s74, v0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: s_lshr_b32 s4, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s5, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s6, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s7, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s8, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s9, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s10, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s11, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s12, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s16, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s56, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s58, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s60, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s62, 16 |
| ; GFX9-NEXT: s_lshr_b32 s73, s72, 16 |
| ; GFX9-NEXT: s_lshr_b32 s75, s74, 16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s42, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_cmp_lg_u32 s42, 0 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s50, s74, s75 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s51, s72, s73 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s52, s62, s63 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s53, s60, s61 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s54, s58, s59 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s55, s56, s57 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB51_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB51_4 |
| ; GFX9-NEXT: .LBB51_2: ; %cmp.true |
| ; GFX9-NEXT: v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_branch .LBB51_5 |
| ; GFX9-NEXT: .LBB51_3: |
| ; GFX9-NEXT: s_branch .LBB51_2 |
| ; GFX9-NEXT: .LBB51_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB51_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v40i16_to_v10f64_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s45, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s46, v0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s56, v2 |
| ; GFX11-NEXT: s_lshr_b32 s41, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s6, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s5, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s4, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: s_lshr_b32 s58, s45, 16 |
| ; GFX11-NEXT: s_lshr_b32 s59, s46, 16 |
| ; GFX11-NEXT: s_mov_b32 s40, 0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s56, 0 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s57 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s47 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s44 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s18, s46, s59 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s19, s45, s58 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB51_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB51_4 |
| ; GFX11-NEXT: .LBB51_2: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v18, s18, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: v_pk_add_u16 v19, s19, 3 op_sel_hi:[1,0] |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB51_3: |
| ; GFX11-NEXT: s_branch .LBB51_2 |
| ; GFX11-NEXT: .LBB51_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31 |
| ; 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 <40 x i16> %a, splat (i16 3) |
| %a2 = bitcast <40 x i16> %a1 to <10 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x i16> %a to <10 x double> |
| br label %end |
| |
| end: |
| %phi = phi <10 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x double> %phi |
| } |
| |
| define <40 x half> @bitcast_v10f64_to_v40f16(<10 x double> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v10f64_to_v40f16: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr29 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr26 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr25 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr24 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr23 |
| ; SI-NEXT: ; implicit-def: $vgpr31 |
| ; SI-NEXT: ; implicit-def: $vgpr22 |
| ; SI-NEXT: ; implicit-def: $vgpr30 |
| ; SI-NEXT: ; implicit-def: $vgpr21 |
| ; SI-NEXT: ; implicit-def: $vgpr28 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr27 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB52_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v32, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v35, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB52_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB52_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; SI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; SI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; SI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; SI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; SI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; SI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; SI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; SI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0 |
| ; SI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; SI-NEXT: v_alignbit_b32 v20, v19, v18, 16 |
| ; SI-NEXT: v_alignbit_b32 v21, v17, v16, 16 |
| ; SI-NEXT: v_alignbit_b32 v22, v15, v14, 16 |
| ; SI-NEXT: v_alignbit_b32 v23, v13, v12, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v11, v10, 16 |
| ; SI-NEXT: v_alignbit_b32 v25, v9, v8, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v7, v6, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v5, v4, 16 |
| ; SI-NEXT: v_alignbit_b32 v32, v3, v2, 16 |
| ; SI-NEXT: v_alignbit_b32 v35, v1, v0, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; SI-NEXT: .LBB52_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v35 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v32 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v29 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v25 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v34 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v23 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v31 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v22 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v30 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v28 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v27 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v35 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v32 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v29 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v26 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v25 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v24 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10f64_to_v40f16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; VI-NEXT: ; implicit-def: $vgpr28 |
| ; VI-NEXT: ; implicit-def: $vgpr27 |
| ; VI-NEXT: ; implicit-def: $vgpr26 |
| ; 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: 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 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; 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_add_f64 v[18:19], v[18:19], 1.0 |
| ; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; 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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; VI-NEXT: .LBB52_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10f64_to_v40f16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; 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: ; implicit-def: $vgpr29 |
| ; GFX9-NEXT: ; implicit-def: $vgpr28 |
| ; GFX9-NEXT: ; implicit-def: $vgpr27 |
| ; GFX9-NEXT: ; implicit-def: $vgpr26 |
| ; GFX9-NEXT: ; implicit-def: $vgpr25 |
| ; GFX9-NEXT: ; implicit-def: $vgpr24 |
| ; GFX9-NEXT: ; implicit-def: $vgpr23 |
| ; GFX9-NEXT: ; implicit-def: $vgpr22 |
| ; GFX9-NEXT: ; implicit-def: $vgpr21 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; 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 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; 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: v_add_f64 v[18:19], v[18:19], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; 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: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: .LBB52_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_mov_b32 s4, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v39, v0, s4 |
| ; GFX9-NEXT: v_perm_b32 v1, v38, v1, s4 |
| ; GFX9-NEXT: v_perm_b32 v2, v37, v2, s4 |
| ; GFX9-NEXT: v_perm_b32 v3, v36, v3, s4 |
| ; GFX9-NEXT: v_perm_b32 v4, v35, v4, s4 |
| ; GFX9-NEXT: v_perm_b32 v5, v34, v5, s4 |
| ; GFX9-NEXT: v_perm_b32 v6, v33, v6, s4 |
| ; GFX9-NEXT: v_perm_b32 v7, v32, v7, s4 |
| ; GFX9-NEXT: v_perm_b32 v8, v31, v8, s4 |
| ; GFX9-NEXT: v_perm_b32 v9, v30, v9, s4 |
| ; GFX9-NEXT: v_perm_b32 v10, v29, v10, s4 |
| ; GFX9-NEXT: v_perm_b32 v11, v28, v11, s4 |
| ; GFX9-NEXT: v_perm_b32 v12, v27, v12, s4 |
| ; GFX9-NEXT: v_perm_b32 v13, v26, v13, s4 |
| ; GFX9-NEXT: v_perm_b32 v14, v25, v14, s4 |
| ; GFX9-NEXT: v_perm_b32 v15, v24, v15, s4 |
| ; GFX9-NEXT: v_perm_b32 v16, v23, v16, s4 |
| ; GFX9-NEXT: v_perm_b32 v17, v22, v17, s4 |
| ; GFX9-NEXT: v_perm_b32 v18, v21, v18, s4 |
| ; GFX9-NEXT: v_perm_b32 v19, v20, v19, s4 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v10f64_to_v40f16: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB52_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[18:19], v[18:19], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX11-TRUE16-NEXT: .LBB52_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v10f64_to_v40f16: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr27 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr23 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr22 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr21 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr20 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB52_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB52_2: ; %Flow |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB52_4 |
| ; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[18:19], v[18:19], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[16:17], v[16:17], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[14:15], v[14:15], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[12:13], v[12:13], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[10:11], v[10:11], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[8:9], v[8:9], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[6:7], v[6:7], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[4:5], v[4:5], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[2:3], v[2:3], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[0:1], v[0:1], 1.0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-FAKE16-NEXT: .LBB52_4: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v39, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v37, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v36, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v35, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v34, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v33, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v32, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v31, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v30, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v29, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v28, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v27, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v26, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v25, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v24, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v23, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v22, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v21, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v20, v19, 0x5040100 |
| ; GFX11-FAKE16-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 <10 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <10 x double> %a1 to <40 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x double> %a to <40 x half> |
| br label %end |
| |
| end: |
| %phi = phi <40 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x half> %phi |
| } |
| |
| define inreg <40 x half> @bitcast_v10f64_to_v40f16_scalar(<10 x double> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v10f64_to_v40f16_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_readfirstlane_b32 s9, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s5, v1 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB53_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_lshr_b32 s89, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s5, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s73, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s72, s17, 16 |
| ; SI-NEXT: s_lshr_b64 s[10:11], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[12:13], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[14:15], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b64 s[40:41], s[28:29], 16 |
| ; SI-NEXT: s_lshr_b64 s[42:43], s[26:27], 16 |
| ; SI-NEXT: s_lshr_b64 s[44:45], s[24:25], 16 |
| ; SI-NEXT: s_lshr_b64 s[46:47], s[22:23], 16 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[20:21], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[18:19], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[16:17], 16 |
| ; SI-NEXT: s_cbranch_execnz .LBB53_4 |
| ; SI-NEXT: .LBB53_2: ; %cmp.true |
| ; SI-NEXT: v_add_f64 v[18:19], s[8:9], 1.0 |
| ; SI-NEXT: v_add_f64 v[16:17], s[6:7], 1.0 |
| ; SI-NEXT: v_add_f64 v[14:15], s[4:5], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[20:21], v[18:19], 16 |
| ; SI-NEXT: v_add_f64 v[12:13], s[28:29], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[21:22], v[16:17], 16 |
| ; SI-NEXT: v_add_f64 v[10:11], s[26:27], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[22:23], v[14:15], 16 |
| ; SI-NEXT: v_add_f64 v[8:9], s[24:25], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[23:24], v[12:13], 16 |
| ; SI-NEXT: v_add_f64 v[6:7], s[22:23], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[24:25], v[10:11], 16 |
| ; SI-NEXT: v_add_f64 v[4:5], s[20:21], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[25:26], v[8:9], 16 |
| ; SI-NEXT: v_add_f64 v[2:3], s[18:19], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[26:27], v[6:7], 16 |
| ; SI-NEXT: v_add_f64 v[0:1], s[16:17], 1.0 |
| ; SI-NEXT: v_lshr_b64 v[27:28], v[4:5], 16 |
| ; SI-NEXT: v_lshr_b64 v[28:29], v[2:3], 16 |
| ; SI-NEXT: v_lshr_b64 v[29:30], v[0:1], 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v1 |
| ; SI-NEXT: s_branch .LBB53_5 |
| ; SI-NEXT: .LBB53_3: |
| ; SI-NEXT: ; implicit-def: $sgpr60 |
| ; SI-NEXT: ; implicit-def: $sgpr72 |
| ; SI-NEXT: ; implicit-def: $sgpr58 |
| ; SI-NEXT: ; implicit-def: $sgpr73 |
| ; SI-NEXT: ; implicit-def: $sgpr56 |
| ; SI-NEXT: ; implicit-def: $sgpr74 |
| ; SI-NEXT: ; implicit-def: $sgpr46 |
| ; SI-NEXT: ; implicit-def: $sgpr75 |
| ; SI-NEXT: ; implicit-def: $sgpr44 |
| ; SI-NEXT: ; implicit-def: $sgpr76 |
| ; SI-NEXT: ; implicit-def: $sgpr42 |
| ; SI-NEXT: ; implicit-def: $sgpr77 |
| ; SI-NEXT: ; implicit-def: $sgpr40 |
| ; SI-NEXT: ; implicit-def: $sgpr78 |
| ; SI-NEXT: ; implicit-def: $sgpr14 |
| ; SI-NEXT: ; implicit-def: $sgpr79 |
| ; SI-NEXT: ; implicit-def: $sgpr12 |
| ; SI-NEXT: ; implicit-def: $sgpr88 |
| ; SI-NEXT: ; implicit-def: $sgpr10 |
| ; SI-NEXT: ; implicit-def: $sgpr89 |
| ; SI-NEXT: s_branch .LBB53_2 |
| ; SI-NEXT: .LBB53_4: |
| ; SI-NEXT: v_mov_b32_e32 v1, s17 |
| ; SI-NEXT: v_mov_b32_e32 v0, s16 |
| ; SI-NEXT: v_mov_b32_e32 v3, s19 |
| ; SI-NEXT: v_mov_b32_e32 v2, s18 |
| ; SI-NEXT: v_mov_b32_e32 v5, s21 |
| ; SI-NEXT: v_mov_b32_e32 v4, s20 |
| ; SI-NEXT: v_mov_b32_e32 v19, s9 |
| ; SI-NEXT: v_mov_b32_e32 v17, s7 |
| ; SI-NEXT: v_mov_b32_e32 v15, s5 |
| ; SI-NEXT: v_mov_b32_e32 v13, s29 |
| ; SI-NEXT: v_mov_b32_e32 v11, s27 |
| ; SI-NEXT: v_mov_b32_e32 v9, s25 |
| ; SI-NEXT: v_mov_b32_e32 v7, s23 |
| ; SI-NEXT: v_mov_b32_e32 v6, s22 |
| ; SI-NEXT: v_mov_b32_e32 v8, s24 |
| ; SI-NEXT: v_mov_b32_e32 v10, s26 |
| ; SI-NEXT: v_mov_b32_e32 v12, s28 |
| ; SI-NEXT: v_mov_b32_e32 v14, s4 |
| ; SI-NEXT: v_mov_b32_e32 v16, s6 |
| ; SI-NEXT: v_mov_b32_e32 v18, s8 |
| ; SI-NEXT: v_mov_b32_e32 v31, s89 |
| ; SI-NEXT: v_mov_b32_e32 v32, s88 |
| ; SI-NEXT: v_mov_b32_e32 v33, s79 |
| ; SI-NEXT: v_mov_b32_e32 v34, s78 |
| ; SI-NEXT: v_mov_b32_e32 v35, s77 |
| ; SI-NEXT: v_mov_b32_e32 v36, s76 |
| ; SI-NEXT: v_mov_b32_e32 v37, s75 |
| ; SI-NEXT: v_mov_b32_e32 v38, s74 |
| ; SI-NEXT: v_mov_b32_e32 v39, s73 |
| ; SI-NEXT: v_mov_b32_e32 v48, s72 |
| ; SI-NEXT: v_mov_b32_e32 v29, s60 |
| ; SI-NEXT: v_mov_b32_e32 v28, s58 |
| ; SI-NEXT: v_mov_b32_e32 v27, s56 |
| ; SI-NEXT: v_mov_b32_e32 v26, s46 |
| ; SI-NEXT: v_mov_b32_e32 v25, s44 |
| ; SI-NEXT: v_mov_b32_e32 v24, s42 |
| ; SI-NEXT: v_mov_b32_e32 v23, s40 |
| ; SI-NEXT: v_mov_b32_e32 v22, s14 |
| ; SI-NEXT: v_mov_b32_e32 v21, s12 |
| ; SI-NEXT: v_mov_b32_e32 v20, s10 |
| ; SI-NEXT: .LBB53_5: ; %end |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v29 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v48 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v28 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v27 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v26 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v25 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v24 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v35 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v23 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v34 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v22 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v21 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v32 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v31 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v29 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v28 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v27 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v26 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v25 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v24 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v22 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v10f64_to_v40f16_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s7, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s5, v1 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB53_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_lshr_b32 s12, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s46, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s5, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s4, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: s_cbranch_execnz .LBB53_4 |
| ; VI-NEXT: .LBB53_2: ; %cmp.true |
| ; VI-NEXT: v_add_f64 v[18:19], s[8:9], 1.0 |
| ; VI-NEXT: v_add_f64 v[16:17], s[6:7], 1.0 |
| ; VI-NEXT: v_add_f64 v[14:15], s[4:5], 1.0 |
| ; VI-NEXT: v_add_f64 v[12:13], s[28:29], 1.0 |
| ; VI-NEXT: v_add_f64 v[10:11], s[26:27], 1.0 |
| ; VI-NEXT: v_add_f64 v[8:9], s[24:25], 1.0 |
| ; VI-NEXT: v_add_f64 v[6:7], s[22:23], 1.0 |
| ; VI-NEXT: v_add_f64 v[4:5], s[20:21], 1.0 |
| ; VI-NEXT: v_add_f64 v[2:3], s[18:19], 1.0 |
| ; VI-NEXT: v_add_f64 v[0:1], s[16:17], 1.0 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v5 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; VI-NEXT: s_branch .LBB53_5 |
| ; VI-NEXT: .LBB53_3: |
| ; VI-NEXT: ; implicit-def: $sgpr63 |
| ; VI-NEXT: ; implicit-def: $sgpr45 |
| ; VI-NEXT: ; implicit-def: $sgpr62 |
| ; VI-NEXT: ; implicit-def: $sgpr44 |
| ; VI-NEXT: ; implicit-def: $sgpr61 |
| ; VI-NEXT: ; implicit-def: $sgpr43 |
| ; VI-NEXT: ; implicit-def: $sgpr60 |
| ; VI-NEXT: ; implicit-def: $sgpr42 |
| ; VI-NEXT: ; implicit-def: $sgpr59 |
| ; VI-NEXT: ; implicit-def: $sgpr41 |
| ; VI-NEXT: ; implicit-def: $sgpr58 |
| ; VI-NEXT: ; implicit-def: $sgpr40 |
| ; VI-NEXT: ; implicit-def: $sgpr57 |
| ; VI-NEXT: ; implicit-def: $sgpr15 |
| ; VI-NEXT: ; implicit-def: $sgpr56 |
| ; VI-NEXT: ; implicit-def: $sgpr14 |
| ; VI-NEXT: ; implicit-def: $sgpr47 |
| ; VI-NEXT: ; implicit-def: $sgpr13 |
| ; VI-NEXT: ; implicit-def: $sgpr46 |
| ; VI-NEXT: ; implicit-def: $sgpr12 |
| ; VI-NEXT: s_branch .LBB53_2 |
| ; VI-NEXT: .LBB53_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s16 |
| ; VI-NEXT: v_mov_b32_e32 v2, s18 |
| ; VI-NEXT: v_mov_b32_e32 v4, s20 |
| ; VI-NEXT: v_mov_b32_e32 v6, s22 |
| ; VI-NEXT: v_mov_b32_e32 v8, s24 |
| ; VI-NEXT: v_mov_b32_e32 v10, s26 |
| ; VI-NEXT: v_mov_b32_e32 v12, s28 |
| ; VI-NEXT: v_mov_b32_e32 v14, s4 |
| ; VI-NEXT: v_mov_b32_e32 v16, s6 |
| ; VI-NEXT: v_mov_b32_e32 v18, s8 |
| ; VI-NEXT: v_mov_b32_e32 v1, s17 |
| ; VI-NEXT: v_mov_b32_e32 v3, s19 |
| ; VI-NEXT: v_mov_b32_e32 v5, s21 |
| ; VI-NEXT: v_mov_b32_e32 v7, s23 |
| ; VI-NEXT: v_mov_b32_e32 v9, s25 |
| ; VI-NEXT: v_mov_b32_e32 v11, s27 |
| ; VI-NEXT: v_mov_b32_e32 v13, s29 |
| ; VI-NEXT: v_mov_b32_e32 v15, s5 |
| ; VI-NEXT: v_mov_b32_e32 v17, s7 |
| ; VI-NEXT: v_mov_b32_e32 v19, s9 |
| ; VI-NEXT: v_mov_b32_e32 v37, s63 |
| ; VI-NEXT: v_mov_b32_e32 v35, s62 |
| ; VI-NEXT: v_mov_b32_e32 v33, s61 |
| ; VI-NEXT: v_mov_b32_e32 v32, s60 |
| ; VI-NEXT: v_mov_b32_e32 v31, s59 |
| ; VI-NEXT: v_mov_b32_e32 v27, s58 |
| ; VI-NEXT: v_mov_b32_e32 v25, s57 |
| ; VI-NEXT: v_mov_b32_e32 v23, s56 |
| ; VI-NEXT: v_mov_b32_e32 v21, s47 |
| ; VI-NEXT: v_mov_b32_e32 v20, s46 |
| ; VI-NEXT: v_mov_b32_e32 v39, s45 |
| ; VI-NEXT: v_mov_b32_e32 v38, s44 |
| ; VI-NEXT: v_mov_b32_e32 v36, s43 |
| ; VI-NEXT: v_mov_b32_e32 v34, s42 |
| ; VI-NEXT: v_mov_b32_e32 v30, s41 |
| ; VI-NEXT: v_mov_b32_e32 v29, s40 |
| ; VI-NEXT: v_mov_b32_e32 v28, s15 |
| ; VI-NEXT: v_mov_b32_e32 v26, s14 |
| ; VI-NEXT: v_mov_b32_e32 v24, s13 |
| ; VI-NEXT: v_mov_b32_e32 v22, s12 |
| ; VI-NEXT: .LBB53_5: ; %end |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v39 |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v38 |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v36 |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v29 |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v28 |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v26 |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v24 |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v22 |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v10f64_to_v40f16_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: v_readfirstlane_b32 s9, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s8, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s7, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s5, v1 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB53_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_lshr_b32 s12, s9, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s8, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s7, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s6, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s5, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s4, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s58, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s60, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s43, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s62, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s16, 16 |
| ; GFX9-NEXT: s_cbranch_execnz .LBB53_4 |
| ; GFX9-NEXT: .LBB53_2: ; %cmp.true |
| ; GFX9-NEXT: v_add_f64 v[18:19], s[8:9], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[16:17], s[6:7], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[14:15], s[4:5], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[12:13], s[28:29], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[10:11], s[26:27], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[8:9], s[24:25], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[6:7], s[22:23], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[4:5], s[20:21], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[2:3], s[18:19], 1.0 |
| ; GFX9-NEXT: v_add_f64 v[0:1], s[16:17], 1.0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; GFX9-NEXT: s_branch .LBB53_5 |
| ; GFX9-NEXT: .LBB53_3: |
| ; GFX9-NEXT: ; implicit-def: $sgpr63 |
| ; GFX9-NEXT: ; implicit-def: $sgpr45 |
| ; GFX9-NEXT: ; implicit-def: $sgpr62 |
| ; GFX9-NEXT: ; implicit-def: $sgpr44 |
| ; GFX9-NEXT: ; implicit-def: $sgpr61 |
| ; GFX9-NEXT: ; implicit-def: $sgpr43 |
| ; GFX9-NEXT: ; implicit-def: $sgpr60 |
| ; GFX9-NEXT: ; implicit-def: $sgpr42 |
| ; GFX9-NEXT: ; implicit-def: $sgpr59 |
| ; GFX9-NEXT: ; implicit-def: $sgpr41 |
| ; GFX9-NEXT: ; implicit-def: $sgpr58 |
| ; GFX9-NEXT: ; implicit-def: $sgpr40 |
| ; GFX9-NEXT: ; implicit-def: $sgpr57 |
| ; GFX9-NEXT: ; implicit-def: $sgpr15 |
| ; GFX9-NEXT: ; implicit-def: $sgpr56 |
| ; GFX9-NEXT: ; implicit-def: $sgpr14 |
| ; GFX9-NEXT: ; implicit-def: $sgpr47 |
| ; GFX9-NEXT: ; implicit-def: $sgpr13 |
| ; GFX9-NEXT: ; implicit-def: $sgpr46 |
| ; GFX9-NEXT: ; implicit-def: $sgpr12 |
| ; GFX9-NEXT: s_branch .LBB53_2 |
| ; GFX9-NEXT: .LBB53_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s28 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s4 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s6 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s29 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s5 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v33, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v32, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s15 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s14 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s13 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s12 |
| ; GFX9-NEXT: .LBB53_5: ; %end |
| ; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX9-NEXT: v_lshl_or_b32 v0, v37, 16, v0 |
| ; GFX9-NEXT: v_lshl_or_b32 v1, v39, 16, v1 |
| ; GFX9-NEXT: v_lshl_or_b32 v2, v35, 16, v2 |
| ; GFX9-NEXT: v_lshl_or_b32 v3, v38, 16, v3 |
| ; GFX9-NEXT: v_lshl_or_b32 v4, v33, 16, v4 |
| ; GFX9-NEXT: v_lshl_or_b32 v5, v36, 16, v5 |
| ; GFX9-NEXT: v_lshl_or_b32 v6, v32, 16, v6 |
| ; GFX9-NEXT: v_lshl_or_b32 v7, v34, 16, v7 |
| ; GFX9-NEXT: v_lshl_or_b32 v8, v31, 16, v8 |
| ; GFX9-NEXT: v_lshl_or_b32 v9, v30, 16, v9 |
| ; GFX9-NEXT: v_lshl_or_b32 v10, v27, 16, v10 |
| ; GFX9-NEXT: v_lshl_or_b32 v11, v29, 16, v11 |
| ; GFX9-NEXT: v_lshl_or_b32 v12, v25, 16, v12 |
| ; GFX9-NEXT: v_lshl_or_b32 v13, v28, 16, v13 |
| ; GFX9-NEXT: v_lshl_or_b32 v14, v23, 16, v14 |
| ; GFX9-NEXT: v_lshl_or_b32 v15, v26, 16, v15 |
| ; GFX9-NEXT: v_lshl_or_b32 v16, v21, 16, v16 |
| ; GFX9-NEXT: v_lshl_or_b32 v17, v24, 16, v17 |
| ; GFX9-NEXT: v_lshl_or_b32 v18, v20, 16, v18 |
| ; GFX9-NEXT: v_lshl_or_b32 v19, v22, 16, v19 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v10f64_to_v40f16_scalar: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v1 |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 |
| ; GFX11-TRUE16-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB53_3 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s5, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s4, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s29, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s27, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s26, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s25, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s24, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s22, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s21, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s20, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s19, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s47, s18, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s17, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s56, s16, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s3, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s57, s2, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s1, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s58, s0, 16 |
| ; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB53_4 |
| ; GFX11-TRUE16-NEXT: .LBB53_2: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[18:19], s[4:5], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[16:17], s[28:29], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[14:15], s[26:27], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[12:13], s[24:25], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[10:11], s[22:23], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[8:9], s[20:21], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[6:7], s[18:19], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[4:5], s[16:17], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[2:3], s[2:3], 1.0 |
| ; GFX11-TRUE16-NEXT: v_add_f64 v[0:1], s[0:1], 1.0 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 16, v19 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 16, v18 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v17 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 16, v16 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v15 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 16, v14 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v13 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v12 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v11 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v10 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v9 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v8 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v7 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v6 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v5 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v2 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v1 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; GFX11-TRUE16-NEXT: s_branch .LBB53_5 |
| ; GFX11-TRUE16-NEXT: .LBB53_3: |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr58 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr40 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr57 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr15 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr56 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr14 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr47 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr13 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr46 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr12 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr45 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr11 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr44 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr10 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr43 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr9 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr8 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr41 |
| ; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr7 |
| ; GFX11-TRUE16-NEXT: s_branch .LBB53_2 |
| ; GFX11-TRUE16-NEXT: .LBB53_4: |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v37, s58 :: v_dual_mov_b32 v20, s41 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v35, s57 :: v_dual_mov_b32 v38, s15 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v33, s56 :: v_dual_mov_b32 v36, s14 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, s47 :: v_dual_mov_b32 v34, s13 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, s46 :: v_dual_mov_b32 v32, s12 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, s45 :: v_dual_mov_b32 v30, s11 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, s44 :: v_dual_mov_b32 v28, s10 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, s43 :: v_dual_mov_b32 v26, s9 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, s42 :: v_dual_mov_b32 v24, s8 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v39, s40 :: v_dual_mov_b32 v22, s7 |
| ; GFX11-TRUE16-NEXT: .LBB53_5: ; %end |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v37.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v39.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v35.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v38.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v33.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v36.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v31.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v34.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v29.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v32.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v27.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v30.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v25.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v28.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v23.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v26.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v21.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v24.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v20.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v22.l |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v10f64_to_v40f16_scalar: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s6, v2 |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s5, v1 |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s4, v0 |
| ; GFX11-FAKE16-NEXT: s_cmp_lg_u32 s6, 0 |
| ; GFX11-FAKE16-NEXT: s_mov_b32 s6, 0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB53_3 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s5, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s4, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s29, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s27, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s26, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s25, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s24, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s23, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s22, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s21, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s20, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s19, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s47, s18, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s17, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s56, s16, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s3, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s57, s2, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s40, s1, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s58, s0, 16 |
| ; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s6 |
| ; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB53_4 |
| ; GFX11-FAKE16-NEXT: .LBB53_2: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[15:16], s[4:5], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[17:18], s[28:29], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[19:20], s[26:27], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[10:11], s[24:25], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[12:13], s[22:23], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[5:6], s[20:21], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[7:8], s[18:19], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[21:22], s[16:17], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[0:1], s[2:3], 1.0 |
| ; GFX11-FAKE16-NEXT: v_add_f64 v[2:3], s[0:1], 1.0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v14, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v9, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v22 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v4, 16, v21 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: s_branch .LBB53_5 |
| ; GFX11-FAKE16-NEXT: .LBB53_3: |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr58 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr40 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr57 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr15 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr56 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr14 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr47 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr13 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr46 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr12 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr45 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr11 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr44 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr10 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr43 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr9 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr8 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr41 |
| ; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr7 |
| ; GFX11-FAKE16-NEXT: s_branch .LBB53_2 |
| ; GFX11-FAKE16-NEXT: .LBB53_4: |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v21, s16 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v7, s18 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, s20 :: v_dual_mov_b32 v12, s22 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v19, s26 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v22, s17 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v15, s4 :: v_dual_mov_b32 v8, s19 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v6, s21 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v1, s3 :: v_dual_mov_b32 v20, s27 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v13, s23 :: v_dual_mov_b32 v18, s29 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s25 :: v_dual_mov_b32 v16, s5 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, s58 :: v_dual_mov_b32 v36, s57 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s56 :: v_dual_mov_b32 v9, s46 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v32, s47 :: v_dual_mov_b32 v29, s45 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v28, s44 :: v_dual_mov_b32 v23, s41 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s43 :: v_dual_mov_b32 v39, s40 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s42 :: v_dual_mov_b32 v35, s14 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v38, s15 :: v_dual_mov_b32 v33, s12 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v34, s13 :: v_dual_mov_b32 v31, s11 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v30, s10 :: v_dual_mov_b32 v27, s9 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v26, s8 :: v_dual_mov_b32 v25, s7 |
| ; GFX11-FAKE16-NEXT: .LBB53_5: ; %end |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v0 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v4, 16, v21 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v22 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v5 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v35, 16, v21 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v10 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v1 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v36, 16, v48 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xffff, v6 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v34, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v9, 16, v22 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v11 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v29, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v28, 16, v21 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v14, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v20 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v15 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v16 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v39, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v38, 16, v49 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v33, 16, v36 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v31, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v30, 16, v22 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v27, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v24, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v26, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v23, 16, v20 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v25, 16, v21 |
| ; GFX11-FAKE16-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 <10 x double> %a, splat (double 1.000000e+00) |
| %a2 = bitcast <10 x double> %a1 to <40 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <10 x double> %a to <40 x half> |
| br label %end |
| |
| end: |
| %phi = phi <40 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x half> %phi |
| } |
| |
| define <10 x double> @bitcast_v40f16_to_v10f64(<40 x half> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v40f16_to_v10f64: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_mov_b32_e32 v32, v19 |
| ; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: v_mov_b32_e32 v33, v18 |
| ; SI-NEXT: v_mov_b32_e32 v43, v0 |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; SI-NEXT: v_mov_b32_e32 v34, v17 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; SI-NEXT: v_mov_b32_e32 v35, v16 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; SI-NEXT: v_mov_b32_e32 v36, v15 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v35 |
| ; SI-NEXT: v_mov_b32_e32 v37, v14 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v36 |
| ; SI-NEXT: v_mov_b32_e32 v38, v13 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v37 |
| ; SI-NEXT: v_mov_b32_e32 v39, v12 |
| ; SI-NEXT: v_mov_b32_e32 v48, v11 |
| ; SI-NEXT: v_mov_b32_e32 v49, v10 |
| ; SI-NEXT: v_mov_b32_e32 v50, v9 |
| ; SI-NEXT: v_mov_b32_e32 v51, v8 |
| ; SI-NEXT: v_mov_b32_e32 v52, v7 |
| ; SI-NEXT: v_mov_b32_e32 v53, v6 |
| ; SI-NEXT: v_mov_b32_e32 v54, v5 |
| ; SI-NEXT: v_mov_b32_e32 v55, v4 |
| ; SI-NEXT: v_mov_b32_e32 v40, v3 |
| ; SI-NEXT: v_mov_b32_e32 v41, v2 |
| ; SI-NEXT: v_mov_b32_e32 v42, v1 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v38 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v39 |
| ; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v48 |
| ; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v49 |
| ; SI-NEXT: v_lshrrev_b32_e32 v62, 16, v50 |
| ; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v51 |
| ; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v52 |
| ; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v53 |
| ; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v54 |
| ; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v55 |
| ; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v40 |
| ; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v41 |
| ; SI-NEXT: v_lshrrev_b32_e32 v58, 16, v42 |
| ; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v43 |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB54_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v43 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v59 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v42 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v58 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; SI-NEXT: v_or_b32_e32 v1, v2, v3 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v41 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v57 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v40 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v56 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v55 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v47 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v54 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v46 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v53 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v45 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v52 |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v44 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v51 |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v63 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v50 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v62 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v49 |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v61 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v60 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v39 |
| ; SI-NEXT: ; implicit-def: $vgpr43 |
| ; SI-NEXT: ; implicit-def: $vgpr42 |
| ; SI-NEXT: ; implicit-def: $vgpr41 |
| ; SI-NEXT: ; implicit-def: $vgpr40 |
| ; SI-NEXT: ; implicit-def: $vgpr55 |
| ; SI-NEXT: ; implicit-def: $vgpr54 |
| ; SI-NEXT: ; implicit-def: $vgpr53 |
| ; SI-NEXT: ; implicit-def: $vgpr52 |
| ; SI-NEXT: ; implicit-def: $vgpr51 |
| ; SI-NEXT: ; implicit-def: $vgpr50 |
| ; SI-NEXT: ; implicit-def: $vgpr49 |
| ; SI-NEXT: ; implicit-def: $vgpr48 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: ; implicit-def: $vgpr59 |
| ; SI-NEXT: ; implicit-def: $vgpr58 |
| ; SI-NEXT: ; implicit-def: $vgpr57 |
| ; SI-NEXT: ; implicit-def: $vgpr56 |
| ; SI-NEXT: ; implicit-def: $vgpr47 |
| ; SI-NEXT: ; implicit-def: $vgpr46 |
| ; SI-NEXT: ; implicit-def: $vgpr45 |
| ; SI-NEXT: ; implicit-def: $vgpr44 |
| ; SI-NEXT: ; implicit-def: $vgpr63 |
| ; SI-NEXT: ; implicit-def: $vgpr62 |
| ; SI-NEXT: ; implicit-def: $vgpr61 |
| ; SI-NEXT: ; implicit-def: $vgpr60 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v38 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v37 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v36 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v35 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v34 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v33 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v19 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v32 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr37 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr34 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: .LBB54_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB54_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v0, v59 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, v58 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, v43 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v42 |
| ; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; SI-NEXT: v_or_b32_e32 v1, v3, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, v57 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v41 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, v40 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, v55 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v56 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, v54 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, v53 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, v47 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, v44 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, v52 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4 |
| ; SI-NEXT: v_or_b32_e32 v4, v5, v4 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, v46 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, v51 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, v61 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, v45 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, v49 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; SI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, v63 |
| ; SI-NEXT: v_or_b32_e32 v7, v9, v7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, v62 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; SI-NEXT: v_or_b32_e32 v8, v10, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, v50 |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, v38 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, v37 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v11 |
| ; SI-NEXT: v_or_b32_e32 v10, v12, v10 |
| ; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, v60 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, v35 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, v34 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; SI-NEXT: v_or_b32_e32 v11, v13, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, v39 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v21, v32 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, v12 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; SI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v14 |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: v_or_b32_e32 v13, v15, v13 |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; SI-NEXT: s_waitcnt vmcnt(2) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v20, v20 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; SI-NEXT: v_or_b32_e32 v14, v16, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, v36 |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v17 |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: v_or_b32_e32 v16, v18, v16 |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(1) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_or_b32_e32 v17, v19, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, v33 |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_or_b32_e32 v18, v19, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v19, v21, v19 |
| ; SI-NEXT: .LBB54_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40f16_to_v10f64: |
| ; 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:12 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: v_mov_b32_e32 v32, v19 |
| ; VI-NEXT: v_mov_b32_e32 v33, v18 |
| ; VI-NEXT: v_mov_b32_e32 v34, v17 |
| ; VI-NEXT: v_mov_b32_e32 v35, v16 |
| ; VI-NEXT: v_mov_b32_e32 v36, v15 |
| ; VI-NEXT: v_mov_b32_e32 v37, v14 |
| ; VI-NEXT: v_mov_b32_e32 v38, v13 |
| ; VI-NEXT: v_mov_b32_e32 v39, v12 |
| ; VI-NEXT: v_mov_b32_e32 v48, v11 |
| ; VI-NEXT: v_mov_b32_e32 v49, v10 |
| ; VI-NEXT: v_mov_b32_e32 v50, v9 |
| ; VI-NEXT: v_mov_b32_e32 v51, v8 |
| ; VI-NEXT: v_mov_b32_e32 v52, v7 |
| ; VI-NEXT: v_mov_b32_e32 v53, v6 |
| ; VI-NEXT: v_mov_b32_e32 v54, v5 |
| ; VI-NEXT: v_mov_b32_e32 v55, v4 |
| ; VI-NEXT: v_mov_b32_e32 v40, v3 |
| ; VI-NEXT: v_mov_b32_e32 v41, v2 |
| ; VI-NEXT: v_mov_b32_e32 v42, v1 |
| ; VI-NEXT: v_mov_b32_e32 v43, v0 |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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_mov_b32_e32 v19, 16 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v0, v19, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v1, v19, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v2, v19, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v3, v19, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v4, v19, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v5, v19, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v6, v19, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v7, v19, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v8, v19, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v9, v19, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v10, v19, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v11, v19, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v12, v19, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v13, v19, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v14, v19, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v15, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v16, v19, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v17, v19, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v18, v19, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_lshlrev_b32_sdwa v19, v19, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 |
| ; VI-NEXT: v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v40, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v55, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v54, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v53, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v52, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v51, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v50, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v49, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v48, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v39, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v38, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v37, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v36, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v35, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v34, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v33, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v32, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; 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: .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_mov_b32_e32 v19, 0x200 |
| ; VI-NEXT: v_add_f16_sdwa v0, v43, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v1, 0x200, v43 |
| ; VI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; VI-NEXT: v_add_f16_sdwa v1, v42, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v2, 0x200, v42 |
| ; VI-NEXT: v_or_b32_e32 v1, v2, v1 |
| ; VI-NEXT: v_add_f16_sdwa v2, v41, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, 0x200, v41 |
| ; VI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; VI-NEXT: v_add_f16_sdwa v3, v40, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v4, 0x200, v40 |
| ; VI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; VI-NEXT: v_add_f16_sdwa v4, v55, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v5, 0x200, v55 |
| ; VI-NEXT: v_or_b32_e32 v4, v5, v4 |
| ; VI-NEXT: v_add_f16_sdwa v5, v54, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v6, 0x200, v54 |
| ; VI-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; VI-NEXT: v_add_f16_sdwa v6, v53, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v7, 0x200, v53 |
| ; VI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; VI-NEXT: v_add_f16_sdwa v7, v52, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v8, 0x200, v52 |
| ; VI-NEXT: v_or_b32_e32 v7, v8, v7 |
| ; VI-NEXT: v_add_f16_sdwa v8, v51, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v9, 0x200, v51 |
| ; VI-NEXT: v_or_b32_e32 v8, v9, v8 |
| ; VI-NEXT: v_add_f16_sdwa v9, v50, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v10, 0x200, v50 |
| ; VI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; VI-NEXT: v_add_f16_sdwa v10, v49, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v11, 0x200, v49 |
| ; VI-NEXT: v_or_b32_e32 v10, v11, v10 |
| ; VI-NEXT: v_add_f16_sdwa v11, v48, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v12, 0x200, v48 |
| ; VI-NEXT: v_or_b32_e32 v11, v12, v11 |
| ; VI-NEXT: v_add_f16_sdwa v12, v39, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v13, 0x200, v39 |
| ; VI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; VI-NEXT: v_add_f16_sdwa v13, v38, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v14, 0x200, v38 |
| ; VI-NEXT: v_or_b32_e32 v13, v14, v13 |
| ; VI-NEXT: v_add_f16_sdwa v14, v37, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v15, 0x200, v37 |
| ; VI-NEXT: v_or_b32_e32 v14, v15, v14 |
| ; VI-NEXT: v_add_f16_sdwa v15, v36, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v16, 0x200, v36 |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: v_add_f16_sdwa v16, v35, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v17, 0x200, v35 |
| ; VI-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; VI-NEXT: v_add_f16_sdwa v17, v34, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v18, 0x200, v34 |
| ; VI-NEXT: v_or_b32_e32 v17, v18, v17 |
| ; VI-NEXT: v_add_f16_sdwa v18, v33, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v20, 0x200, v33 |
| ; VI-NEXT: v_or_b32_e32 v18, v20, v18 |
| ; VI-NEXT: v_add_f16_sdwa v19, v32, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v20, 0x200, v32 |
| ; VI-NEXT: v_or_b32_e32 v19, v20, v19 |
| ; VI-NEXT: .LBB54_4: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40f16_to_v10f64: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_mov_b32_e32 v32, v19 |
| ; 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: v_mov_b32_e32 v33, v18 |
| ; GFX9-NEXT: v_mov_b32_e32 v43, v0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v32 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, v17 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, v16 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, v15 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, v14 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, v13 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, v12 |
| ; GFX9-NEXT: v_mov_b32_e32 v48, v11 |
| ; GFX9-NEXT: v_mov_b32_e32 v49, v10 |
| ; GFX9-NEXT: v_mov_b32_e32 v50, v9 |
| ; GFX9-NEXT: v_mov_b32_e32 v51, v8 |
| ; GFX9-NEXT: v_mov_b32_e32 v52, v7 |
| ; GFX9-NEXT: v_mov_b32_e32 v53, v6 |
| ; GFX9-NEXT: v_mov_b32_e32 v54, v5 |
| ; GFX9-NEXT: v_mov_b32_e32 v55, v4 |
| ; GFX9-NEXT: v_mov_b32_e32 v40, v3 |
| ; GFX9-NEXT: v_mov_b32_e32 v41, v2 |
| ; GFX9-NEXT: v_mov_b32_e32 v42, v1 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v48 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v49 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v50 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v51 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v52 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v53 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v54 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v55 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43 |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 |
| ; 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: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v39 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v38 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v37 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v36 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v35 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v34 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v33 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v32 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; 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: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr20 |
| ; GFX9-NEXT: ; kill: killed $vgpr20 |
| ; GFX9-NEXT: ; implicit-def: $vgpr60 |
| ; GFX9-NEXT: ; implicit-def: $vgpr61 |
| ; GFX9-NEXT: ; implicit-def: $vgpr62 |
| ; GFX9-NEXT: ; implicit-def: $vgpr63 |
| ; GFX9-NEXT: ; implicit-def: $vgpr44 |
| ; GFX9-NEXT: ; implicit-def: $vgpr45 |
| ; GFX9-NEXT: ; implicit-def: $vgpr46 |
| ; GFX9-NEXT: ; implicit-def: $vgpr47 |
| ; GFX9-NEXT: ; implicit-def: $vgpr56 |
| ; GFX9-NEXT: ; implicit-def: $vgpr57 |
| ; GFX9-NEXT: ; implicit-def: $vgpr58 |
| ; GFX9-NEXT: ; implicit-def: $vgpr59 |
| ; 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: buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v59, v43, s6 |
| ; GFX9-NEXT: s_movk_i32 s7, 0x200 |
| ; GFX9-NEXT: v_perm_b32 v1, v58, v42, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v57, v41, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v56, v40, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v47, v55, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v46, v54, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v45, v53, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v44, v52, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v63, v51, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v62, v50, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v61, v49, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v60, v48, s6 |
| ; GFX9-NEXT: v_pk_add_f16 v0, v0, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, v1, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, v2, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, v3, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, v4, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, v5, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_waitcnt vmcnt(7) |
| ; GFX9-NEXT: v_perm_b32 v12, v12, v39, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(6) |
| ; GFX9-NEXT: v_perm_b32 v13, v13, v38, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(5) |
| ; GFX9-NEXT: v_perm_b32 v14, v14, v37, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(4) |
| ; GFX9-NEXT: v_perm_b32 v15, v15, v36, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(3) |
| ; GFX9-NEXT: v_perm_b32 v16, v16, v35, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(2) |
| ; GFX9-NEXT: v_perm_b32 v17, v17, v34, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(1) |
| ; GFX9-NEXT: v_perm_b32 v18, v18, v33, s6 |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: v_perm_b32 v19, v19, v32, s6 |
| ; GFX9-NEXT: v_pk_add_f16 v12, v12, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, v13, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, v14, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v15, v15, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v16, v16, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v17, v17, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v18, v18, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v19, v19, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: .LBB54_4: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; 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-TRUE16-LABEL: bitcast_v40f16_to_v10f64: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB54_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: .LBB54_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v40f16_to_v10f64: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v20 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v37, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v38, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v39, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v48, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v36, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v35, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v34, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v33, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v32, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v31, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v30, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v29, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v28, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v27, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v26, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v25, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v24, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v23, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v22, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v21, v19, 0x5040100 |
| ; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB54_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: .LBB54_2: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-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 <40 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <40 x half> %a1 to <10 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x half> %a to <10 x double> |
| br label %end |
| |
| end: |
| %phi = phi <10 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x double> %phi |
| } |
| |
| define inreg <10 x double> @bitcast_v40f16_to_v10f64_scalar(<40 x half> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v40f16_to_v10f64_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v32, s36, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s37, 1 |
| ; SI-NEXT: v_writelane_b32 v32, s38, 2 |
| ; SI-NEXT: v_writelane_b32 v32, s39, 3 |
| ; SI-NEXT: v_writelane_b32 v32, s48, 4 |
| ; SI-NEXT: v_writelane_b32 v32, s49, 5 |
| ; SI-NEXT: v_writelane_b32 v32, s50, 6 |
| ; SI-NEXT: v_writelane_b32 v32, s51, 7 |
| ; SI-NEXT: v_writelane_b32 v32, s52, 8 |
| ; SI-NEXT: v_writelane_b32 v32, s53, 9 |
| ; SI-NEXT: v_writelane_b32 v32, s54, 10 |
| ; SI-NEXT: v_writelane_b32 v32, s55, 11 |
| ; SI-NEXT: v_writelane_b32 v32, s64, 12 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s10, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s12, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s14, v1 |
| ; SI-NEXT: v_readfirstlane_b32 s73, v0 |
| ; SI-NEXT: v_writelane_b32 v32, s65, 13 |
| ; SI-NEXT: s_lshr_b32 s72, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s75, s28, 16 |
| ; SI-NEXT: s_lshr_b32 s76, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s77, s26, 16 |
| ; SI-NEXT: s_lshr_b32 s78, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; SI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; SI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; SI-NEXT: s_lshr_b32 s92, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s93, s18, 16 |
| ; SI-NEXT: s_lshr_b32 s94, s17, 16 |
| ; SI-NEXT: s_lshr_b32 s95, s16, 16 |
| ; SI-NEXT: s_lshr_b32 s7, s6, 16 |
| ; SI-NEXT: s_lshr_b32 s9, s8, 16 |
| ; SI-NEXT: s_lshr_b32 s11, s10, 16 |
| ; SI-NEXT: s_lshr_b32 s13, s12, 16 |
| ; SI-NEXT: s_lshr_b32 s15, s14, 16 |
| ; SI-NEXT: s_lshr_b32 s74, s73, 16 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: v_writelane_b32 v32, s66, 14 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: v_writelane_b32 v32, s67, 15 |
| ; SI-NEXT: s_cbranch_scc0 .LBB55_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_and_b32 s4, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s95, 16 |
| ; SI-NEXT: s_or_b32 s36, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s94, 16 |
| ; SI-NEXT: s_or_b32 s37, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s93, 16 |
| ; SI-NEXT: s_or_b32 s38, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s92, 16 |
| ; SI-NEXT: s_or_b32 s39, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; SI-NEXT: s_or_b32 s40, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; SI-NEXT: s_or_b32 s41, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; SI-NEXT: s_or_b32 s42, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; SI-NEXT: s_or_b32 s43, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; SI-NEXT: s_or_b32 s44, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; SI-NEXT: s_or_b32 s45, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s77, 16 |
| ; SI-NEXT: s_or_b32 s46, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s76, 16 |
| ; SI-NEXT: s_or_b32 s47, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; SI-NEXT: s_or_b32 s48, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; SI-NEXT: s_or_b32 s49, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s73, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s74, 16 |
| ; SI-NEXT: s_or_b32 s50, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s14, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s15, 16 |
| ; SI-NEXT: s_or_b32 s51, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; SI-NEXT: s_or_b32 s52, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s10, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s11, 16 |
| ; SI-NEXT: s_or_b32 s53, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s8, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s9, 16 |
| ; SI-NEXT: s_or_b32 s54, s4, s5 |
| ; SI-NEXT: s_and_b32 s4, s6, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s7, 16 |
| ; SI-NEXT: s_or_b32 s55, s4, s5 |
| ; SI-NEXT: s_cbranch_execnz .LBB55_4 |
| ; SI-NEXT: .LBB55_2: ; %cmp.true |
| ; SI-NEXT: v_cvt_f32_f16_e32 v0, s95 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, s16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, s94 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s17 |
| ; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, s18 |
| ; SI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, s93 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_or_b32_e32 v1, v3, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s92 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_or_b32_e32 v2, v4, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, s19 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s91 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, s20 |
| ; SI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s90 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, s21 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_or_b32_e32 v4, v6, v4 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, s89 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_or_b32_e32 v5, v7, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, s22 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, s88 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, s23 |
| ; SI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, s79 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, s24 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_or_b32_e32 v7, v9, v7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, s78 |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8 |
| ; SI-NEXT: v_or_b32_e32 v8, v10, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, s25 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, s77 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, s26 |
| ; SI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, s76 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, s27 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_or_b32_e32 v10, v12, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, s75 |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; SI-NEXT: v_or_b32_e32 v11, v13, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, s28 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, s72 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, s29 |
| ; SI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, s74 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, s73 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_or_b32_e32 v13, v15, v13 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, s15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14 |
| ; SI-NEXT: v_or_b32_e32 v14, v16, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, s14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, s13 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, s12 |
| ; SI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, s11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, s10 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_or_b32_e32 v16, v18, v16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, s9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_or_b32_e32 v17, v19, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, s8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v20, s7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v21, s6 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_or_b32_e32 v18, v19, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v19, v21, v19 |
| ; SI-NEXT: s_branch .LBB55_5 |
| ; SI-NEXT: .LBB55_3: |
| ; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; SI-NEXT: s_branch .LBB55_2 |
| ; SI-NEXT: .LBB55_4: |
| ; SI-NEXT: v_mov_b32_e32 v0, s36 |
| ; SI-NEXT: v_mov_b32_e32 v1, s37 |
| ; SI-NEXT: v_mov_b32_e32 v2, s38 |
| ; SI-NEXT: v_mov_b32_e32 v3, s39 |
| ; SI-NEXT: v_mov_b32_e32 v4, s40 |
| ; SI-NEXT: v_mov_b32_e32 v5, s41 |
| ; SI-NEXT: v_mov_b32_e32 v6, s42 |
| ; SI-NEXT: v_mov_b32_e32 v7, s43 |
| ; SI-NEXT: v_mov_b32_e32 v8, s44 |
| ; SI-NEXT: v_mov_b32_e32 v9, s45 |
| ; SI-NEXT: v_mov_b32_e32 v10, s46 |
| ; SI-NEXT: v_mov_b32_e32 v11, s47 |
| ; SI-NEXT: v_mov_b32_e32 v12, s48 |
| ; SI-NEXT: v_mov_b32_e32 v13, s49 |
| ; SI-NEXT: v_mov_b32_e32 v14, s50 |
| ; SI-NEXT: v_mov_b32_e32 v15, s51 |
| ; SI-NEXT: v_mov_b32_e32 v16, s52 |
| ; SI-NEXT: v_mov_b32_e32 v17, s53 |
| ; SI-NEXT: v_mov_b32_e32 v18, s54 |
| ; SI-NEXT: v_mov_b32_e32 v19, s55 |
| ; SI-NEXT: v_mov_b32_e32 v20, s56 |
| ; SI-NEXT: v_mov_b32_e32 v21, s57 |
| ; SI-NEXT: v_mov_b32_e32 v22, s58 |
| ; SI-NEXT: v_mov_b32_e32 v23, s59 |
| ; SI-NEXT: v_mov_b32_e32 v24, s60 |
| ; SI-NEXT: v_mov_b32_e32 v25, s61 |
| ; SI-NEXT: v_mov_b32_e32 v26, s62 |
| ; SI-NEXT: v_mov_b32_e32 v27, s63 |
| ; SI-NEXT: v_mov_b32_e32 v28, s64 |
| ; SI-NEXT: v_mov_b32_e32 v29, s65 |
| ; SI-NEXT: v_mov_b32_e32 v30, s66 |
| ; SI-NEXT: v_mov_b32_e32 v31, s67 |
| ; SI-NEXT: .LBB55_5: ; %end |
| ; SI-NEXT: v_readlane_b32 s67, v32, 15 |
| ; SI-NEXT: v_readlane_b32 s66, v32, 14 |
| ; SI-NEXT: v_readlane_b32 s65, v32, 13 |
| ; SI-NEXT: v_readlane_b32 s64, v32, 12 |
| ; SI-NEXT: v_readlane_b32 s55, v32, 11 |
| ; SI-NEXT: v_readlane_b32 s54, v32, 10 |
| ; SI-NEXT: v_readlane_b32 s53, v32, 9 |
| ; SI-NEXT: v_readlane_b32 s52, v32, 8 |
| ; SI-NEXT: v_readlane_b32 s51, v32, 7 |
| ; SI-NEXT: v_readlane_b32 s50, v32, 6 |
| ; SI-NEXT: v_readlane_b32 s49, v32, 5 |
| ; SI-NEXT: v_readlane_b32 s48, v32, 4 |
| ; SI-NEXT: v_readlane_b32 s39, v32, 3 |
| ; SI-NEXT: v_readlane_b32 s38, v32, 2 |
| ; SI-NEXT: v_readlane_b32 s37, v32, 1 |
| ; SI-NEXT: v_readlane_b32 s36, v32, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40f16_to_v10f64_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: v_writelane_b32 v32, s30, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s31, 1 |
| ; VI-NEXT: v_writelane_b32 v32, s34, 2 |
| ; VI-NEXT: v_writelane_b32 v32, s35, 3 |
| ; VI-NEXT: v_writelane_b32 v32, s36, 4 |
| ; VI-NEXT: v_writelane_b32 v32, s37, 5 |
| ; VI-NEXT: v_writelane_b32 v32, s38, 6 |
| ; VI-NEXT: v_writelane_b32 v32, s39, 7 |
| ; VI-NEXT: v_writelane_b32 v32, s48, 8 |
| ; VI-NEXT: v_writelane_b32 v32, s49, 9 |
| ; VI-NEXT: v_writelane_b32 v32, s50, 10 |
| ; VI-NEXT: v_writelane_b32 v32, s51, 11 |
| ; VI-NEXT: v_writelane_b32 v32, s52, 12 |
| ; VI-NEXT: v_writelane_b32 v32, s53, 13 |
| ; VI-NEXT: v_writelane_b32 v32, s54, 14 |
| ; VI-NEXT: v_writelane_b32 v32, s55, 15 |
| ; VI-NEXT: v_writelane_b32 v32, s64, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s6, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s8, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s11, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s14, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s73, v1 |
| ; VI-NEXT: v_readfirstlane_b32 s76, v0 |
| ; VI-NEXT: v_writelane_b32 v32, s65, 17 |
| ; VI-NEXT: s_lshr_b32 s10, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s72, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s74, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s77, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s79, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s88, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s89, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s90, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s91, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s30, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s31, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s34, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s35, s16, 16 |
| ; VI-NEXT: s_lshr_b32 s7, s6, 16 |
| ; VI-NEXT: s_lshr_b32 s9, s8, 16 |
| ; VI-NEXT: s_lshr_b32 s12, s11, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s14, 16 |
| ; VI-NEXT: s_lshr_b32 s75, s73, 16 |
| ; VI-NEXT: s_lshr_b32 s78, s76, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: v_writelane_b32 v32, s66, 18 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: v_writelane_b32 v32, s67, 19 |
| ; VI-NEXT: s_cbranch_scc0 .LBB55_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s35, 16 |
| ; VI-NEXT: s_or_b32 s36, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s5, s34, 16 |
| ; VI-NEXT: s_or_b32 s37, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s5, s31, 16 |
| ; VI-NEXT: s_or_b32 s38, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s5, s30, 16 |
| ; VI-NEXT: s_or_b32 s39, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s5, s91, 16 |
| ; VI-NEXT: s_or_b32 s40, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s5, s90, 16 |
| ; VI-NEXT: s_or_b32 s41, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s5, s89, 16 |
| ; VI-NEXT: s_or_b32 s42, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s5, s88, 16 |
| ; VI-NEXT: s_or_b32 s43, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s5, s79, 16 |
| ; VI-NEXT: s_or_b32 s44, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s5, s77, 16 |
| ; VI-NEXT: s_or_b32 s45, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s5, s74, 16 |
| ; VI-NEXT: s_or_b32 s46, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s5, s72, 16 |
| ; VI-NEXT: s_or_b32 s47, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s5, s13, 16 |
| ; VI-NEXT: s_or_b32 s48, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s5, s10, 16 |
| ; VI-NEXT: s_or_b32 s49, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s76 |
| ; VI-NEXT: s_lshl_b32 s5, s78, 16 |
| ; VI-NEXT: s_or_b32 s50, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s73 |
| ; VI-NEXT: s_lshl_b32 s5, s75, 16 |
| ; VI-NEXT: s_or_b32 s51, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s14 |
| ; VI-NEXT: s_lshl_b32 s5, s15, 16 |
| ; VI-NEXT: s_or_b32 s52, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s11 |
| ; VI-NEXT: s_lshl_b32 s5, s12, 16 |
| ; VI-NEXT: s_or_b32 s53, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s8 |
| ; VI-NEXT: s_lshl_b32 s5, s9, 16 |
| ; VI-NEXT: s_or_b32 s54, s4, s5 |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s6 |
| ; VI-NEXT: s_lshl_b32 s5, s7, 16 |
| ; VI-NEXT: s_or_b32 s55, s4, s5 |
| ; VI-NEXT: s_cbranch_execnz .LBB55_4 |
| ; VI-NEXT: .LBB55_2: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v19, 0x200 |
| ; VI-NEXT: v_mov_b32_e32 v0, s35 |
| ; VI-NEXT: v_mov_b32_e32 v2, s34 |
| ; VI-NEXT: v_add_f16_sdwa v0, v0, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v1, s16, v19 |
| ; VI-NEXT: v_add_f16_sdwa v2, v2, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, s17, v19 |
| ; VI-NEXT: v_or_b32_e32 v0, v1, v0 |
| ; VI-NEXT: v_or_b32_e32 v1, v3, v2 |
| ; VI-NEXT: v_mov_b32_e32 v2, s31 |
| ; VI-NEXT: v_add_f16_sdwa v2, v2, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v3, s18, v19 |
| ; VI-NEXT: v_or_b32_e32 v2, v3, v2 |
| ; VI-NEXT: v_mov_b32_e32 v3, s30 |
| ; VI-NEXT: v_add_f16_sdwa v3, v3, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v4, s19, v19 |
| ; VI-NEXT: v_or_b32_e32 v3, v4, v3 |
| ; VI-NEXT: v_mov_b32_e32 v4, s91 |
| ; VI-NEXT: v_add_f16_sdwa v4, v4, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v5, s20, v19 |
| ; VI-NEXT: v_or_b32_e32 v4, v5, v4 |
| ; VI-NEXT: v_mov_b32_e32 v5, s90 |
| ; VI-NEXT: v_add_f16_sdwa v5, v5, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v6, s21, v19 |
| ; VI-NEXT: v_or_b32_e32 v5, v6, v5 |
| ; VI-NEXT: v_mov_b32_e32 v6, s89 |
| ; VI-NEXT: v_add_f16_sdwa v6, v6, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v7, s22, v19 |
| ; VI-NEXT: v_or_b32_e32 v6, v7, v6 |
| ; VI-NEXT: v_mov_b32_e32 v7, s88 |
| ; VI-NEXT: v_add_f16_sdwa v7, v7, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v8, s23, v19 |
| ; VI-NEXT: v_or_b32_e32 v7, v8, v7 |
| ; VI-NEXT: v_mov_b32_e32 v8, s79 |
| ; VI-NEXT: v_add_f16_sdwa v8, v8, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v9, s24, v19 |
| ; VI-NEXT: v_or_b32_e32 v8, v9, v8 |
| ; VI-NEXT: v_mov_b32_e32 v9, s77 |
| ; VI-NEXT: v_add_f16_sdwa v9, v9, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v10, s25, v19 |
| ; VI-NEXT: v_or_b32_e32 v9, v10, v9 |
| ; VI-NEXT: v_mov_b32_e32 v10, s74 |
| ; VI-NEXT: v_add_f16_sdwa v10, v10, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v11, s26, v19 |
| ; VI-NEXT: v_or_b32_e32 v10, v11, v10 |
| ; VI-NEXT: v_mov_b32_e32 v11, s72 |
| ; VI-NEXT: v_add_f16_sdwa v11, v11, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v12, s27, v19 |
| ; VI-NEXT: v_or_b32_e32 v11, v12, v11 |
| ; VI-NEXT: v_mov_b32_e32 v12, s13 |
| ; VI-NEXT: v_add_f16_sdwa v12, v12, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v13, s28, v19 |
| ; VI-NEXT: v_or_b32_e32 v12, v13, v12 |
| ; VI-NEXT: v_mov_b32_e32 v13, s10 |
| ; VI-NEXT: v_add_f16_sdwa v13, v13, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v14, s29, v19 |
| ; VI-NEXT: v_or_b32_e32 v13, v14, v13 |
| ; VI-NEXT: v_mov_b32_e32 v14, s78 |
| ; VI-NEXT: v_add_f16_sdwa v14, v14, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v15, s76, v19 |
| ; VI-NEXT: v_or_b32_e32 v14, v15, v14 |
| ; VI-NEXT: v_mov_b32_e32 v15, s75 |
| ; VI-NEXT: v_add_f16_sdwa v15, v15, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v16, s73, v19 |
| ; VI-NEXT: v_or_b32_e32 v15, v16, v15 |
| ; VI-NEXT: v_mov_b32_e32 v16, s15 |
| ; VI-NEXT: v_add_f16_sdwa v16, v16, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v17, s14, v19 |
| ; VI-NEXT: v_or_b32_e32 v16, v17, v16 |
| ; VI-NEXT: v_mov_b32_e32 v17, s12 |
| ; VI-NEXT: v_add_f16_sdwa v17, v17, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v18, s11, v19 |
| ; VI-NEXT: v_or_b32_e32 v17, v18, v17 |
| ; VI-NEXT: v_mov_b32_e32 v18, s9 |
| ; VI-NEXT: v_add_f16_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v20, s8, v19 |
| ; VI-NEXT: v_or_b32_e32 v18, v20, v18 |
| ; VI-NEXT: v_mov_b32_e32 v20, s7 |
| ; VI-NEXT: v_add_f16_sdwa v20, v20, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD |
| ; VI-NEXT: v_add_f16_e32 v19, s6, v19 |
| ; VI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; VI-NEXT: s_branch .LBB55_5 |
| ; VI-NEXT: .LBB55_3: |
| ; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 |
| ; VI-NEXT: s_branch .LBB55_2 |
| ; VI-NEXT: .LBB55_4: |
| ; VI-NEXT: v_mov_b32_e32 v0, s36 |
| ; VI-NEXT: v_mov_b32_e32 v1, s37 |
| ; VI-NEXT: v_mov_b32_e32 v2, s38 |
| ; VI-NEXT: v_mov_b32_e32 v3, s39 |
| ; VI-NEXT: v_mov_b32_e32 v4, s40 |
| ; VI-NEXT: v_mov_b32_e32 v5, s41 |
| ; VI-NEXT: v_mov_b32_e32 v6, s42 |
| ; VI-NEXT: v_mov_b32_e32 v7, s43 |
| ; VI-NEXT: v_mov_b32_e32 v8, s44 |
| ; VI-NEXT: v_mov_b32_e32 v9, s45 |
| ; VI-NEXT: v_mov_b32_e32 v10, s46 |
| ; VI-NEXT: v_mov_b32_e32 v11, s47 |
| ; VI-NEXT: v_mov_b32_e32 v12, s48 |
| ; VI-NEXT: v_mov_b32_e32 v13, s49 |
| ; VI-NEXT: v_mov_b32_e32 v14, s50 |
| ; VI-NEXT: v_mov_b32_e32 v15, s51 |
| ; VI-NEXT: v_mov_b32_e32 v16, s52 |
| ; VI-NEXT: v_mov_b32_e32 v17, s53 |
| ; VI-NEXT: v_mov_b32_e32 v18, s54 |
| ; VI-NEXT: v_mov_b32_e32 v19, s55 |
| ; VI-NEXT: v_mov_b32_e32 v20, s56 |
| ; VI-NEXT: v_mov_b32_e32 v21, s57 |
| ; VI-NEXT: v_mov_b32_e32 v22, s58 |
| ; VI-NEXT: v_mov_b32_e32 v23, s59 |
| ; VI-NEXT: v_mov_b32_e32 v24, s60 |
| ; VI-NEXT: v_mov_b32_e32 v25, s61 |
| ; VI-NEXT: v_mov_b32_e32 v26, s62 |
| ; VI-NEXT: v_mov_b32_e32 v27, s63 |
| ; VI-NEXT: v_mov_b32_e32 v28, s64 |
| ; VI-NEXT: v_mov_b32_e32 v29, s65 |
| ; VI-NEXT: v_mov_b32_e32 v30, s66 |
| ; VI-NEXT: v_mov_b32_e32 v31, s67 |
| ; VI-NEXT: .LBB55_5: ; %end |
| ; VI-NEXT: v_readlane_b32 s67, v32, 19 |
| ; VI-NEXT: v_readlane_b32 s66, v32, 18 |
| ; VI-NEXT: v_readlane_b32 s65, v32, 17 |
| ; VI-NEXT: v_readlane_b32 s64, v32, 16 |
| ; VI-NEXT: v_readlane_b32 s55, v32, 15 |
| ; VI-NEXT: v_readlane_b32 s54, v32, 14 |
| ; VI-NEXT: v_readlane_b32 s53, v32, 13 |
| ; VI-NEXT: v_readlane_b32 s52, v32, 12 |
| ; VI-NEXT: v_readlane_b32 s51, v32, 11 |
| ; VI-NEXT: v_readlane_b32 s50, v32, 10 |
| ; VI-NEXT: v_readlane_b32 s49, v32, 9 |
| ; VI-NEXT: v_readlane_b32 s48, v32, 8 |
| ; VI-NEXT: v_readlane_b32 s39, v32, 7 |
| ; VI-NEXT: v_readlane_b32 s38, v32, 6 |
| ; VI-NEXT: v_readlane_b32 s37, v32, 5 |
| ; VI-NEXT: v_readlane_b32 s36, v32, 4 |
| ; VI-NEXT: v_readlane_b32 s35, v32, 3 |
| ; VI-NEXT: v_readlane_b32 s34, v32, 2 |
| ; VI-NEXT: v_readlane_b32 s31, v32, 1 |
| ; VI-NEXT: v_readlane_b32 s30, v32, 0 |
| ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; VI-NEXT: s_mov_b64 exec, s[4:5] |
| ; VI-NEXT: s_waitcnt vmcnt(0) |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40f16_to_v10f64_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: v_writelane_b32 v32, s36, 0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s37, 1 |
| ; GFX9-NEXT: v_writelane_b32 v32, s38, 2 |
| ; GFX9-NEXT: v_writelane_b32 v32, s39, 3 |
| ; GFX9-NEXT: v_writelane_b32 v32, s48, 4 |
| ; GFX9-NEXT: v_writelane_b32 v32, s49, 5 |
| ; GFX9-NEXT: v_writelane_b32 v32, s50, 6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s51, 7 |
| ; GFX9-NEXT: v_writelane_b32 v32, s52, 8 |
| ; GFX9-NEXT: v_writelane_b32 v32, s53, 9 |
| ; GFX9-NEXT: v_readfirstlane_b32 s56, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s58, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s60, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s62, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s72, v1 |
| ; GFX9-NEXT: v_readfirstlane_b32 s74, v0 |
| ; GFX9-NEXT: v_writelane_b32 v32, s54, 10 |
| ; GFX9-NEXT: s_lshr_b32 s4, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s5, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s6, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s7, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s8, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s9, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s10, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s11, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s12, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s16, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s56, 16 |
| ; GFX9-NEXT: s_lshr_b32 s59, s58, 16 |
| ; GFX9-NEXT: s_lshr_b32 s61, s60, 16 |
| ; GFX9-NEXT: s_lshr_b32 s63, s62, 16 |
| ; GFX9-NEXT: s_lshr_b32 s73, s72, 16 |
| ; GFX9-NEXT: s_lshr_b32 s75, s74, 16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s42, v6 |
| ; GFX9-NEXT: v_writelane_b32 v32, s55, 11 |
| ; GFX9-NEXT: s_cmp_lg_u32 s42, 0 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s50, s74, s75 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s51, s72, s73 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s52, s62, s63 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s53, s60, s61 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s54, s58, s59 |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s55, s56, s57 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB55_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB55_4 |
| ; GFX9-NEXT: .LBB55_2: ; %cmp.true |
| ; GFX9-NEXT: v_mov_b32_e32 v19, 0x200 |
| ; GFX9-NEXT: v_pk_add_f16 v0, s36, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, s37, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, s38, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, s39, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, s40, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, s41, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, s42, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, s43, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, s44, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, s45, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, s46, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, s47, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v12, s48, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, s49, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, s50, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v15, s51, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v16, s52, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v17, s53, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v18, s54, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v19, s55, v19 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_branch .LBB55_5 |
| ; GFX9-NEXT: .LBB55_3: |
| ; GFX9-NEXT: s_branch .LBB55_2 |
| ; GFX9-NEXT: .LBB55_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s48 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s49 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s50 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s51 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s52 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s53 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s54 |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s55 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s64 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s65 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s66 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s67 |
| ; GFX9-NEXT: .LBB55_5: ; %end |
| ; GFX9-NEXT: v_readlane_b32 s55, v32, 11 |
| ; GFX9-NEXT: v_readlane_b32 s54, v32, 10 |
| ; GFX9-NEXT: v_readlane_b32 s53, v32, 9 |
| ; GFX9-NEXT: v_readlane_b32 s52, v32, 8 |
| ; GFX9-NEXT: v_readlane_b32 s51, v32, 7 |
| ; GFX9-NEXT: v_readlane_b32 s50, v32, 6 |
| ; GFX9-NEXT: v_readlane_b32 s49, v32, 5 |
| ; GFX9-NEXT: v_readlane_b32 s48, v32, 4 |
| ; GFX9-NEXT: v_readlane_b32 s39, v32, 3 |
| ; GFX9-NEXT: v_readlane_b32 s38, v32, 2 |
| ; GFX9-NEXT: v_readlane_b32 s37, v32, 1 |
| ; GFX9-NEXT: v_readlane_b32 s36, v32, 0 |
| ; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; GFX9-NEXT: s_mov_b64 exec, s[4:5] |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-LABEL: bitcast_v40f16_to_v10f64_scalar: |
| ; GFX11: ; %bb.0: |
| ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-NEXT: v_readfirstlane_b32 s45, v1 |
| ; GFX11-NEXT: v_readfirstlane_b32 s46, v0 |
| ; GFX11-NEXT: v_readfirstlane_b32 s56, v2 |
| ; GFX11-NEXT: s_lshr_b32 s41, s29, 16 |
| ; GFX11-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX11-NEXT: s_lshr_b32 s15, s27, 16 |
| ; GFX11-NEXT: s_lshr_b32 s14, s26, 16 |
| ; GFX11-NEXT: s_lshr_b32 s13, s25, 16 |
| ; GFX11-NEXT: s_lshr_b32 s12, s24, 16 |
| ; GFX11-NEXT: s_lshr_b32 s11, s23, 16 |
| ; GFX11-NEXT: s_lshr_b32 s10, s22, 16 |
| ; GFX11-NEXT: s_lshr_b32 s9, s21, 16 |
| ; GFX11-NEXT: s_lshr_b32 s8, s20, 16 |
| ; GFX11-NEXT: s_lshr_b32 s7, s19, 16 |
| ; GFX11-NEXT: s_lshr_b32 s6, s18, 16 |
| ; GFX11-NEXT: s_lshr_b32 s5, s17, 16 |
| ; GFX11-NEXT: s_lshr_b32 s4, s16, 16 |
| ; GFX11-NEXT: s_lshr_b32 s43, s3, 16 |
| ; GFX11-NEXT: s_lshr_b32 s44, s2, 16 |
| ; GFX11-NEXT: s_lshr_b32 s47, s1, 16 |
| ; GFX11-NEXT: s_lshr_b32 s57, s0, 16 |
| ; GFX11-NEXT: s_lshr_b32 s58, s45, 16 |
| ; GFX11-NEXT: s_lshr_b32 s59, s46, 16 |
| ; GFX11-NEXT: s_mov_b32 s40, 0 |
| ; GFX11-NEXT: s_cmp_lg_u32 s56, 0 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s57 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s47 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s44 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s18, s46, s59 |
| ; GFX11-NEXT: s_pack_ll_b32_b16 s19, s45, s58 |
| ; GFX11-NEXT: s_cbranch_scc0 .LBB55_3 |
| ; GFX11-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40 |
| ; GFX11-NEXT: s_cbranch_vccnz .LBB55_4 |
| ; GFX11-NEXT: .LBB55_2: ; %cmp.true |
| ; GFX11-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v18, 0x200, s18 op_sel_hi:[0,1] |
| ; GFX11-NEXT: v_pk_add_f16 v19, 0x200, s19 op_sel_hi:[0,1] |
| ; GFX11-NEXT: s_setpc_b64 s[30:31] |
| ; GFX11-NEXT: .LBB55_3: |
| ; GFX11-NEXT: s_branch .LBB55_2 |
| ; GFX11-NEXT: .LBB55_4: |
| ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 |
| ; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3 |
| ; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5 |
| ; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7 |
| ; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9 |
| ; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11 |
| ; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13 |
| ; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15 |
| ; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17 |
| ; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19 |
| ; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21 |
| ; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23 |
| ; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25 |
| ; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27 |
| ; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29 |
| ; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31 |
| ; 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 <40 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <40 x half> %a1 to <10 x double> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x half> %a to <10 x double> |
| br label %end |
| |
| end: |
| %phi = phi <10 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <10 x double> %phi |
| } |
| |
| define <40 x half> @bitcast_v40i16_to_v40f16(<40 x i16> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v40i16_to_v40f16: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: v_lshrrev_b32_e32 v21, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v18 |
| ; SI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v14 |
| ; SI-NEXT: v_lshrrev_b32_e32 v24, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v12 |
| ; SI-NEXT: v_lshrrev_b32_e32 v26, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v8 |
| ; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v6 |
| ; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v4 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v2 |
| ; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v1 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v39 |
| ; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v37 |
| ; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v38 |
| ; SI-NEXT: v_lshlrev_b32_e32 v42, 16, v34 |
| ; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v36 |
| ; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v31 |
| ; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v35 |
| ; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v28 |
| ; SI-NEXT: v_lshlrev_b32_e32 v41, 16, v33 |
| ; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v26 |
| ; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v29 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v24 |
| ; SI-NEXT: v_lshlrev_b32_e32 v47, 16, v32 |
| ; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v23 |
| ; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v30 |
| ; SI-NEXT: v_lshlrev_b32_e32 v50, 16, v22 |
| ; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v27 |
| ; SI-NEXT: v_lshlrev_b32_e32 v52, 16, v21 |
| ; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v25 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr59 |
| ; SI-NEXT: ; implicit-def: $vgpr53 |
| ; SI-NEXT: ; implicit-def: $vgpr56 |
| ; SI-NEXT: ; implicit-def: $vgpr51 |
| ; SI-NEXT: ; implicit-def: $vgpr43 |
| ; SI-NEXT: ; implicit-def: $vgpr49 |
| ; SI-NEXT: ; implicit-def: $vgpr55 |
| ; SI-NEXT: ; implicit-def: $vgpr38 |
| ; SI-NEXT: ; implicit-def: $vgpr36 |
| ; SI-NEXT: ; implicit-def: $vgpr33 |
| ; SI-NEXT: ; implicit-def: $vgpr30 |
| ; SI-NEXT: ; implicit-def: $vgpr27 |
| ; SI-NEXT: ; implicit-def: $vgpr25 |
| ; SI-NEXT: ; kill: killed $vgpr20 |
| ; SI-NEXT: ; implicit-def: $vgpr20 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB56_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_or_b32_e32 v53, v1, v63 |
| ; SI-NEXT: v_alignbit_b32 v1, v53, v60, 16 |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v3 |
| ; SI-NEXT: v_or_b32_e32 v51, v1, v40 |
| ; SI-NEXT: v_alignbit_b32 v1, v51, v61, 16 |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v5 |
| ; SI-NEXT: v_or_b32_e32 v49, v1, v42 |
| ; SI-NEXT: v_alignbit_b32 v1, v49, v62, 16 |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v7 |
| ; SI-NEXT: v_or_b32_e32 v38, v1, v45 |
| ; SI-NEXT: v_alignbit_b32 v1, v38, v54, 16 |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v9 |
| ; SI-NEXT: v_or_b32_e32 v36, v1, v46 |
| ; SI-NEXT: v_alignbit_b32 v1, v36, v41, 16 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v11 |
| ; SI-NEXT: v_or_b32_e32 v59, v0, v60 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v2 |
| ; SI-NEXT: v_or_b32_e32 v33, v1, v57 |
| ; SI-NEXT: v_or_b32_e32 v56, v0, v61 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v4 |
| ; SI-NEXT: v_alignbit_b32 v1, v33, v44, 16 |
| ; SI-NEXT: v_or_b32_e32 v43, v0, v62 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v6 |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v13 |
| ; SI-NEXT: v_or_b32_e32 v55, v0, v54 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v8 |
| ; SI-NEXT: v_or_b32_e32 v30, v1, v29 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v41 |
| ; SI-NEXT: v_alignbit_b32 v1, v30, v47, 16 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v10 |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v15 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v44 |
| ; SI-NEXT: v_or_b32_e32 v27, v1, v35 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v12 |
| ; SI-NEXT: v_alignbit_b32 v1, v27, v58, 16 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v47 |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v17 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v14 |
| ; SI-NEXT: v_or_b32_e32 v25, v1, v50 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v58 |
| ; SI-NEXT: v_alignbit_b32 v1, v25, v32, 16 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v16 |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v19 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v32 |
| ; SI-NEXT: v_or_b32_e32 v20, v1, v52 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v18 |
| ; SI-NEXT: v_alignbit_b32 v1, v20, v39, 16 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v39 |
| ; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; SI-NEXT: ; implicit-def: $vgpr0 |
| ; SI-NEXT: ; implicit-def: $vgpr1 |
| ; SI-NEXT: ; implicit-def: $vgpr2 |
| ; SI-NEXT: ; implicit-def: $vgpr3 |
| ; SI-NEXT: ; implicit-def: $vgpr4 |
| ; SI-NEXT: ; implicit-def: $vgpr5 |
| ; SI-NEXT: ; implicit-def: $vgpr6 |
| ; SI-NEXT: ; implicit-def: $vgpr7 |
| ; SI-NEXT: ; implicit-def: $vgpr8 |
| ; SI-NEXT: ; implicit-def: $vgpr9 |
| ; SI-NEXT: ; implicit-def: $vgpr10 |
| ; SI-NEXT: ; implicit-def: $vgpr11 |
| ; SI-NEXT: ; implicit-def: $vgpr12 |
| ; SI-NEXT: ; implicit-def: $vgpr13 |
| ; SI-NEXT: ; implicit-def: $vgpr14 |
| ; SI-NEXT: ; implicit-def: $vgpr15 |
| ; SI-NEXT: ; implicit-def: $vgpr16 |
| ; SI-NEXT: ; implicit-def: $vgpr17 |
| ; SI-NEXT: ; implicit-def: $vgpr18 |
| ; SI-NEXT: ; implicit-def: $vgpr19 |
| ; SI-NEXT: ; implicit-def: $vgpr63 |
| ; SI-NEXT: ; implicit-def: $vgpr60 |
| ; SI-NEXT: ; implicit-def: $vgpr40 |
| ; SI-NEXT: ; implicit-def: $vgpr61 |
| ; SI-NEXT: ; implicit-def: $vgpr42 |
| ; SI-NEXT: ; implicit-def: $vgpr62 |
| ; SI-NEXT: ; implicit-def: $vgpr45 |
| ; SI-NEXT: ; implicit-def: $vgpr54 |
| ; SI-NEXT: ; implicit-def: $vgpr46 |
| ; SI-NEXT: ; implicit-def: $vgpr41 |
| ; SI-NEXT: ; implicit-def: $vgpr57 |
| ; SI-NEXT: ; implicit-def: $vgpr44 |
| ; SI-NEXT: ; implicit-def: $vgpr29 |
| ; SI-NEXT: ; implicit-def: $vgpr47 |
| ; SI-NEXT: ; implicit-def: $vgpr35 |
| ; SI-NEXT: ; implicit-def: $vgpr58 |
| ; SI-NEXT: ; implicit-def: $vgpr50 |
| ; SI-NEXT: ; implicit-def: $vgpr32 |
| ; SI-NEXT: ; implicit-def: $vgpr52 |
| ; SI-NEXT: ; implicit-def: $vgpr39 |
| ; SI-NEXT: .LBB56_2: ; %Flow |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB56_4 |
| ; SI-NEXT: ; %bb.3: ; %cmp.true |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_or_b32_e32 v18, v39, v18 |
| ; SI-NEXT: v_add_i32_e32 v21, vcc, 0x30000, v18 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v19 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: s_mov_b32 s6, 0x30000 |
| ; SI-NEXT: v_or_b32_e32 v18, v52, v18 |
| ; SI-NEXT: v_or_b32_e32 v16, v32, v16 |
| ; SI-NEXT: v_add_i32_e32 v20, vcc, s6, v18 |
| ; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v16 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v17 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_or_b32_e32 v16, v50, v16 |
| ; SI-NEXT: v_or_b32_e32 v14, v58, v14 |
| ; SI-NEXT: v_add_i32_e32 v25, vcc, s6, v16 |
| ; SI-NEXT: v_add_i32_e32 v16, vcc, s6, v14 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v15 |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_or_b32_e32 v14, v35, v14 |
| ; SI-NEXT: v_or_b32_e32 v12, v47, v12 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0 |
| ; SI-NEXT: v_add_i32_e32 v27, vcc, s6, v14 |
| ; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v12 |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v13 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_or_b32_e32 v0, v60, v0 |
| ; SI-NEXT: v_or_b32_e32 v12, v29, v12 |
| ; SI-NEXT: v_or_b32_e32 v10, v44, v10 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_or_b32_e32 v2, v61, v2 |
| ; SI-NEXT: v_add_i32_e32 v59, vcc, s6, v0 |
| ; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v1 |
| ; SI-NEXT: v_add_i32_e32 v30, vcc, s6, v12 |
| ; SI-NEXT: v_add_i32_e32 v12, vcc, s6, v10 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v11 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_or_b32_e32 v4, v62, v4 |
| ; SI-NEXT: v_add_i32_e32 v56, vcc, s6, v2 |
| ; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v3 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_or_b32_e32 v6, v54, v6 |
| ; SI-NEXT: v_add_i32_e32 v43, vcc, s6, v4 |
| ; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v5 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_or_b32_e32 v0, v63, v0 |
| ; SI-NEXT: v_or_b32_e32 v10, v57, v10 |
| ; SI-NEXT: v_or_b32_e32 v8, v41, v8 |
| ; SI-NEXT: v_add_i32_e32 v55, vcc, s6, v6 |
| ; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v7 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_or_b32_e32 v2, v40, v2 |
| ; SI-NEXT: v_add_i32_e32 v53, vcc, s6, v0 |
| ; SI-NEXT: v_add_i32_e32 v33, vcc, s6, v10 |
| ; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v8 |
| ; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v9 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_or_b32_e32 v4, v42, v4 |
| ; SI-NEXT: v_add_i32_e32 v51, vcc, s6, v2 |
| ; SI-NEXT: v_alignbit_b32 v0, v53, v59, 16 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_or_b32_e32 v6, v45, v6 |
| ; SI-NEXT: v_add_i32_e32 v49, vcc, s6, v4 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_alignbit_b32 v0, v51, v56, 16 |
| ; SI-NEXT: v_or_b32_e32 v8, v46, v8 |
| ; SI-NEXT: v_add_i32_e32 v38, vcc, s6, v6 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_alignbit_b32 v0, v49, v43, 16 |
| ; SI-NEXT: v_add_i32_e32 v36, vcc, s6, v8 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_alignbit_b32 v0, v38, v55, 16 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(1) |
| ; SI-NEXT: v_alignbit_b32 v0, v36, v10, 16 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(1) |
| ; SI-NEXT: v_alignbit_b32 v0, v33, v12, 16 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(1) |
| ; SI-NEXT: v_alignbit_b32 v0, v30, v14, 16 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(1) |
| ; SI-NEXT: v_alignbit_b32 v0, v27, v16, 16 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(1) |
| ; SI-NEXT: v_alignbit_b32 v0, v25, v18, 16 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill |
| ; SI-NEXT: s_waitcnt expcnt(1) |
| ; SI-NEXT: v_alignbit_b32 v0, v20, v21, 16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v53 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v51 |
| ; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v49 |
| ; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v38 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v36 |
| ; SI-NEXT: v_lshrrev_b32_e32 v26, 16, v33 |
| ; SI-NEXT: v_lshrrev_b32_e32 v24, 16, v30 |
| ; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v27 |
| ; SI-NEXT: v_lshrrev_b32_e32 v22, 16, v25 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v21, 16, v20 |
| ; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill |
| ; SI-NEXT: .LBB56_4: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(1) |
| ; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v59 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v37 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v34 |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v31 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v28 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v26 |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v24 |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v23 |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v22 |
| ; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(9) |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v1 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v53 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v2 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v56 |
| ; SI-NEXT: s_waitcnt vmcnt(8) |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v3 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v51 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v43 |
| ; SI-NEXT: s_waitcnt vmcnt(7) |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v5 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v49 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v55 |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v7 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v38 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(6) |
| ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9 |
| ; SI-NEXT: s_waitcnt vmcnt(5) |
| ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11 |
| ; SI-NEXT: s_waitcnt vmcnt(4) |
| ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13 |
| ; SI-NEXT: s_waitcnt vmcnt(3) |
| ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 |
| ; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(14) |
| ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v9 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v36 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v11 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v33 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v13 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v30 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v15 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v27 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v17 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v25 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v19 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v20 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v21 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40i16_to_v40f16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v5 |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; 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 .LBB56_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_u16_e32 v0, 3, v0 |
| ; VI-NEXT: v_add_u16_e32 v39, 3, v39 |
| ; VI-NEXT: v_add_u16_e32 v1, 3, v1 |
| ; VI-NEXT: v_add_u16_e32 v38, 3, v38 |
| ; VI-NEXT: v_add_u16_e32 v2, 3, v2 |
| ; VI-NEXT: v_add_u16_e32 v37, 3, v37 |
| ; VI-NEXT: v_add_u16_e32 v3, 3, v3 |
| ; VI-NEXT: v_add_u16_e32 v36, 3, v36 |
| ; VI-NEXT: v_add_u16_e32 v4, 3, v4 |
| ; VI-NEXT: v_add_u16_e32 v20, 3, v20 |
| ; VI-NEXT: v_add_u16_e32 v5, 3, v5 |
| ; VI-NEXT: v_add_u16_e32 v35, 3, v35 |
| ; VI-NEXT: v_add_u16_e32 v6, 3, v6 |
| ; VI-NEXT: v_add_u16_e32 v34, 3, v34 |
| ; VI-NEXT: v_add_u16_e32 v7, 3, v7 |
| ; VI-NEXT: v_add_u16_e32 v33, 3, v33 |
| ; VI-NEXT: v_add_u16_e32 v8, 3, v8 |
| ; VI-NEXT: v_add_u16_e32 v32, 3, v32 |
| ; VI-NEXT: v_add_u16_e32 v9, 3, v9 |
| ; VI-NEXT: v_add_u16_e32 v31, 3, v31 |
| ; VI-NEXT: v_add_u16_e32 v10, 3, v10 |
| ; VI-NEXT: v_add_u16_e32 v30, 3, v30 |
| ; VI-NEXT: v_add_u16_e32 v11, 3, v11 |
| ; VI-NEXT: v_add_u16_e32 v29, 3, v29 |
| ; VI-NEXT: v_add_u16_e32 v12, 3, v12 |
| ; VI-NEXT: v_add_u16_e32 v28, 3, v28 |
| ; VI-NEXT: v_add_u16_e32 v13, 3, v13 |
| ; VI-NEXT: v_add_u16_e32 v27, 3, v27 |
| ; VI-NEXT: v_add_u16_e32 v14, 3, v14 |
| ; VI-NEXT: v_add_u16_e32 v26, 3, v26 |
| ; VI-NEXT: v_add_u16_e32 v15, 3, v15 |
| ; VI-NEXT: v_add_u16_e32 v25, 3, v25 |
| ; VI-NEXT: v_add_u16_e32 v16, 3, v16 |
| ; VI-NEXT: v_add_u16_e32 v24, 3, v24 |
| ; VI-NEXT: v_add_u16_e32 v17, 3, v17 |
| ; VI-NEXT: v_add_u16_e32 v23, 3, v23 |
| ; VI-NEXT: v_add_u16_e32 v18, 3, v18 |
| ; VI-NEXT: v_add_u16_e32 v22, 3, v22 |
| ; VI-NEXT: v_add_u16_e32 v19, 3, v19 |
| ; VI-NEXT: v_add_u16_e32 v21, 3, v21 |
| ; VI-NEXT: .LBB56_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v35 |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v34 |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v33 |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v32 |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v31 |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v30 |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v29 |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v28 |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v27 |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v26 |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v25 |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v24 |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v23 |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v21 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40i16_to_v40f16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v5 |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v0 |
| ; 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 .LBB56_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v19, v39, v19, s6 |
| ; GFX9-NEXT: v_perm_b32 v18, v38, v18, s6 |
| ; GFX9-NEXT: v_perm_b32 v17, v37, v17, s6 |
| ; GFX9-NEXT: v_perm_b32 v16, v36, v16, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v35, v15, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v34, v14, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v33, v13, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v32, v12, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v31, v11, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v30, v10, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v29, v9, s6 |
| ; GFX9-NEXT: v_perm_b32 v8, v28, v8, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v27, v7, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v26, v6, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v24, v5, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v25, v4, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v23, v3, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v22, v2, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v21, v1, s6 |
| ; GFX9-NEXT: v_perm_b32 v0, v20, v0, s6 |
| ; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; 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_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: v_lshrrev_b32_e32 v20, 16, v0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v19 |
| ; GFX9-NEXT: .LBB56_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_mov_b32 s4, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v20, v0, s4 |
| ; GFX9-NEXT: v_perm_b32 v1, v21, v1, s4 |
| ; GFX9-NEXT: v_perm_b32 v2, v22, v2, s4 |
| ; GFX9-NEXT: v_perm_b32 v3, v23, v3, s4 |
| ; GFX9-NEXT: v_perm_b32 v4, v25, v4, s4 |
| ; GFX9-NEXT: v_perm_b32 v5, v24, v5, s4 |
| ; GFX9-NEXT: v_perm_b32 v6, v26, v6, s4 |
| ; GFX9-NEXT: v_perm_b32 v7, v27, v7, s4 |
| ; GFX9-NEXT: v_perm_b32 v8, v28, v8, s4 |
| ; GFX9-NEXT: v_perm_b32 v9, v29, v9, s4 |
| ; GFX9-NEXT: v_perm_b32 v10, v30, v10, s4 |
| ; GFX9-NEXT: v_perm_b32 v11, v31, v11, s4 |
| ; GFX9-NEXT: v_perm_b32 v12, v32, v12, s4 |
| ; GFX9-NEXT: v_perm_b32 v13, v33, v13, s4 |
| ; GFX9-NEXT: v_perm_b32 v14, v34, v14, s4 |
| ; GFX9-NEXT: v_perm_b32 v15, v35, v15, s4 |
| ; GFX9-NEXT: v_perm_b32 v16, v36, v16, s4 |
| ; GFX9-NEXT: v_perm_b32 v17, v37, v17, s4 |
| ; GFX9-NEXT: v_perm_b32 v18, v38, v18, s4 |
| ; GFX9-NEXT: v_perm_b32 v19, v39, v19, s4 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v40i16_to_v40f16: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB56_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: .LBB56_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v40i16_to_v40f16: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v0 |
| ; GFX11-FAKE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-FAKE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB56_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v48, v19, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v39, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v38, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v37, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v36, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v35, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v34, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v33, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v32, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v31, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v30, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v29, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v28, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v27, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v26, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v21, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v22, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v23, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v24, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v25, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v19 |
| ; GFX11-FAKE16-NEXT: .LBB56_2: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v21, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v22, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v23, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v24, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v25, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v26, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v27, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v28, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v29, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v30, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v31, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v32, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v33, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v34, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v35, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v36, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v37, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v38, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v39, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v48, v19, 0x5040100 |
| ; GFX11-FAKE16-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 <40 x i16> %a, splat (i16 3) |
| %a2 = bitcast <40 x i16> %a1 to <40 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x i16> %a to <40 x half> |
| br label %end |
| |
| end: |
| %phi = phi <40 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x half> %phi |
| } |
| |
| define inreg <40 x half> @bitcast_v40i16_to_v40f16_scalar(<40 x i16> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v40i16_to_v40f16_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_writelane_b32 v20, s30, 0 |
| ; SI-NEXT: v_writelane_b32 v20, s31, 1 |
| ; SI-NEXT: v_writelane_b32 v20, s34, 2 |
| ; SI-NEXT: v_writelane_b32 v20, s35, 3 |
| ; SI-NEXT: v_writelane_b32 v20, s36, 4 |
| ; SI-NEXT: v_writelane_b32 v20, s37, 5 |
| ; SI-NEXT: v_writelane_b32 v20, s38, 6 |
| ; SI-NEXT: v_writelane_b32 v20, s39, 7 |
| ; SI-NEXT: v_writelane_b32 v20, s48, 8 |
| ; SI-NEXT: v_writelane_b32 v20, s49, 9 |
| ; SI-NEXT: v_writelane_b32 v20, s50, 10 |
| ; SI-NEXT: v_writelane_b32 v20, s51, 11 |
| ; SI-NEXT: v_writelane_b32 v20, s52, 12 |
| ; SI-NEXT: v_writelane_b32 v20, s53, 13 |
| ; SI-NEXT: v_writelane_b32 v20, s54, 14 |
| ; SI-NEXT: v_writelane_b32 v20, s55, 15 |
| ; SI-NEXT: v_writelane_b32 v20, s64, 16 |
| ; SI-NEXT: v_writelane_b32 v20, s65, 17 |
| ; SI-NEXT: v_writelane_b32 v20, s66, 18 |
| ; SI-NEXT: v_writelane_b32 v20, s67, 19 |
| ; SI-NEXT: v_writelane_b32 v20, s68, 20 |
| ; SI-NEXT: v_writelane_b32 v20, s69, 21 |
| ; SI-NEXT: v_writelane_b32 v20, s70, 22 |
| ; SI-NEXT: v_writelane_b32 v20, s71, 23 |
| ; SI-NEXT: v_readfirstlane_b32 s69, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s71, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s66, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s68, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s55, v1 |
| ; SI-NEXT: v_readfirstlane_b32 s65, v0 |
| ; SI-NEXT: s_lshr_b32 s36, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s54, s28, 16 |
| ; SI-NEXT: s_lshr_b32 s35, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s53, s26, 16 |
| ; SI-NEXT: s_lshr_b32 s34, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s52, s24, 16 |
| ; SI-NEXT: s_lshr_b32 s31, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s51, s22, 16 |
| ; SI-NEXT: s_lshr_b32 s30, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s50, s20, 16 |
| ; SI-NEXT: s_lshr_b32 s95, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s49, s18, 16 |
| ; SI-NEXT: s_lshr_b32 s94, s17, 16 |
| ; SI-NEXT: s_lshr_b32 s48, s16, 16 |
| ; SI-NEXT: s_lshr_b32 s39, s69, 16 |
| ; SI-NEXT: s_lshr_b32 s70, s71, 16 |
| ; SI-NEXT: s_lshr_b32 s38, s66, 16 |
| ; SI-NEXT: s_lshr_b32 s67, s68, 16 |
| ; SI-NEXT: s_lshr_b32 s37, s55, 16 |
| ; SI-NEXT: s_lshr_b32 s64, s65, 16 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: s_cbranch_scc0 .LBB57_4 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_and_b32 s5, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s7, s94, 16 |
| ; SI-NEXT: s_or_b32 s47, s5, s7 |
| ; SI-NEXT: s_and_b32 s5, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s7, s95, 16 |
| ; SI-NEXT: s_or_b32 s43, s5, s7 |
| ; SI-NEXT: s_and_b32 s5, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s7, s30, 16 |
| ; SI-NEXT: s_or_b32 s41, s5, s7 |
| ; SI-NEXT: s_and_b32 s5, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s7, s31, 16 |
| ; SI-NEXT: s_or_b32 s15, s5, s7 |
| ; SI-NEXT: s_and_b32 s5, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s7, s34, 16 |
| ; SI-NEXT: s_or_b32 s13, s5, s7 |
| ; SI-NEXT: s_and_b32 s5, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s7, s35, 16 |
| ; SI-NEXT: s_or_b32 s11, s5, s7 |
| ; SI-NEXT: s_and_b32 s5, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s7, s36, 16 |
| ; SI-NEXT: s_or_b32 s9, s5, s7 |
| ; SI-NEXT: s_and_b32 s5, s55, 0xffff |
| ; SI-NEXT: s_lshl_b32 s7, s37, 16 |
| ; SI-NEXT: s_or_b32 s7, s5, s7 |
| ; SI-NEXT: s_and_b32 s5, s66, 0xffff |
| ; SI-NEXT: s_lshl_b32 s45, s38, 16 |
| ; SI-NEXT: s_and_b32 s44, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s46, s48, 16 |
| ; SI-NEXT: s_or_b32 s5, s5, s45 |
| ; SI-NEXT: s_and_b32 s45, s69, 0xffff |
| ; SI-NEXT: s_lshl_b32 s56, s39, 16 |
| ; SI-NEXT: s_lshl_b32 s42, s49, 16 |
| ; SI-NEXT: s_or_b32 s91, s45, s56 |
| ; SI-NEXT: s_or_b32 s44, s44, s46 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[46:47], 16 |
| ; SI-NEXT: s_and_b32 s46, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s40, s50, 16 |
| ; SI-NEXT: s_or_b32 s46, s46, s42 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[42:43], 16 |
| ; SI-NEXT: s_and_b32 s42, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s14, s51, 16 |
| ; SI-NEXT: s_or_b32 s42, s42, s40 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[40:41], 16 |
| ; SI-NEXT: s_and_b32 s40, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s12, s52, 16 |
| ; SI-NEXT: s_or_b32 s40, s40, s14 |
| ; SI-NEXT: s_lshr_b64 s[62:63], s[14:15], 16 |
| ; SI-NEXT: s_and_b32 s14, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s10, s53, 16 |
| ; SI-NEXT: s_or_b32 s14, s14, s12 |
| ; SI-NEXT: s_lshr_b64 s[72:73], s[12:13], 16 |
| ; SI-NEXT: s_and_b32 s12, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s8, s54, 16 |
| ; SI-NEXT: s_or_b32 s12, s12, s10 |
| ; SI-NEXT: s_lshr_b64 s[74:75], s[10:11], 16 |
| ; SI-NEXT: s_and_b32 s10, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s6, s64, 16 |
| ; SI-NEXT: s_or_b32 s10, s10, s8 |
| ; SI-NEXT: s_lshr_b64 s[76:77], s[8:9], 16 |
| ; SI-NEXT: s_and_b32 s8, s65, 0xffff |
| ; SI-NEXT: s_lshl_b32 s4, s67, 16 |
| ; SI-NEXT: s_or_b32 s8, s8, s6 |
| ; SI-NEXT: s_lshr_b64 s[78:79], s[6:7], 16 |
| ; SI-NEXT: s_and_b32 s6, s68, 0xffff |
| ; SI-NEXT: s_lshl_b32 s90, s70, 16 |
| ; SI-NEXT: s_or_b32 s6, s6, s4 |
| ; SI-NEXT: s_lshr_b64 s[88:89], s[4:5], 16 |
| ; SI-NEXT: s_and_b32 s4, s71, 0xffff |
| ; SI-NEXT: s_mov_b32 s45, s47 |
| ; SI-NEXT: s_mov_b32 s47, s43 |
| ; SI-NEXT: s_mov_b32 s43, s41 |
| ; SI-NEXT: s_mov_b32 s41, s15 |
| ; SI-NEXT: s_mov_b32 s15, s13 |
| ; SI-NEXT: s_mov_b32 s13, s11 |
| ; SI-NEXT: s_mov_b32 s11, s9 |
| ; SI-NEXT: s_mov_b32 s9, s7 |
| ; SI-NEXT: s_mov_b32 s7, s5 |
| ; SI-NEXT: s_or_b32 s4, s4, s90 |
| ; SI-NEXT: s_mov_b32 s5, s91 |
| ; SI-NEXT: s_lshr_b64 s[90:91], s[90:91], 16 |
| ; SI-NEXT: s_cbranch_execnz .LBB57_3 |
| ; SI-NEXT: .LBB57_2: ; %cmp.true |
| ; SI-NEXT: s_add_i32 s71, s71, 3 |
| ; SI-NEXT: s_and_b32 s4, s71, 0xffff |
| ; SI-NEXT: s_lshl_b32 s5, s70, 16 |
| ; SI-NEXT: s_add_i32 s69, s69, 3 |
| ; SI-NEXT: s_or_b32 s4, s5, s4 |
| ; SI-NEXT: s_and_b32 s5, s69, 0xffff |
| ; SI-NEXT: s_lshl_b32 s6, s39, 16 |
| ; SI-NEXT: s_add_i32 s68, s68, 3 |
| ; SI-NEXT: s_or_b32 s5, s6, s5 |
| ; SI-NEXT: s_and_b32 s6, s68, 0xffff |
| ; SI-NEXT: s_lshl_b32 s7, s67, 16 |
| ; SI-NEXT: s_add_i32 s66, s66, 3 |
| ; SI-NEXT: s_or_b32 s6, s7, s6 |
| ; SI-NEXT: s_and_b32 s7, s66, 0xffff |
| ; SI-NEXT: s_lshl_b32 s8, s38, 16 |
| ; SI-NEXT: s_add_i32 s65, s65, 3 |
| ; SI-NEXT: s_or_b32 s7, s8, s7 |
| ; SI-NEXT: s_and_b32 s8, s65, 0xffff |
| ; SI-NEXT: s_lshl_b32 s9, s64, 16 |
| ; SI-NEXT: s_add_i32 s55, s55, 3 |
| ; SI-NEXT: s_or_b32 s8, s9, s8 |
| ; SI-NEXT: s_and_b32 s9, s55, 0xffff |
| ; SI-NEXT: s_lshl_b32 s10, s37, 16 |
| ; SI-NEXT: s_add_i32 s28, s28, 3 |
| ; SI-NEXT: s_or_b32 s9, s10, s9 |
| ; SI-NEXT: s_and_b32 s10, s28, 0xffff |
| ; SI-NEXT: s_lshl_b32 s11, s54, 16 |
| ; SI-NEXT: s_add_i32 s29, s29, 3 |
| ; SI-NEXT: s_or_b32 s10, s11, s10 |
| ; SI-NEXT: s_and_b32 s11, s29, 0xffff |
| ; SI-NEXT: s_lshl_b32 s12, s36, 16 |
| ; SI-NEXT: s_add_i32 s26, s26, 3 |
| ; SI-NEXT: s_or_b32 s11, s12, s11 |
| ; SI-NEXT: s_and_b32 s12, s26, 0xffff |
| ; SI-NEXT: s_lshl_b32 s13, s53, 16 |
| ; SI-NEXT: s_add_i32 s27, s27, 3 |
| ; SI-NEXT: s_or_b32 s12, s13, s12 |
| ; SI-NEXT: s_and_b32 s13, s27, 0xffff |
| ; SI-NEXT: s_lshl_b32 s14, s35, 16 |
| ; SI-NEXT: s_add_i32 s24, s24, 3 |
| ; SI-NEXT: s_or_b32 s13, s14, s13 |
| ; SI-NEXT: s_and_b32 s14, s24, 0xffff |
| ; SI-NEXT: s_lshl_b32 s15, s52, 16 |
| ; SI-NEXT: s_add_i32 s25, s25, 3 |
| ; SI-NEXT: s_or_b32 s14, s15, s14 |
| ; SI-NEXT: s_and_b32 s15, s25, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s34, 16 |
| ; SI-NEXT: s_add_i32 s22, s22, 3 |
| ; SI-NEXT: s_or_b32 s15, s24, s15 |
| ; SI-NEXT: s_and_b32 s22, s22, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s51, 16 |
| ; SI-NEXT: s_or_b32 s22, s24, s22 |
| ; SI-NEXT: s_add_i32 s23, s23, 3 |
| ; SI-NEXT: s_add_i32 s40, s22, 0x30000 |
| ; SI-NEXT: s_and_b32 s22, s23, 0xffff |
| ; SI-NEXT: s_lshl_b32 s23, s31, 16 |
| ; SI-NEXT: s_or_b32 s22, s23, s22 |
| ; SI-NEXT: s_add_i32 s20, s20, 3 |
| ; SI-NEXT: s_add_i32 s41, s22, 0x30000 |
| ; SI-NEXT: s_and_b32 s20, s20, 0xffff |
| ; SI-NEXT: s_lshl_b32 s22, s50, 16 |
| ; SI-NEXT: s_or_b32 s20, s22, s20 |
| ; SI-NEXT: s_add_i32 s21, s21, 3 |
| ; SI-NEXT: s_add_i32 s42, s20, 0x30000 |
| ; SI-NEXT: s_and_b32 s20, s21, 0xffff |
| ; SI-NEXT: s_lshl_b32 s21, s30, 16 |
| ; SI-NEXT: s_or_b32 s20, s21, s20 |
| ; SI-NEXT: s_add_i32 s18, s18, 3 |
| ; SI-NEXT: s_add_i32 s43, s20, 0x30000 |
| ; SI-NEXT: s_and_b32 s18, s18, 0xffff |
| ; SI-NEXT: s_lshl_b32 s20, s49, 16 |
| ; SI-NEXT: s_or_b32 s18, s20, s18 |
| ; SI-NEXT: s_add_i32 s19, s19, 3 |
| ; SI-NEXT: s_add_i32 s46, s18, 0x30000 |
| ; SI-NEXT: s_and_b32 s18, s19, 0xffff |
| ; SI-NEXT: s_lshl_b32 s19, s95, 16 |
| ; SI-NEXT: s_or_b32 s18, s19, s18 |
| ; SI-NEXT: s_add_i32 s16, s16, 3 |
| ; SI-NEXT: s_add_i32 s47, s18, 0x30000 |
| ; SI-NEXT: s_and_b32 s16, s16, 0xffff |
| ; SI-NEXT: s_lshl_b32 s18, s48, 16 |
| ; SI-NEXT: s_or_b32 s16, s18, s16 |
| ; SI-NEXT: s_add_i32 s17, s17, 3 |
| ; SI-NEXT: s_add_i32 s44, s16, 0x30000 |
| ; SI-NEXT: s_and_b32 s16, s17, 0xffff |
| ; SI-NEXT: s_lshl_b32 s17, s94, 16 |
| ; SI-NEXT: s_or_b32 s16, s17, s16 |
| ; SI-NEXT: s_add_i32 s4, s4, 0x30000 |
| ; SI-NEXT: s_add_i32 s5, s5, 0x30000 |
| ; SI-NEXT: s_add_i32 s6, s6, 0x30000 |
| ; SI-NEXT: s_add_i32 s7, s7, 0x30000 |
| ; SI-NEXT: s_add_i32 s8, s8, 0x30000 |
| ; SI-NEXT: s_add_i32 s9, s9, 0x30000 |
| ; SI-NEXT: s_add_i32 s10, s10, 0x30000 |
| ; SI-NEXT: s_add_i32 s11, s11, 0x30000 |
| ; SI-NEXT: s_add_i32 s12, s12, 0x30000 |
| ; SI-NEXT: s_add_i32 s13, s13, 0x30000 |
| ; SI-NEXT: s_add_i32 s14, s14, 0x30000 |
| ; SI-NEXT: s_add_i32 s15, s15, 0x30000 |
| ; SI-NEXT: s_add_i32 s45, s16, 0x30000 |
| ; SI-NEXT: s_lshr_b64 s[56:57], s[44:45], 16 |
| ; SI-NEXT: s_lshr_b64 s[58:59], s[46:47], 16 |
| ; SI-NEXT: s_lshr_b64 s[60:61], s[42:43], 16 |
| ; SI-NEXT: s_lshr_b64 s[62:63], s[40:41], 16 |
| ; SI-NEXT: s_lshr_b64 s[72:73], s[14:15], 16 |
| ; SI-NEXT: s_lshr_b64 s[74:75], s[12:13], 16 |
| ; SI-NEXT: s_lshr_b64 s[76:77], s[10:11], 16 |
| ; SI-NEXT: s_lshr_b64 s[78:79], s[8:9], 16 |
| ; SI-NEXT: s_lshr_b64 s[88:89], s[6:7], 16 |
| ; SI-NEXT: s_lshr_b64 s[90:91], s[4:5], 16 |
| ; SI-NEXT: s_lshr_b32 s94, s45, 16 |
| ; SI-NEXT: s_lshr_b32 s95, s47, 16 |
| ; SI-NEXT: s_lshr_b32 s30, s43, 16 |
| ; SI-NEXT: s_lshr_b32 s31, s41, 16 |
| ; SI-NEXT: s_lshr_b32 s34, s15, 16 |
| ; SI-NEXT: s_lshr_b32 s35, s13, 16 |
| ; SI-NEXT: s_lshr_b32 s36, s11, 16 |
| ; SI-NEXT: s_lshr_b32 s37, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s38, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s39, s5, 16 |
| ; SI-NEXT: .LBB57_3: ; %end |
| ; SI-NEXT: s_and_b32 s16, s44, 0xffff |
| ; SI-NEXT: s_lshl_b32 s17, s56, 16 |
| ; SI-NEXT: s_or_b32 s16, s16, s17 |
| ; SI-NEXT: s_and_b32 s17, s45, 0xffff |
| ; SI-NEXT: s_lshl_b32 s18, s94, 16 |
| ; SI-NEXT: s_or_b32 s17, s17, s18 |
| ; SI-NEXT: s_and_b32 s18, s46, 0xffff |
| ; SI-NEXT: s_lshl_b32 s19, s58, 16 |
| ; SI-NEXT: s_or_b32 s18, s18, s19 |
| ; SI-NEXT: s_and_b32 s19, s47, 0xffff |
| ; SI-NEXT: s_lshl_b32 s20, s95, 16 |
| ; SI-NEXT: s_or_b32 s19, s19, s20 |
| ; SI-NEXT: s_and_b32 s20, s42, 0xffff |
| ; SI-NEXT: s_lshl_b32 s21, s60, 16 |
| ; SI-NEXT: s_or_b32 s20, s20, s21 |
| ; SI-NEXT: s_and_b32 s21, s43, 0xffff |
| ; SI-NEXT: s_lshl_b32 s22, s30, 16 |
| ; SI-NEXT: s_or_b32 s21, s21, s22 |
| ; SI-NEXT: s_and_b32 s22, s40, 0xffff |
| ; SI-NEXT: s_lshl_b32 s23, s62, 16 |
| ; SI-NEXT: s_or_b32 s22, s22, s23 |
| ; SI-NEXT: s_and_b32 s23, s41, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s31, 16 |
| ; SI-NEXT: s_or_b32 s23, s23, s24 |
| ; SI-NEXT: s_and_b32 s14, s14, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s72, 16 |
| ; SI-NEXT: s_or_b32 s14, s14, s24 |
| ; SI-NEXT: s_and_b32 s15, s15, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s34, 16 |
| ; SI-NEXT: s_or_b32 s15, s15, s24 |
| ; SI-NEXT: s_and_b32 s12, s12, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s74, 16 |
| ; SI-NEXT: s_or_b32 s12, s12, s24 |
| ; SI-NEXT: s_and_b32 s13, s13, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s35, 16 |
| ; SI-NEXT: s_or_b32 s13, s13, s24 |
| ; SI-NEXT: s_and_b32 s10, s10, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s76, 16 |
| ; SI-NEXT: s_or_b32 s10, s10, s24 |
| ; SI-NEXT: s_and_b32 s11, s11, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s36, 16 |
| ; SI-NEXT: s_or_b32 s11, s11, s24 |
| ; SI-NEXT: s_and_b32 s8, s8, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s78, 16 |
| ; SI-NEXT: s_or_b32 s8, s8, s24 |
| ; SI-NEXT: s_and_b32 s9, s9, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s37, 16 |
| ; SI-NEXT: s_or_b32 s9, s9, s24 |
| ; SI-NEXT: s_and_b32 s6, s6, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s88, 16 |
| ; SI-NEXT: s_or_b32 s6, s6, s24 |
| ; SI-NEXT: s_and_b32 s7, s7, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s38, 16 |
| ; SI-NEXT: s_or_b32 s7, s7, s24 |
| ; SI-NEXT: s_and_b32 s4, s4, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s90, 16 |
| ; SI-NEXT: s_or_b32 s4, s4, s24 |
| ; SI-NEXT: s_and_b32 s5, s5, 0xffff |
| ; SI-NEXT: s_lshl_b32 s24, s39, 16 |
| ; SI-NEXT: s_or_b32 s5, s5, s24 |
| ; SI-NEXT: v_mov_b32_e32 v0, s16 |
| ; SI-NEXT: v_mov_b32_e32 v1, s17 |
| ; SI-NEXT: v_mov_b32_e32 v2, s18 |
| ; SI-NEXT: v_mov_b32_e32 v3, s19 |
| ; SI-NEXT: v_mov_b32_e32 v4, s20 |
| ; SI-NEXT: v_mov_b32_e32 v5, s21 |
| ; SI-NEXT: v_mov_b32_e32 v6, s22 |
| ; SI-NEXT: v_mov_b32_e32 v7, s23 |
| ; SI-NEXT: v_mov_b32_e32 v8, s14 |
| ; SI-NEXT: v_mov_b32_e32 v9, s15 |
| ; SI-NEXT: v_mov_b32_e32 v10, s12 |
| ; SI-NEXT: v_mov_b32_e32 v11, s13 |
| ; SI-NEXT: v_mov_b32_e32 v12, s10 |
| ; SI-NEXT: v_mov_b32_e32 v13, s11 |
| ; SI-NEXT: v_mov_b32_e32 v14, s8 |
| ; SI-NEXT: v_mov_b32_e32 v15, s9 |
| ; SI-NEXT: v_mov_b32_e32 v16, s6 |
| ; SI-NEXT: v_mov_b32_e32 v17, s7 |
| ; SI-NEXT: v_mov_b32_e32 v18, s4 |
| ; SI-NEXT: v_mov_b32_e32 v19, s5 |
| ; SI-NEXT: v_readlane_b32 s71, v20, 23 |
| ; SI-NEXT: v_readlane_b32 s70, v20, 22 |
| ; SI-NEXT: v_readlane_b32 s69, v20, 21 |
| ; SI-NEXT: v_readlane_b32 s68, v20, 20 |
| ; SI-NEXT: v_readlane_b32 s67, v20, 19 |
| ; SI-NEXT: v_readlane_b32 s66, v20, 18 |
| ; SI-NEXT: v_readlane_b32 s65, v20, 17 |
| ; SI-NEXT: v_readlane_b32 s64, v20, 16 |
| ; SI-NEXT: v_readlane_b32 s55, v20, 15 |
| ; SI-NEXT: v_readlane_b32 s54, v20, 14 |
| ; SI-NEXT: v_readlane_b32 s53, v20, 13 |
| ; SI-NEXT: v_readlane_b32 s52, v20, 12 |
| ; SI-NEXT: v_readlane_b32 s51, v20, 11 |
| ; SI-NEXT: v_readlane_b32 s50, v20, 10 |
| ; SI-NEXT: v_readlane_b32 s49, v20, 9 |
| ; SI-NEXT: v_readlane_b32 s48, v20, 8 |
| ; SI-NEXT: v_readlane_b32 s39, v20, 7 |
| ; SI-NEXT: v_readlane_b32 s38, v20, 6 |
| ; SI-NEXT: v_readlane_b32 s37, v20, 5 |
| ; SI-NEXT: v_readlane_b32 s36, v20, 4 |
| ; SI-NEXT: v_readlane_b32 s35, v20, 3 |
| ; SI-NEXT: v_readlane_b32 s34, v20, 2 |
| ; SI-NEXT: v_readlane_b32 s31, v20, 1 |
| ; SI-NEXT: v_readlane_b32 s30, v20, 0 |
| ; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1 |
| ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: s_mov_b64 exec, s[4:5] |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; SI-NEXT: .LBB57_4: |
| ; SI-NEXT: ; implicit-def: $sgpr44 |
| ; SI-NEXT: ; implicit-def: $sgpr56 |
| ; SI-NEXT: ; implicit-def: $sgpr46 |
| ; SI-NEXT: ; implicit-def: $sgpr58 |
| ; SI-NEXT: ; implicit-def: $sgpr42 |
| ; SI-NEXT: ; implicit-def: $sgpr60 |
| ; SI-NEXT: ; implicit-def: $sgpr40 |
| ; SI-NEXT: ; implicit-def: $sgpr62 |
| ; SI-NEXT: ; implicit-def: $sgpr14 |
| ; SI-NEXT: ; implicit-def: $sgpr72 |
| ; SI-NEXT: ; implicit-def: $sgpr12 |
| ; SI-NEXT: ; implicit-def: $sgpr74 |
| ; SI-NEXT: ; implicit-def: $sgpr10 |
| ; SI-NEXT: ; implicit-def: $sgpr76 |
| ; SI-NEXT: ; implicit-def: $sgpr8 |
| ; SI-NEXT: ; implicit-def: $sgpr78 |
| ; SI-NEXT: ; implicit-def: $sgpr6 |
| ; SI-NEXT: ; implicit-def: $sgpr88 |
| ; SI-NEXT: ; implicit-def: $sgpr4 |
| ; SI-NEXT: ; implicit-def: $sgpr90 |
| ; SI-NEXT: s_branch .LBB57_2 |
| ; |
| ; VI-LABEL: bitcast_v40i16_to_v40f16_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s7, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s9, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s12, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s40, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s43, v1 |
| ; VI-NEXT: v_readfirstlane_b32 s46, v0 |
| ; VI-NEXT: s_lshr_b32 s8, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s11, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s56, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s58, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s60, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s62, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; VI-NEXT: s_lshr_b32 s6, s7, 16 |
| ; VI-NEXT: s_lshr_b32 s10, s9, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s12, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s40, 16 |
| ; VI-NEXT: s_lshr_b32 s44, s43, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s46, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB57_4 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB57_3 |
| ; VI-NEXT: .LBB57_2: ; %cmp.true |
| ; VI-NEXT: s_add_i32 s16, s16, 3 |
| ; VI-NEXT: s_add_i32 s63, s63, 3 |
| ; VI-NEXT: s_add_i32 s17, s17, 3 |
| ; VI-NEXT: s_add_i32 s62, s62, 3 |
| ; VI-NEXT: s_add_i32 s18, s18, 3 |
| ; VI-NEXT: s_add_i32 s61, s61, 3 |
| ; VI-NEXT: s_add_i32 s19, s19, 3 |
| ; VI-NEXT: s_add_i32 s60, s60, 3 |
| ; VI-NEXT: s_add_i32 s20, s20, 3 |
| ; VI-NEXT: s_add_i32 s59, s59, 3 |
| ; VI-NEXT: s_add_i32 s21, s21, 3 |
| ; VI-NEXT: s_add_i32 s58, s58, 3 |
| ; VI-NEXT: s_add_i32 s22, s22, 3 |
| ; VI-NEXT: s_add_i32 s57, s57, 3 |
| ; VI-NEXT: s_add_i32 s23, s23, 3 |
| ; VI-NEXT: s_add_i32 s56, s56, 3 |
| ; VI-NEXT: s_add_i32 s24, s24, 3 |
| ; VI-NEXT: s_add_i32 s45, s45, 3 |
| ; VI-NEXT: s_add_i32 s25, s25, 3 |
| ; VI-NEXT: s_add_i32 s42, s42, 3 |
| ; VI-NEXT: s_add_i32 s26, s26, 3 |
| ; VI-NEXT: s_add_i32 s15, s15, 3 |
| ; VI-NEXT: s_add_i32 s27, s27, 3 |
| ; VI-NEXT: s_add_i32 s14, s14, 3 |
| ; VI-NEXT: s_add_i32 s28, s28, 3 |
| ; VI-NEXT: s_add_i32 s11, s11, 3 |
| ; VI-NEXT: s_add_i32 s29, s29, 3 |
| ; VI-NEXT: s_add_i32 s8, s8, 3 |
| ; VI-NEXT: s_add_i32 s46, s46, 3 |
| ; VI-NEXT: s_add_i32 s47, s47, 3 |
| ; VI-NEXT: s_add_i32 s43, s43, 3 |
| ; VI-NEXT: s_add_i32 s44, s44, 3 |
| ; VI-NEXT: s_add_i32 s40, s40, 3 |
| ; VI-NEXT: s_add_i32 s41, s41, 3 |
| ; VI-NEXT: s_add_i32 s12, s12, 3 |
| ; VI-NEXT: s_add_i32 s13, s13, 3 |
| ; VI-NEXT: s_add_i32 s9, s9, 3 |
| ; VI-NEXT: s_add_i32 s10, s10, 3 |
| ; VI-NEXT: s_add_i32 s7, s7, 3 |
| ; VI-NEXT: s_add_i32 s6, s6, 3 |
| ; VI-NEXT: .LBB57_3: ; %end |
| ; VI-NEXT: s_and_b32 s4, 0xffff, s16 |
| ; VI-NEXT: s_lshl_b32 s5, s63, 16 |
| ; VI-NEXT: s_or_b32 s4, s4, s5 |
| ; VI-NEXT: s_and_b32 s5, 0xffff, s17 |
| ; VI-NEXT: s_lshl_b32 s16, s62, 16 |
| ; VI-NEXT: s_or_b32 s5, s5, s16 |
| ; VI-NEXT: s_and_b32 s16, 0xffff, s18 |
| ; VI-NEXT: s_lshl_b32 s17, s61, 16 |
| ; VI-NEXT: s_or_b32 s16, s16, s17 |
| ; VI-NEXT: s_and_b32 s17, 0xffff, s19 |
| ; VI-NEXT: s_lshl_b32 s18, s60, 16 |
| ; VI-NEXT: s_or_b32 s17, s17, s18 |
| ; VI-NEXT: s_and_b32 s18, 0xffff, s20 |
| ; VI-NEXT: s_lshl_b32 s19, s59, 16 |
| ; VI-NEXT: s_or_b32 s18, s18, s19 |
| ; VI-NEXT: s_and_b32 s19, 0xffff, s21 |
| ; VI-NEXT: s_lshl_b32 s20, s58, 16 |
| ; VI-NEXT: s_or_b32 s19, s19, s20 |
| ; VI-NEXT: s_and_b32 s20, 0xffff, s22 |
| ; VI-NEXT: s_lshl_b32 s21, s57, 16 |
| ; VI-NEXT: s_or_b32 s20, s20, s21 |
| ; VI-NEXT: s_and_b32 s21, 0xffff, s23 |
| ; VI-NEXT: s_lshl_b32 s22, s56, 16 |
| ; VI-NEXT: s_or_b32 s21, s21, s22 |
| ; VI-NEXT: s_and_b32 s22, 0xffff, s24 |
| ; VI-NEXT: s_lshl_b32 s23, s45, 16 |
| ; VI-NEXT: s_or_b32 s22, s22, s23 |
| ; VI-NEXT: s_and_b32 s23, 0xffff, s25 |
| ; VI-NEXT: s_lshl_b32 s24, s42, 16 |
| ; VI-NEXT: s_or_b32 s23, s23, s24 |
| ; VI-NEXT: s_and_b32 s24, 0xffff, s26 |
| ; VI-NEXT: s_lshl_b32 s15, s15, 16 |
| ; VI-NEXT: s_or_b32 s15, s24, s15 |
| ; VI-NEXT: s_and_b32 s24, 0xffff, s27 |
| ; VI-NEXT: s_lshl_b32 s14, s14, 16 |
| ; VI-NEXT: s_or_b32 s14, s24, s14 |
| ; VI-NEXT: s_and_b32 s24, 0xffff, s28 |
| ; VI-NEXT: s_lshl_b32 s11, s11, 16 |
| ; VI-NEXT: s_or_b32 s11, s24, s11 |
| ; VI-NEXT: s_and_b32 s24, 0xffff, s29 |
| ; VI-NEXT: s_lshl_b32 s8, s8, 16 |
| ; VI-NEXT: s_or_b32 s8, s24, s8 |
| ; VI-NEXT: s_and_b32 s24, 0xffff, s46 |
| ; VI-NEXT: s_lshl_b32 s25, s47, 16 |
| ; VI-NEXT: s_or_b32 s24, s24, s25 |
| ; VI-NEXT: s_and_b32 s25, 0xffff, s43 |
| ; VI-NEXT: s_lshl_b32 s26, s44, 16 |
| ; VI-NEXT: s_or_b32 s25, s25, s26 |
| ; VI-NEXT: s_and_b32 s26, 0xffff, s40 |
| ; VI-NEXT: s_lshl_b32 s27, s41, 16 |
| ; VI-NEXT: s_and_b32 s12, 0xffff, s12 |
| ; VI-NEXT: s_lshl_b32 s13, s13, 16 |
| ; VI-NEXT: s_and_b32 s9, 0xffff, s9 |
| ; VI-NEXT: s_lshl_b32 s10, s10, 16 |
| ; VI-NEXT: s_and_b32 s7, 0xffff, s7 |
| ; VI-NEXT: s_lshl_b32 s6, s6, 16 |
| ; VI-NEXT: s_or_b32 s26, s26, s27 |
| ; VI-NEXT: s_or_b32 s12, s12, s13 |
| ; VI-NEXT: s_or_b32 s9, s9, s10 |
| ; VI-NEXT: s_or_b32 s6, s7, s6 |
| ; VI-NEXT: v_mov_b32_e32 v0, s4 |
| ; VI-NEXT: v_mov_b32_e32 v1, s5 |
| ; VI-NEXT: v_mov_b32_e32 v2, s16 |
| ; VI-NEXT: v_mov_b32_e32 v3, s17 |
| ; VI-NEXT: v_mov_b32_e32 v4, s18 |
| ; VI-NEXT: v_mov_b32_e32 v5, s19 |
| ; VI-NEXT: v_mov_b32_e32 v6, s20 |
| ; VI-NEXT: v_mov_b32_e32 v7, s21 |
| ; VI-NEXT: v_mov_b32_e32 v8, s22 |
| ; VI-NEXT: v_mov_b32_e32 v9, s23 |
| ; VI-NEXT: v_mov_b32_e32 v10, s15 |
| ; VI-NEXT: v_mov_b32_e32 v11, s14 |
| ; VI-NEXT: v_mov_b32_e32 v12, s11 |
| ; VI-NEXT: v_mov_b32_e32 v13, s8 |
| ; VI-NEXT: v_mov_b32_e32 v14, s24 |
| ; VI-NEXT: v_mov_b32_e32 v15, s25 |
| ; VI-NEXT: v_mov_b32_e32 v16, s26 |
| ; VI-NEXT: v_mov_b32_e32 v17, s12 |
| ; VI-NEXT: v_mov_b32_e32 v18, s9 |
| ; VI-NEXT: v_mov_b32_e32 v19, s6 |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; VI-NEXT: .LBB57_4: |
| ; VI-NEXT: s_branch .LBB57_2 |
| ; |
| ; GFX9-LABEL: bitcast_v40i16_to_v40f16_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s63, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s62, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s61, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s60, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s59, v1 |
| ; GFX9-NEXT: v_readfirstlane_b32 s58, v0 |
| ; GFX9-NEXT: s_lshr_b32 s43, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s12, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s11, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s10, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s9, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s8, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s7, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s6, s16, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s63, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s62, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s61, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s60, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s59, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s58, 16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB57_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB57_4 |
| ; GFX9-NEXT: .LBB57_2: ; %cmp.true |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s63, s57 |
| ; GFX9-NEXT: v_pk_add_u16 v19, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s62, s56 |
| ; GFX9-NEXT: v_pk_add_u16 v18, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s61, s47 |
| ; GFX9-NEXT: v_pk_add_u16 v17, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s60, s46 |
| ; GFX9-NEXT: v_pk_add_u16 v16, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s59, s45 |
| ; GFX9-NEXT: v_pk_add_u16 v15, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s58, s44 |
| ; GFX9-NEXT: v_pk_add_u16 v14, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s29, s43 |
| ; GFX9-NEXT: v_pk_add_u16 v13, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s28, s42 |
| ; GFX9-NEXT: v_pk_add_u16 v12, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s27, s41 |
| ; GFX9-NEXT: v_pk_add_u16 v11, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s26, s40 |
| ; GFX9-NEXT: v_pk_add_u16 v10, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s25, s15 |
| ; GFX9-NEXT: v_pk_add_u16 v9, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s24, s14 |
| ; GFX9-NEXT: v_pk_add_u16 v8, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s23, s13 |
| ; GFX9-NEXT: v_pk_add_u16 v7, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s22, s12 |
| ; GFX9-NEXT: v_pk_add_u16 v6, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s21, s11 |
| ; GFX9-NEXT: v_pk_add_u16 v5, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s20, s10 |
| ; GFX9-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s19, s9 |
| ; GFX9-NEXT: v_pk_add_u16 v3, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s18, s8 |
| ; GFX9-NEXT: v_pk_add_u16 v2, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s17, s7 |
| ; GFX9-NEXT: v_pk_add_u16 v1, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s6 |
| ; GFX9-NEXT: v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: s_branch .LBB57_5 |
| ; GFX9-NEXT: .LBB57_3: |
| ; GFX9-NEXT: s_branch .LBB57_2 |
| ; GFX9-NEXT: .LBB57_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s29 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s28 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s15 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s14 |
| ; GFX9-NEXT: v_mov_b32_e32 v32, s13 |
| ; GFX9-NEXT: v_mov_b32_e32 v33, s12 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, s6 |
| ; GFX9-NEXT: .LBB57_5: ; %end |
| ; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX9-NEXT: v_lshl_or_b32 v0, v39, 16, v0 |
| ; GFX9-NEXT: v_lshl_or_b32 v1, v38, 16, v1 |
| ; GFX9-NEXT: v_lshl_or_b32 v2, v37, 16, v2 |
| ; GFX9-NEXT: v_lshl_or_b32 v3, v36, 16, v3 |
| ; GFX9-NEXT: v_lshl_or_b32 v4, v35, 16, v4 |
| ; GFX9-NEXT: v_lshl_or_b32 v5, v34, 16, v5 |
| ; GFX9-NEXT: v_lshl_or_b32 v6, v33, 16, v6 |
| ; GFX9-NEXT: v_lshl_or_b32 v7, v32, 16, v7 |
| ; GFX9-NEXT: v_lshl_or_b32 v8, v31, 16, v8 |
| ; GFX9-NEXT: v_lshl_or_b32 v9, v30, 16, v9 |
| ; GFX9-NEXT: v_lshl_or_b32 v10, v29, 16, v10 |
| ; GFX9-NEXT: v_lshl_or_b32 v11, v28, 16, v11 |
| ; GFX9-NEXT: v_lshl_or_b32 v12, v27, 16, v12 |
| ; GFX9-NEXT: v_lshl_or_b32 v13, v26, 16, v13 |
| ; GFX9-NEXT: v_lshl_or_b32 v14, v25, 16, v14 |
| ; GFX9-NEXT: v_lshl_or_b32 v15, v24, 16, v15 |
| ; GFX9-NEXT: v_lshl_or_b32 v16, v23, 16, v16 |
| ; GFX9-NEXT: v_lshl_or_b32 v17, v22, 16, v17 |
| ; GFX9-NEXT: v_lshl_or_b32 v18, v21, 16, v18 |
| ; GFX9-NEXT: v_lshl_or_b32 v19, v20, 16, v19 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v40i16_to_v40f16_scalar: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s57, v1 |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s56, v0 |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s58, v2 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s29, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s28, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s27, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s26, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s25, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s24, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s23, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s22, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s21, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s20, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s19, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s18, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s17, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s16, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s3, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s2, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s1, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s0, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s47, s57, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s56, 16 |
| ; GFX11-TRUE16-NEXT: s_cmp_lg_u32 s58, 0 |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s58, 0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB57_3 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s58 |
| ; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB57_4 |
| ; GFX11-TRUE16-NEXT: .LBB57_2: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s47, s57, s47 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s46, s56, s46 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s29, s29, s45 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s28, s28, s44 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s27, s27, s43 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s26, s26, s42 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s25, s25, s41 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s24, s24, s40 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s23, s15 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s22, s14 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s21, s13 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s20, s12 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s19, s11 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s18, s10 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s17, s9 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s16, s8 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s6 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s4 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s7 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s5 |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, s47, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, s46, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, s29, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, s28, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s26, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, s25, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, s24, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, s15, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, s14, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s13, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s12, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s11, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s10, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s9, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, s8, 3 op_sel_hi:[1,0] |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-TRUE16-NEXT: s_branch .LBB57_5 |
| ; GFX11-TRUE16-NEXT: .LBB57_3: |
| ; GFX11-TRUE16-NEXT: s_branch .LBB57_2 |
| ; GFX11-TRUE16-NEXT: .LBB57_4: |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, s57 :: v_dual_mov_b32 v18, s56 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, s29 :: v_dual_mov_b32 v16, s28 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v14, s26 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v13, s25 :: v_dual_mov_b32 v12, s24 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, s23 :: v_dual_mov_b32 v10, s22 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, s21 :: v_dual_mov_b32 v8, s20 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, s19 :: v_dual_mov_b32 v6, s18 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v4, s16 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v2, s2 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v0, s0 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s47 :: v_dual_mov_b32 v21, s46 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v23, s44 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v24, s43 :: v_dual_mov_b32 v25, s42 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v27, s40 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, s13 :: v_dual_mov_b32 v31, s12 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v32, s11 :: v_dual_mov_b32 v33, s10 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v34, s9 :: v_dual_mov_b32 v35, s8 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v36, s6 :: v_dual_mov_b32 v37, s4 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v38, s5 :: v_dual_mov_b32 v39, s7 |
| ; GFX11-TRUE16-NEXT: .LBB57_5: ; %end |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v39.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v38.l |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v37.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v36.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v35.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v34.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v33.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v32.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v31.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v30.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v29.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v28.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v27.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v26.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v25.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v24.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v23.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v22.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v21.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v20.l |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v40i16_to_v40f16_scalar: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s57, v1 |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s56, v0 |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s58, v2 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s29, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s28, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s27, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s26, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s25, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s40, s24, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s23, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s22, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s21, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s20, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s19, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s18, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s17, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s16, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s3, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s2, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s1, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s0, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s47, s57, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s56, 16 |
| ; GFX11-FAKE16-NEXT: s_cmp_lg_u32 s58, 0 |
| ; GFX11-FAKE16-NEXT: s_mov_b32 s58, 0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB57_3 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s58 |
| ; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB57_4 |
| ; GFX11-FAKE16-NEXT: .LBB57_2: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s47, s57, s47 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s46, s56, s46 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s29, s29, s45 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s28, s28, s44 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s27, s27, s43 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s26, s26, s42 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s25, s25, s41 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s24, s24, s40 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s15, s23, s15 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s14, s22, s14 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s13, s21, s13 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s12, s20, s12 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s11, s19, s11 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s10, s18, s10 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s9, s17, s9 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s16, s8 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s6 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s4 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s0, s7 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s1, s5 |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, s47, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, s46, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, s29, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, s28, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, s27, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, s26, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, s25, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, s24, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, s15, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, s13, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, s12, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, s11, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, s10, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, s3, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, s8, 3 op_sel_hi:[1,0] |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v15 |
| ; GFX11-FAKE16-NEXT: s_branch .LBB57_5 |
| ; GFX11-FAKE16-NEXT: .LBB57_3: |
| ; GFX11-FAKE16-NEXT: s_branch .LBB57_2 |
| ; GFX11-FAKE16-NEXT: .LBB57_4: |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v15, s57 :: v_dual_mov_b32 v16, s56 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v17, s29 :: v_dual_mov_b32 v18, s28 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v19, s27 :: v_dual_mov_b32 v10, s26 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s25 :: v_dual_mov_b32 v12, s24 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v13, s23 :: v_dual_mov_b32 v14, s22 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, s21 :: v_dual_mov_b32 v6, s20 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v7, s19 :: v_dual_mov_b32 v8, s18 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v9, s17 :: v_dual_mov_b32 v0, s16 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v1, s3 :: v_dual_mov_b32 v2, s2 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v4, s0 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v20, s47 :: v_dual_mov_b32 v21, s46 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v23, s44 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s43 :: v_dual_mov_b32 v25, s42 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v27, s40 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v30, s13 :: v_dual_mov_b32 v31, s12 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v32, s11 :: v_dual_mov_b32 v33, s10 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v34, s9 :: v_dual_mov_b32 v35, s8 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v36, s6 :: v_dual_mov_b32 v37, s4 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v38, s5 :: v_dual_mov_b32 v39, s7 |
| ; GFX11-FAKE16-NEXT: .LBB57_5: ; %end |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v1 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v0 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v39, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v38, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v36, 16, v48 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v35, 16, v49 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v6 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xffff, v5 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v34, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v33, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v31, 16, v35 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v30, 16, v36 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v30, 0xffff, v11 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xffff, v10 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v29, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v28, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v26, 16, v30 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v25, 16, v31 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v16 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v15 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v24, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v23, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v21, 16, v25 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v20, 16, v26 |
| ; GFX11-FAKE16-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 <40 x i16> %a, splat (i16 3) |
| %a2 = bitcast <40 x i16> %a1 to <40 x half> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x i16> %a to <40 x half> |
| br label %end |
| |
| end: |
| %phi = phi <40 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x half> %phi |
| } |
| |
| define <40 x i16> @bitcast_v40f16_to_v40i16(<40 x half> %a, i32 %b) { |
| ; SI-LABEL: bitcast_v40f16_to_v40i16: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_lshrrev_b32_e32 v21, 16, v19 |
| ; SI-NEXT: v_lshrrev_b32_e32 v24, 16, v18 |
| ; SI-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; SI-NEXT: v_lshrrev_b32_e32 v26, 16, v16 |
| ; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v15 |
| ; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v14 |
| ; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v13 |
| ; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v12 |
| ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v11 |
| ; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v10 |
| ; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v9 |
| ; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v8 |
| ; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v7 |
| ; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v6 |
| ; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v5 |
| ; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v4 |
| ; SI-NEXT: v_lshrrev_b32_e32 v20, 16, v3 |
| ; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v2 |
| ; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v1 |
| ; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc |
| ; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5] |
| ; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5] |
| ; SI-NEXT: s_cbranch_execz .LBB58_2 |
| ; SI-NEXT: ; %bb.1: ; %cmp.true |
| ; SI-NEXT: v_cvt_f32_f16_e32 v21, v21 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, v19 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v22, v22 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v23, v23 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v15, v15 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v25, v25 |
| ; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v13, v13 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v27, v27 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v11, v11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v28, v28 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v22, v22 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v9, v9 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v30, v30 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v23, v23 |
| ; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15 |
| ; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, v7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v25, v25 |
| ; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13 |
| ; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v20, v20 |
| ; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v21 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v27, v27 |
| ; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11 |
| ; SI-NEXT: v_add_f32_e32 v28, 0x38000000, v28 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v34, v34 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v22 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v28, v28 |
| ; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9 |
| ; SI-NEXT: v_add_f32_e32 v30, 0x38000000, v30 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v39, v39 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v38, v38 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v37, v37 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v36, v36 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v35, v35 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v33, v33 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v32, v32 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v29, v29 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v26, v26 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v24, v24 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v23 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v30, v30 |
| ; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7 |
| ; SI-NEXT: v_add_f32_e32 v31, 0x38000000, v31 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v4, v4 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v6, v6 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v8, v8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v12, v12 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v14, v14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v16, v16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v18, v18 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v25 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v27 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_add_f32_e32 v34, 0x38000000, v34 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v28 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v34, v34 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_add_f32_e32 v39, 0x38000000, v39 |
| ; SI-NEXT: v_add_f32_e32 v38, 0x38000000, v38 |
| ; SI-NEXT: v_add_f32_e32 v37, 0x38000000, v37 |
| ; SI-NEXT: v_add_f32_e32 v36, 0x38000000, v36 |
| ; SI-NEXT: v_add_f32_e32 v35, 0x38000000, v35 |
| ; SI-NEXT: v_add_f32_e32 v33, 0x38000000, v33 |
| ; SI-NEXT: v_add_f32_e32 v32, 0x38000000, v32 |
| ; SI-NEXT: v_add_f32_e32 v29, 0x38000000, v29 |
| ; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26 |
| ; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v30 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v39, v39 |
| ; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v38, v38 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v37, v37 |
| ; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v36, v36 |
| ; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v35, v35 |
| ; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v33, v33 |
| ; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32 |
| ; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v29, v29 |
| ; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v26, v26 |
| ; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v24, v24 |
| ; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v31 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v34 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v48 |
| ; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v39 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v38 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v37 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v36 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v35 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v33 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v32 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v29 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v26 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v24 |
| ; SI-NEXT: v_alignbit_b32 v39, v1, v39, 16 |
| ; SI-NEXT: v_alignbit_b32 v38, v3, v38, 16 |
| ; SI-NEXT: v_alignbit_b32 v37, v5, v37, 16 |
| ; SI-NEXT: v_alignbit_b32 v36, v7, v36, 16 |
| ; SI-NEXT: v_alignbit_b32 v35, v9, v35, 16 |
| ; SI-NEXT: v_alignbit_b32 v33, v11, v33, 16 |
| ; SI-NEXT: v_alignbit_b32 v32, v13, v32, 16 |
| ; SI-NEXT: v_alignbit_b32 v29, v15, v29, 16 |
| ; SI-NEXT: v_alignbit_b32 v26, v17, v26, 16 |
| ; SI-NEXT: v_alignbit_b32 v24, v19, v24, 16 |
| ; SI-NEXT: .LBB58_2: ; %end |
| ; SI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v20 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v37 |
| ; SI-NEXT: v_or_b32_e32 v4, v4, v20 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v31 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v20 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v36 |
| ; SI-NEXT: v_or_b32_e32 v6, v6, v20 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v30 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v20 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v35 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v20 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v28 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v20 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v33 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v20 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v27 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v20 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v32 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v20 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v25 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v20 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v29 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v20 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v23 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v20 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v26 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v20 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v22 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v20 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v24 |
| ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v34 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v38 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v21 |
| ; SI-NEXT: v_or_b32_e32 v0, v0, v39 |
| ; SI-NEXT: v_or_b32_e32 v2, v2, v34 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40f16_to_v40i16: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v19 |
| ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v18 |
| ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v17 |
| ; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v16 |
| ; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v15 |
| ; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v14 |
| ; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v13 |
| ; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v12 |
| ; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v11 |
| ; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v9 |
| ; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v8 |
| ; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v7 |
| ; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v6 |
| ; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v5 |
| ; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v4 |
| ; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; 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 .LBB58_2 |
| ; VI-NEXT: ; %bb.1: ; %cmp.true |
| ; VI-NEXT: v_add_f16_e32 v0, 0x200, v0 |
| ; VI-NEXT: v_add_f16_e32 v39, 0x200, v39 |
| ; VI-NEXT: v_add_f16_e32 v1, 0x200, v1 |
| ; VI-NEXT: v_add_f16_e32 v38, 0x200, v38 |
| ; VI-NEXT: v_add_f16_e32 v2, 0x200, v2 |
| ; VI-NEXT: v_add_f16_e32 v37, 0x200, v37 |
| ; VI-NEXT: v_add_f16_e32 v3, 0x200, v3 |
| ; VI-NEXT: v_add_f16_e32 v36, 0x200, v36 |
| ; VI-NEXT: v_add_f16_e32 v4, 0x200, v4 |
| ; VI-NEXT: v_add_f16_e32 v20, 0x200, v20 |
| ; VI-NEXT: v_add_f16_e32 v5, 0x200, v5 |
| ; VI-NEXT: v_add_f16_e32 v35, 0x200, v35 |
| ; VI-NEXT: v_add_f16_e32 v6, 0x200, v6 |
| ; VI-NEXT: v_add_f16_e32 v34, 0x200, v34 |
| ; VI-NEXT: v_add_f16_e32 v7, 0x200, v7 |
| ; VI-NEXT: v_add_f16_e32 v33, 0x200, v33 |
| ; VI-NEXT: v_add_f16_e32 v8, 0x200, v8 |
| ; VI-NEXT: v_add_f16_e32 v32, 0x200, v32 |
| ; VI-NEXT: v_add_f16_e32 v9, 0x200, v9 |
| ; VI-NEXT: v_add_f16_e32 v31, 0x200, v31 |
| ; VI-NEXT: v_add_f16_e32 v10, 0x200, v10 |
| ; VI-NEXT: v_add_f16_e32 v30, 0x200, v30 |
| ; VI-NEXT: v_add_f16_e32 v11, 0x200, v11 |
| ; VI-NEXT: v_add_f16_e32 v29, 0x200, v29 |
| ; VI-NEXT: v_add_f16_e32 v12, 0x200, v12 |
| ; VI-NEXT: v_add_f16_e32 v28, 0x200, v28 |
| ; VI-NEXT: v_add_f16_e32 v13, 0x200, v13 |
| ; VI-NEXT: v_add_f16_e32 v27, 0x200, v27 |
| ; VI-NEXT: v_add_f16_e32 v14, 0x200, v14 |
| ; VI-NEXT: v_add_f16_e32 v26, 0x200, v26 |
| ; VI-NEXT: v_add_f16_e32 v15, 0x200, v15 |
| ; VI-NEXT: v_add_f16_e32 v25, 0x200, v25 |
| ; VI-NEXT: v_add_f16_e32 v16, 0x200, v16 |
| ; VI-NEXT: v_add_f16_e32 v24, 0x200, v24 |
| ; VI-NEXT: v_add_f16_e32 v17, 0x200, v17 |
| ; VI-NEXT: v_add_f16_e32 v23, 0x200, v23 |
| ; VI-NEXT: v_add_f16_e32 v18, 0x200, v18 |
| ; VI-NEXT: v_add_f16_e32 v22, 0x200, v22 |
| ; VI-NEXT: v_add_f16_e32 v19, 0x200, v19 |
| ; VI-NEXT: v_add_f16_e32 v21, 0x200, v21 |
| ; VI-NEXT: .LBB58_2: ; %end |
| ; VI-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v35 |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v34 |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v33 |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v32 |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v31 |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v30 |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v29 |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v28 |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v27 |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v26 |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v25 |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v24 |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v23 |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v21 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40f16_to_v40i16: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v19 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v5 |
| ; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v20 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v0 |
| ; 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 .LBB58_2 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX9-NEXT: s_mov_b32 s6, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v19, v39, v19, s6 |
| ; GFX9-NEXT: v_perm_b32 v18, v38, v18, s6 |
| ; GFX9-NEXT: v_perm_b32 v17, v37, v17, s6 |
| ; GFX9-NEXT: v_perm_b32 v16, v36, v16, s6 |
| ; GFX9-NEXT: v_perm_b32 v15, v35, v15, s6 |
| ; GFX9-NEXT: v_perm_b32 v14, v34, v14, s6 |
| ; GFX9-NEXT: v_perm_b32 v13, v33, v13, s6 |
| ; GFX9-NEXT: v_perm_b32 v12, v32, v12, s6 |
| ; GFX9-NEXT: v_perm_b32 v11, v31, v11, s6 |
| ; GFX9-NEXT: v_perm_b32 v10, v30, v10, s6 |
| ; GFX9-NEXT: v_perm_b32 v9, v29, v9, s6 |
| ; GFX9-NEXT: s_movk_i32 s7, 0x200 |
| ; GFX9-NEXT: v_perm_b32 v8, v28, v8, s6 |
| ; GFX9-NEXT: v_perm_b32 v7, v27, v7, s6 |
| ; GFX9-NEXT: v_perm_b32 v6, v26, v6, s6 |
| ; GFX9-NEXT: v_perm_b32 v5, v24, v5, s6 |
| ; GFX9-NEXT: v_perm_b32 v4, v25, v4, s6 |
| ; GFX9-NEXT: v_perm_b32 v3, v23, v3, s6 |
| ; GFX9-NEXT: v_perm_b32 v2, v22, v2, s6 |
| ; GFX9-NEXT: v_perm_b32 v1, v21, v1, s6 |
| ; GFX9-NEXT: v_perm_b32 v0, v20, v0, s6 |
| ; GFX9-NEXT: v_pk_add_f16 v19, v19, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v18, v18, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v17, v17, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v16, v16, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v15, v15, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v14, v14, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v13, v13, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v12, v12, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v5, v5, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v4, v4, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v3, v3, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v2, v2, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v1, v1, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_pk_add_f16 v0, v0, s7 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v19 |
| ; GFX9-NEXT: .LBB58_2: ; %end |
| ; GFX9-NEXT: s_or_b64 exec, exec, s[4:5] |
| ; GFX9-NEXT: s_mov_b32 s4, 0x5040100 |
| ; GFX9-NEXT: v_perm_b32 v0, v20, v0, s4 |
| ; GFX9-NEXT: v_perm_b32 v1, v21, v1, s4 |
| ; GFX9-NEXT: v_perm_b32 v2, v22, v2, s4 |
| ; GFX9-NEXT: v_perm_b32 v3, v23, v3, s4 |
| ; GFX9-NEXT: v_perm_b32 v4, v25, v4, s4 |
| ; GFX9-NEXT: v_perm_b32 v5, v24, v5, s4 |
| ; GFX9-NEXT: v_perm_b32 v6, v26, v6, s4 |
| ; GFX9-NEXT: v_perm_b32 v7, v27, v7, s4 |
| ; GFX9-NEXT: v_perm_b32 v8, v28, v8, s4 |
| ; GFX9-NEXT: v_perm_b32 v9, v29, v9, s4 |
| ; GFX9-NEXT: v_perm_b32 v10, v30, v10, s4 |
| ; GFX9-NEXT: v_perm_b32 v11, v31, v11, s4 |
| ; GFX9-NEXT: v_perm_b32 v12, v32, v12, s4 |
| ; GFX9-NEXT: v_perm_b32 v13, v33, v13, s4 |
| ; GFX9-NEXT: v_perm_b32 v14, v34, v14, s4 |
| ; GFX9-NEXT: v_perm_b32 v15, v35, v15, s4 |
| ; GFX9-NEXT: v_perm_b32 v16, v36, v16, s4 |
| ; GFX9-NEXT: v_perm_b32 v17, v37, v17, s4 |
| ; GFX9-NEXT: v_perm_b32 v18, v38, v18, s4 |
| ; GFX9-NEXT: v_perm_b32 v19, v39, v19, s4 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v40f16_to_v40i16: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB58_2 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: .LBB58_2: ; %end |
| ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v40f16_to_v40i16: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v0 |
| ; GFX11-FAKE16-NEXT: s_mov_b32 s0, exec_lo |
| ; GFX11-FAKE16-NEXT: v_cmpx_ne_u32_e32 0, v20 |
| ; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB58_2 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v48, v19, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v39, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v38, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v37, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v36, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v35, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v34, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v33, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v32, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v31, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v30, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v29, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v28, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v27, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v26, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v21, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v22, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v23, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v24, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v25, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v15 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v19 |
| ; GFX11-FAKE16-NEXT: .LBB58_2: ; %end |
| ; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v0, v21, v0, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v1, v22, v1, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v2, v23, v2, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v3, v24, v3, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v4, v25, v4, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v5, v26, v5, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v6, v27, v6, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v7, v28, v7, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v8, v29, v8, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v9, v30, v9, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v10, v31, v10, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v11, v32, v11, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v12, v33, v12, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v13, v34, v13, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v14, v35, v14, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v15, v36, v15, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v16, v37, v16, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v17, v38, v17, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v18, v39, v18, 0x5040100 |
| ; GFX11-FAKE16-NEXT: v_perm_b32 v19, v48, v19, 0x5040100 |
| ; GFX11-FAKE16-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 <40 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <40 x half> %a1 to <40 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x half> %a to <40 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <40 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x i16> %phi |
| } |
| |
| define inreg <40 x i16> @bitcast_v40f16_to_v40i16_scalar(<40 x half> inreg %a, i32 inreg %b) { |
| ; SI-LABEL: bitcast_v40f16_to_v40i16_scalar: |
| ; SI: ; %bb.0: |
| ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; SI-NEXT: v_readfirstlane_b32 s9, v5 |
| ; SI-NEXT: v_readfirstlane_b32 s6, v4 |
| ; SI-NEXT: v_readfirstlane_b32 s10, v3 |
| ; SI-NEXT: v_readfirstlane_b32 s7, v2 |
| ; SI-NEXT: v_readfirstlane_b32 s41, v1 |
| ; SI-NEXT: v_readfirstlane_b32 s8, v0 |
| ; SI-NEXT: s_lshr_b32 s42, s29, 16 |
| ; SI-NEXT: s_lshr_b32 s46, s28, 16 |
| ; SI-NEXT: s_lshr_b32 s40, s27, 16 |
| ; SI-NEXT: s_lshr_b32 s57, s26, 16 |
| ; SI-NEXT: s_lshr_b32 s15, s25, 16 |
| ; SI-NEXT: s_lshr_b32 s59, s24, 16 |
| ; SI-NEXT: s_lshr_b32 s14, s23, 16 |
| ; SI-NEXT: s_lshr_b32 s60, s22, 16 |
| ; SI-NEXT: s_lshr_b32 s13, s21, 16 |
| ; SI-NEXT: s_lshr_b32 s61, s20, 16 |
| ; SI-NEXT: s_lshr_b32 s12, s19, 16 |
| ; SI-NEXT: s_lshr_b32 s62, s18, 16 |
| ; SI-NEXT: s_lshr_b32 s11, s17, 16 |
| ; SI-NEXT: s_lshr_b32 s63, s16, 16 |
| ; SI-NEXT: s_lshr_b32 s58, s9, 16 |
| ; SI-NEXT: s_lshr_b32 s44, s6, 16 |
| ; SI-NEXT: s_lshr_b32 s43, s10, 16 |
| ; SI-NEXT: s_lshr_b32 s45, s7, 16 |
| ; SI-NEXT: s_lshr_b32 s47, s41, 16 |
| ; SI-NEXT: s_lshr_b32 s56, s8, 16 |
| ; SI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; SI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill |
| ; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 ; 4-byte Folded Spill |
| ; SI-NEXT: s_cbranch_scc0 .LBB59_3 |
| ; SI-NEXT: ; %bb.1: ; %cmp.false |
| ; SI-NEXT: s_cbranch_execnz .LBB59_4 |
| ; SI-NEXT: .LBB59_2: ; %cmp.true |
| ; SI-NEXT: v_cvt_f32_f16_e32 v0, s63 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, s16 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v2, s62 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s18 |
| ; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 |
| ; SI-NEXT: v_or_b32_e32 v48, v1, v0 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, s61 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s20 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 |
| ; SI-NEXT: v_or_b32_e32 v49, v3, v2 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, s60 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v20, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s59 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s22 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, s57 |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s46 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s56 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, s45 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v1 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s44 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v16, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, s58 |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s47 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v18, v3 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v50, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, s41 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s42 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, s29 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v51, v3 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v52, v5 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v51 |
| ; SI-NEXT: v_or_b32_e32 v15, v1, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, s40 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v52 |
| ; SI-NEXT: v_or_b32_e32 v13, v5, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s27 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s15 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v53, v1 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v54, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v53 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s25 |
| ; SI-NEXT: v_or_b32_e32 v11, v1, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s14 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v7, s23 |
| ; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v55, v3 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v7 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v54 |
| ; SI-NEXT: v_or_b32_e32 v9, v5, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v1, s13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v55 |
| ; SI-NEXT: v_or_b32_e32 v7, v3, v5 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s21 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v5, s12 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v41, v1 |
| ; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v5 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v42, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v41 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, s19 |
| ; SI-NEXT: v_or_b32_e32 v5, v1, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v3, s11 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v23, s17 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v43, v3 |
| ; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v23 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v23, v3 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v17, s43 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v42 |
| ; SI-NEXT: v_or_b32_e32 v3, v19, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v19, s10 |
| ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v43 |
| ; SI-NEXT: v_or_b32_e32 v1, v23, v1 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v23, s9 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v40, v17 |
| ; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v19 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17 |
| ; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v23 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v40 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v23 |
| ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v50 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v22, s24 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v23 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v23, s26 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v22, v22 |
| ; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v23, v23 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_or_b32_e32 v59, v20, v4 |
| ; SI-NEXT: v_or_b32_e32 v57, v21, v6 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v20, s28 |
| ; SI-NEXT: v_or_b32_e32 v56, v22, v8 |
| ; SI-NEXT: v_or_b32_e32 v46, v23, v10 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v21, s8 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v22, s7 |
| ; SI-NEXT: v_cvt_f32_f16_e32 v23, s6 |
| ; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20 |
| ; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21 |
| ; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22 |
| ; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v22, v22 |
| ; SI-NEXT: v_cvt_f16_f32_e32 v23, v23 |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16 |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18 |
| ; SI-NEXT: v_or_b32_e32 v58, v20, v12 |
| ; SI-NEXT: v_or_b32_e32 v47, v21, v14 |
| ; SI-NEXT: v_or_b32_e32 v45, v22, v16 |
| ; SI-NEXT: v_or_b32_e32 v44, v23, v18 |
| ; SI-NEXT: v_lshr_b64 v[38:39], v[0:1], 16 |
| ; SI-NEXT: v_lshr_b64 v[36:37], v[2:3], 16 |
| ; SI-NEXT: v_lshr_b64 v[34:35], v[4:5], 16 |
| ; SI-NEXT: v_lshr_b64 v[32:33], v[6:7], 16 |
| ; SI-NEXT: v_lshr_b64 v[30:31], v[8:9], 16 |
| ; SI-NEXT: v_lshr_b64 v[28:29], v[10:11], 16 |
| ; SI-NEXT: v_lshr_b64 v[26:27], v[12:13], 16 |
| ; SI-NEXT: v_lshr_b64 v[24:25], v[14:15], 16 |
| ; SI-NEXT: v_lshr_b64 v[22:23], v[16:17], 16 |
| ; SI-NEXT: v_lshr_b64 v[20:21], v[18:19], 16 |
| ; SI-NEXT: s_branch .LBB59_5 |
| ; SI-NEXT: .LBB59_3: |
| ; SI-NEXT: s_branch .LBB59_2 |
| ; SI-NEXT: .LBB59_4: |
| ; SI-NEXT: v_mov_b32_e32 v50, s58 |
| ; SI-NEXT: v_mov_b32_e32 v40, s43 |
| ; SI-NEXT: v_mov_b32_e32 v51, s47 |
| ; SI-NEXT: v_mov_b32_e32 v52, s42 |
| ; SI-NEXT: v_mov_b32_e32 v53, s40 |
| ; SI-NEXT: v_mov_b32_e32 v54, s15 |
| ; SI-NEXT: v_mov_b32_e32 v55, s14 |
| ; SI-NEXT: v_mov_b32_e32 v41, s13 |
| ; SI-NEXT: v_mov_b32_e32 v42, s12 |
| ; SI-NEXT: v_mov_b32_e32 v43, s11 |
| ; SI-NEXT: v_mov_b32_e32 v44, s6 |
| ; SI-NEXT: s_waitcnt expcnt(6) |
| ; SI-NEXT: v_mov_b32_e32 v45, s7 |
| ; SI-NEXT: s_waitcnt expcnt(4) |
| ; SI-NEXT: v_mov_b32_e32 v47, s8 |
| ; SI-NEXT: s_waitcnt expcnt(1) |
| ; SI-NEXT: v_mov_b32_e32 v58, s28 |
| ; SI-NEXT: v_mov_b32_e32 v46, s26 |
| ; SI-NEXT: v_mov_b32_e32 v56, s24 |
| ; SI-NEXT: v_mov_b32_e32 v57, s22 |
| ; SI-NEXT: s_waitcnt expcnt(0) |
| ; SI-NEXT: v_mov_b32_e32 v59, s20 |
| ; SI-NEXT: v_mov_b32_e32 v49, s18 |
| ; SI-NEXT: v_mov_b32_e32 v48, s16 |
| ; SI-NEXT: v_mov_b32_e32 v1, s17 |
| ; SI-NEXT: v_mov_b32_e32 v3, s19 |
| ; SI-NEXT: v_mov_b32_e32 v5, s21 |
| ; SI-NEXT: v_mov_b32_e32 v7, s23 |
| ; SI-NEXT: v_mov_b32_e32 v9, s25 |
| ; SI-NEXT: v_mov_b32_e32 v11, s27 |
| ; SI-NEXT: v_mov_b32_e32 v13, s29 |
| ; SI-NEXT: v_mov_b32_e32 v15, s41 |
| ; SI-NEXT: v_mov_b32_e32 v17, s10 |
| ; SI-NEXT: v_mov_b32_e32 v19, s9 |
| ; SI-NEXT: v_mov_b32_e32 v38, s63 |
| ; SI-NEXT: v_mov_b32_e32 v36, s62 |
| ; SI-NEXT: v_mov_b32_e32 v34, s61 |
| ; SI-NEXT: v_mov_b32_e32 v32, s60 |
| ; SI-NEXT: v_mov_b32_e32 v30, s59 |
| ; SI-NEXT: v_mov_b32_e32 v28, s57 |
| ; SI-NEXT: v_mov_b32_e32 v26, s46 |
| ; SI-NEXT: v_mov_b32_e32 v24, s56 |
| ; SI-NEXT: v_mov_b32_e32 v22, s45 |
| ; SI-NEXT: v_mov_b32_e32 v20, s44 |
| ; SI-NEXT: .LBB59_5: ; %end |
| ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v38 |
| ; SI-NEXT: v_and_b32_e32 v2, 0xffff, v48 |
| ; SI-NEXT: v_or_b32_e32 v0, v2, v0 |
| ; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v43 |
| ; SI-NEXT: v_or_b32_e32 v1, v1, v2 |
| ; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v36 |
| ; SI-NEXT: v_and_b32_e32 v4, 0xffff, v49 |
| ; SI-NEXT: v_or_b32_e32 v2, v4, v2 |
| ; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v42 |
| ; SI-NEXT: v_or_b32_e32 v3, v3, v4 |
| ; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v34 |
| ; SI-NEXT: v_and_b32_e32 v6, 0xffff, v59 |
| ; SI-NEXT: v_or_b32_e32 v4, v6, v4 |
| ; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v41 |
| ; SI-NEXT: v_or_b32_e32 v5, v5, v6 |
| ; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v32 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v57 |
| ; SI-NEXT: v_or_b32_e32 v6, v8, v6 |
| ; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v55 |
| ; SI-NEXT: v_or_b32_e32 v7, v7, v8 |
| ; SI-NEXT: v_and_b32_e32 v8, 0xffff, v56 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v30 |
| ; SI-NEXT: v_or_b32_e32 v8, v8, v10 |
| ; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v54 |
| ; SI-NEXT: v_or_b32_e32 v9, v9, v10 |
| ; SI-NEXT: v_and_b32_e32 v10, 0xffff, v46 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v28 |
| ; SI-NEXT: v_or_b32_e32 v10, v10, v12 |
| ; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v53 |
| ; SI-NEXT: v_or_b32_e32 v11, v11, v12 |
| ; SI-NEXT: v_and_b32_e32 v12, 0xffff, v58 |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v26 |
| ; SI-NEXT: v_or_b32_e32 v12, v12, v14 |
| ; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v52 |
| ; SI-NEXT: v_or_b32_e32 v13, v13, v14 |
| ; SI-NEXT: v_and_b32_e32 v14, 0xffff, v47 |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v24 |
| ; SI-NEXT: v_or_b32_e32 v14, v14, v16 |
| ; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v51 |
| ; SI-NEXT: v_or_b32_e32 v15, v15, v16 |
| ; SI-NEXT: v_and_b32_e32 v16, 0xffff, v45 |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v22 |
| ; SI-NEXT: v_or_b32_e32 v16, v16, v18 |
| ; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v40 |
| ; SI-NEXT: v_or_b32_e32 v17, v17, v18 |
| ; SI-NEXT: v_and_b32_e32 v18, 0xffff, v44 |
| ; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload |
| ; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; SI-NEXT: v_or_b32_e32 v18, v18, v20 |
| ; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v50 |
| ; SI-NEXT: v_or_b32_e32 v19, v19, v20 |
| ; SI-NEXT: s_waitcnt vmcnt(0) |
| ; SI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; VI-LABEL: bitcast_v40f16_to_v40i16_scalar: |
| ; VI: ; %bb.0: |
| ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; VI-NEXT: v_readfirstlane_b32 s44, v5 |
| ; VI-NEXT: v_readfirstlane_b32 s46, v4 |
| ; VI-NEXT: v_readfirstlane_b32 s56, v3 |
| ; VI-NEXT: v_readfirstlane_b32 s58, v2 |
| ; VI-NEXT: v_readfirstlane_b32 s60, v1 |
| ; VI-NEXT: v_readfirstlane_b32 s62, v0 |
| ; VI-NEXT: s_lshr_b32 s6, s29, 16 |
| ; VI-NEXT: s_lshr_b32 s7, s28, 16 |
| ; VI-NEXT: s_lshr_b32 s8, s27, 16 |
| ; VI-NEXT: s_lshr_b32 s9, s26, 16 |
| ; VI-NEXT: s_lshr_b32 s10, s25, 16 |
| ; VI-NEXT: s_lshr_b32 s11, s24, 16 |
| ; VI-NEXT: s_lshr_b32 s12, s23, 16 |
| ; VI-NEXT: s_lshr_b32 s13, s22, 16 |
| ; VI-NEXT: s_lshr_b32 s14, s21, 16 |
| ; VI-NEXT: s_lshr_b32 s15, s20, 16 |
| ; VI-NEXT: s_lshr_b32 s40, s19, 16 |
| ; VI-NEXT: s_lshr_b32 s41, s18, 16 |
| ; VI-NEXT: s_lshr_b32 s42, s17, 16 |
| ; VI-NEXT: s_lshr_b32 s43, s16, 16 |
| ; VI-NEXT: s_lshr_b32 s45, s44, 16 |
| ; VI-NEXT: s_lshr_b32 s47, s46, 16 |
| ; VI-NEXT: s_lshr_b32 s57, s56, 16 |
| ; VI-NEXT: s_lshr_b32 s59, s58, 16 |
| ; VI-NEXT: s_lshr_b32 s61, s60, 16 |
| ; VI-NEXT: s_lshr_b32 s63, s62, 16 |
| ; VI-NEXT: v_readfirstlane_b32 s4, v6 |
| ; VI-NEXT: s_cmp_lg_u32 s4, 0 |
| ; VI-NEXT: s_cbranch_scc0 .LBB59_3 |
| ; VI-NEXT: ; %bb.1: ; %cmp.false |
| ; VI-NEXT: s_cbranch_execnz .LBB59_4 |
| ; VI-NEXT: .LBB59_2: ; %cmp.true |
| ; VI-NEXT: v_mov_b32_e32 v20, 0x200 |
| ; VI-NEXT: v_add_f16_e32 v0, s16, v20 |
| ; VI-NEXT: v_add_f16_e32 v39, s43, v20 |
| ; VI-NEXT: v_add_f16_e32 v1, s17, v20 |
| ; VI-NEXT: v_add_f16_e32 v38, s42, v20 |
| ; VI-NEXT: v_add_f16_e32 v2, s18, v20 |
| ; VI-NEXT: v_add_f16_e32 v37, s41, v20 |
| ; VI-NEXT: v_add_f16_e32 v3, s19, v20 |
| ; VI-NEXT: v_add_f16_e32 v36, s40, v20 |
| ; VI-NEXT: v_add_f16_e32 v4, s20, v20 |
| ; VI-NEXT: v_add_f16_e32 v35, s15, v20 |
| ; VI-NEXT: v_add_f16_e32 v5, s21, v20 |
| ; VI-NEXT: v_add_f16_e32 v34, s14, v20 |
| ; VI-NEXT: v_add_f16_e32 v6, s22, v20 |
| ; VI-NEXT: v_add_f16_e32 v33, s13, v20 |
| ; VI-NEXT: v_add_f16_e32 v7, s23, v20 |
| ; VI-NEXT: v_add_f16_e32 v32, s12, v20 |
| ; VI-NEXT: v_add_f16_e32 v8, s24, v20 |
| ; VI-NEXT: v_add_f16_e32 v31, s11, v20 |
| ; VI-NEXT: v_add_f16_e32 v9, s25, v20 |
| ; VI-NEXT: v_add_f16_e32 v30, s10, v20 |
| ; VI-NEXT: v_add_f16_e32 v10, s26, v20 |
| ; VI-NEXT: v_add_f16_e32 v29, s9, v20 |
| ; VI-NEXT: v_add_f16_e32 v11, s27, v20 |
| ; VI-NEXT: v_add_f16_e32 v28, s8, v20 |
| ; VI-NEXT: v_add_f16_e32 v12, s28, v20 |
| ; VI-NEXT: v_add_f16_e32 v27, s7, v20 |
| ; VI-NEXT: v_add_f16_e32 v13, s29, v20 |
| ; VI-NEXT: v_add_f16_e32 v26, s6, v20 |
| ; VI-NEXT: v_add_f16_e32 v14, s62, v20 |
| ; VI-NEXT: v_add_f16_e32 v25, s63, v20 |
| ; VI-NEXT: v_add_f16_e32 v15, s60, v20 |
| ; VI-NEXT: v_add_f16_e32 v24, s61, v20 |
| ; VI-NEXT: v_add_f16_e32 v16, s58, v20 |
| ; VI-NEXT: v_add_f16_e32 v23, s59, v20 |
| ; VI-NEXT: v_add_f16_e32 v17, s56, v20 |
| ; VI-NEXT: v_add_f16_e32 v22, s57, v20 |
| ; VI-NEXT: v_add_f16_e32 v18, s46, v20 |
| ; VI-NEXT: v_add_f16_e32 v21, s47, v20 |
| ; VI-NEXT: v_add_f16_e32 v19, s44, v20 |
| ; VI-NEXT: v_add_f16_e32 v20, s45, v20 |
| ; VI-NEXT: s_branch .LBB59_5 |
| ; VI-NEXT: .LBB59_3: |
| ; VI-NEXT: s_branch .LBB59_2 |
| ; VI-NEXT: .LBB59_4: |
| ; VI-NEXT: v_mov_b32_e32 v20, s45 |
| ; VI-NEXT: v_mov_b32_e32 v19, s44 |
| ; VI-NEXT: v_mov_b32_e32 v21, s47 |
| ; VI-NEXT: v_mov_b32_e32 v18, s46 |
| ; VI-NEXT: v_mov_b32_e32 v22, s57 |
| ; VI-NEXT: v_mov_b32_e32 v17, s56 |
| ; VI-NEXT: v_mov_b32_e32 v23, s59 |
| ; VI-NEXT: v_mov_b32_e32 v16, s58 |
| ; VI-NEXT: v_mov_b32_e32 v24, s61 |
| ; VI-NEXT: v_mov_b32_e32 v15, s60 |
| ; VI-NEXT: v_mov_b32_e32 v25, s63 |
| ; VI-NEXT: v_mov_b32_e32 v14, s62 |
| ; VI-NEXT: v_mov_b32_e32 v26, s6 |
| ; VI-NEXT: v_mov_b32_e32 v13, s29 |
| ; VI-NEXT: v_mov_b32_e32 v27, s7 |
| ; VI-NEXT: v_mov_b32_e32 v12, s28 |
| ; VI-NEXT: v_mov_b32_e32 v28, s8 |
| ; VI-NEXT: v_mov_b32_e32 v11, s27 |
| ; VI-NEXT: v_mov_b32_e32 v29, s9 |
| ; VI-NEXT: v_mov_b32_e32 v10, s26 |
| ; VI-NEXT: v_mov_b32_e32 v30, s10 |
| ; VI-NEXT: v_mov_b32_e32 v9, s25 |
| ; VI-NEXT: v_mov_b32_e32 v31, s11 |
| ; VI-NEXT: v_mov_b32_e32 v8, s24 |
| ; VI-NEXT: v_mov_b32_e32 v32, s12 |
| ; VI-NEXT: v_mov_b32_e32 v7, s23 |
| ; VI-NEXT: v_mov_b32_e32 v33, s13 |
| ; VI-NEXT: v_mov_b32_e32 v6, s22 |
| ; VI-NEXT: v_mov_b32_e32 v34, s14 |
| ; VI-NEXT: v_mov_b32_e32 v5, s21 |
| ; VI-NEXT: v_mov_b32_e32 v35, s15 |
| ; VI-NEXT: v_mov_b32_e32 v4, s20 |
| ; VI-NEXT: v_mov_b32_e32 v36, s40 |
| ; VI-NEXT: v_mov_b32_e32 v3, s19 |
| ; VI-NEXT: v_mov_b32_e32 v37, s41 |
| ; VI-NEXT: v_mov_b32_e32 v2, s18 |
| ; VI-NEXT: v_mov_b32_e32 v38, s42 |
| ; VI-NEXT: v_mov_b32_e32 v1, s17 |
| ; VI-NEXT: v_mov_b32_e32 v39, s43 |
| ; VI-NEXT: v_mov_b32_e32 v0, s16 |
| ; VI-NEXT: .LBB59_5: ; %end |
| ; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39 |
| ; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38 |
| ; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37 |
| ; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36 |
| ; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35 |
| ; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34 |
| ; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33 |
| ; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32 |
| ; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31 |
| ; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30 |
| ; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29 |
| ; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28 |
| ; VI-NEXT: v_lshlrev_b32_e32 v27, 16, v27 |
| ; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26 |
| ; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25 |
| ; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24 |
| ; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v23 |
| ; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v22 |
| ; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v21 |
| ; VI-NEXT: v_lshlrev_b32_e32 v20, 16, v20 |
| ; VI-NEXT: v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v1, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v2, v2, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v3, v3, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v5, v5, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v6, v6, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v7, v7, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v8, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v9, v9, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v10, v10, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v12, v12, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v13, v13, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v14, v14, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v16, v16, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v17, v17, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v18, v18, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| ; VI-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX9-LABEL: bitcast_v40f16_to_v40i16_scalar: |
| ; GFX9: ; %bb.0: |
| ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX9-NEXT: v_readfirstlane_b32 s63, v5 |
| ; GFX9-NEXT: v_readfirstlane_b32 s62, v4 |
| ; GFX9-NEXT: v_readfirstlane_b32 s61, v3 |
| ; GFX9-NEXT: v_readfirstlane_b32 s60, v2 |
| ; GFX9-NEXT: v_readfirstlane_b32 s59, v1 |
| ; GFX9-NEXT: v_readfirstlane_b32 s58, v0 |
| ; GFX9-NEXT: s_lshr_b32 s43, s29, 16 |
| ; GFX9-NEXT: s_lshr_b32 s42, s28, 16 |
| ; GFX9-NEXT: s_lshr_b32 s41, s27, 16 |
| ; GFX9-NEXT: s_lshr_b32 s40, s26, 16 |
| ; GFX9-NEXT: s_lshr_b32 s15, s25, 16 |
| ; GFX9-NEXT: s_lshr_b32 s14, s24, 16 |
| ; GFX9-NEXT: s_lshr_b32 s13, s23, 16 |
| ; GFX9-NEXT: s_lshr_b32 s12, s22, 16 |
| ; GFX9-NEXT: s_lshr_b32 s11, s21, 16 |
| ; GFX9-NEXT: s_lshr_b32 s10, s20, 16 |
| ; GFX9-NEXT: s_lshr_b32 s9, s19, 16 |
| ; GFX9-NEXT: s_lshr_b32 s8, s18, 16 |
| ; GFX9-NEXT: s_lshr_b32 s7, s17, 16 |
| ; GFX9-NEXT: s_lshr_b32 s6, s16, 16 |
| ; GFX9-NEXT: s_lshr_b32 s57, s63, 16 |
| ; GFX9-NEXT: s_lshr_b32 s56, s62, 16 |
| ; GFX9-NEXT: s_lshr_b32 s47, s61, 16 |
| ; GFX9-NEXT: s_lshr_b32 s46, s60, 16 |
| ; GFX9-NEXT: s_lshr_b32 s45, s59, 16 |
| ; GFX9-NEXT: s_lshr_b32 s44, s58, 16 |
| ; GFX9-NEXT: v_readfirstlane_b32 s4, v6 |
| ; GFX9-NEXT: s_cmp_lg_u32 s4, 0 |
| ; GFX9-NEXT: s_cbranch_scc0 .LBB59_3 |
| ; GFX9-NEXT: ; %bb.1: ; %cmp.false |
| ; GFX9-NEXT: s_cbranch_execnz .LBB59_4 |
| ; GFX9-NEXT: .LBB59_2: ; %cmp.true |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s63, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v0, 0x200 |
| ; GFX9-NEXT: v_pk_add_f16 v19, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s62, s56 |
| ; GFX9-NEXT: v_pk_add_f16 v18, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s61, s47 |
| ; GFX9-NEXT: v_pk_add_f16 v17, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s60, s46 |
| ; GFX9-NEXT: v_pk_add_f16 v16, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s59, s45 |
| ; GFX9-NEXT: v_pk_add_f16 v15, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s58, s44 |
| ; GFX9-NEXT: v_pk_add_f16 v14, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s29, s43 |
| ; GFX9-NEXT: v_pk_add_f16 v13, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s28, s42 |
| ; GFX9-NEXT: v_pk_add_f16 v12, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s27, s41 |
| ; GFX9-NEXT: v_pk_add_f16 v11, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s26, s40 |
| ; GFX9-NEXT: v_pk_add_f16 v10, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s25, s15 |
| ; GFX9-NEXT: v_pk_add_f16 v9, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s24, s14 |
| ; GFX9-NEXT: v_pk_add_f16 v8, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s23, s13 |
| ; GFX9-NEXT: v_pk_add_f16 v7, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s22, s12 |
| ; GFX9-NEXT: v_pk_add_f16 v6, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s21, s11 |
| ; GFX9-NEXT: v_pk_add_f16 v5, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s20, s10 |
| ; GFX9-NEXT: v_pk_add_f16 v4, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s19, s9 |
| ; GFX9-NEXT: v_pk_add_f16 v3, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s18, s8 |
| ; GFX9-NEXT: v_pk_add_f16 v2, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s17, s7 |
| ; GFX9-NEXT: v_pk_add_f16 v1, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s6 |
| ; GFX9-NEXT: v_pk_add_f16 v0, s4, v0 op_sel_hi:[1,0] |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX9-NEXT: s_branch .LBB59_5 |
| ; GFX9-NEXT: .LBB59_3: |
| ; GFX9-NEXT: s_branch .LBB59_2 |
| ; GFX9-NEXT: .LBB59_4: |
| ; GFX9-NEXT: v_mov_b32_e32 v19, s63 |
| ; GFX9-NEXT: v_mov_b32_e32 v18, s62 |
| ; GFX9-NEXT: v_mov_b32_e32 v17, s61 |
| ; GFX9-NEXT: v_mov_b32_e32 v16, s60 |
| ; GFX9-NEXT: v_mov_b32_e32 v15, s59 |
| ; GFX9-NEXT: v_mov_b32_e32 v14, s58 |
| ; GFX9-NEXT: v_mov_b32_e32 v13, s29 |
| ; GFX9-NEXT: v_mov_b32_e32 v12, s28 |
| ; GFX9-NEXT: v_mov_b32_e32 v11, s27 |
| ; GFX9-NEXT: v_mov_b32_e32 v10, s26 |
| ; GFX9-NEXT: v_mov_b32_e32 v9, s25 |
| ; GFX9-NEXT: v_mov_b32_e32 v8, s24 |
| ; GFX9-NEXT: v_mov_b32_e32 v7, s23 |
| ; GFX9-NEXT: v_mov_b32_e32 v6, s22 |
| ; GFX9-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX9-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX9-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX9-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX9-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX9-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX9-NEXT: v_mov_b32_e32 v20, s57 |
| ; GFX9-NEXT: v_mov_b32_e32 v21, s56 |
| ; GFX9-NEXT: v_mov_b32_e32 v22, s47 |
| ; GFX9-NEXT: v_mov_b32_e32 v23, s46 |
| ; GFX9-NEXT: v_mov_b32_e32 v24, s45 |
| ; GFX9-NEXT: v_mov_b32_e32 v25, s44 |
| ; GFX9-NEXT: v_mov_b32_e32 v26, s43 |
| ; GFX9-NEXT: v_mov_b32_e32 v27, s42 |
| ; GFX9-NEXT: v_mov_b32_e32 v28, s41 |
| ; GFX9-NEXT: v_mov_b32_e32 v29, s40 |
| ; GFX9-NEXT: v_mov_b32_e32 v30, s15 |
| ; GFX9-NEXT: v_mov_b32_e32 v31, s14 |
| ; GFX9-NEXT: v_mov_b32_e32 v32, s13 |
| ; GFX9-NEXT: v_mov_b32_e32 v33, s12 |
| ; GFX9-NEXT: v_mov_b32_e32 v34, s11 |
| ; GFX9-NEXT: v_mov_b32_e32 v35, s10 |
| ; GFX9-NEXT: v_mov_b32_e32 v36, s9 |
| ; GFX9-NEXT: v_mov_b32_e32 v37, s8 |
| ; GFX9-NEXT: v_mov_b32_e32 v38, s7 |
| ; GFX9-NEXT: v_mov_b32_e32 v39, s6 |
| ; GFX9-NEXT: .LBB59_5: ; %end |
| ; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1 |
| ; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5 |
| ; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6 |
| ; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10 |
| ; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11 |
| ; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15 |
| ; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v16 |
| ; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX9-NEXT: v_lshl_or_b32 v0, v39, 16, v0 |
| ; GFX9-NEXT: v_lshl_or_b32 v1, v38, 16, v1 |
| ; GFX9-NEXT: v_lshl_or_b32 v2, v37, 16, v2 |
| ; GFX9-NEXT: v_lshl_or_b32 v3, v36, 16, v3 |
| ; GFX9-NEXT: v_lshl_or_b32 v4, v35, 16, v4 |
| ; GFX9-NEXT: v_lshl_or_b32 v5, v34, 16, v5 |
| ; GFX9-NEXT: v_lshl_or_b32 v6, v33, 16, v6 |
| ; GFX9-NEXT: v_lshl_or_b32 v7, v32, 16, v7 |
| ; GFX9-NEXT: v_lshl_or_b32 v8, v31, 16, v8 |
| ; GFX9-NEXT: v_lshl_or_b32 v9, v30, 16, v9 |
| ; GFX9-NEXT: v_lshl_or_b32 v10, v29, 16, v10 |
| ; GFX9-NEXT: v_lshl_or_b32 v11, v28, 16, v11 |
| ; GFX9-NEXT: v_lshl_or_b32 v12, v27, 16, v12 |
| ; GFX9-NEXT: v_lshl_or_b32 v13, v26, 16, v13 |
| ; GFX9-NEXT: v_lshl_or_b32 v14, v25, 16, v14 |
| ; GFX9-NEXT: v_lshl_or_b32 v15, v24, 16, v15 |
| ; GFX9-NEXT: v_lshl_or_b32 v16, v23, 16, v16 |
| ; GFX9-NEXT: v_lshl_or_b32 v17, v22, 16, v17 |
| ; GFX9-NEXT: v_lshl_or_b32 v18, v21, 16, v18 |
| ; GFX9-NEXT: v_lshl_or_b32 v19, v20, 16, v19 |
| ; GFX9-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-TRUE16-LABEL: bitcast_v40f16_to_v40i16_scalar: |
| ; GFX11-TRUE16: ; %bb.0: |
| ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s57, v1 |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s56, v0 |
| ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s58, v2 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s29, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s28, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s27, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s26, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s25, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s24, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s23, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s22, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s21, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s20, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s19, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s18, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s17, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s16, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s3, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s2, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s1, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s0, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s47, s57, 16 |
| ; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s56, 16 |
| ; GFX11-TRUE16-NEXT: s_cmp_lg_u32 s58, 0 |
| ; GFX11-TRUE16-NEXT: s_mov_b32 s58, 0 |
| ; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB59_3 |
| ; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s58 |
| ; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB59_4 |
| ; GFX11-TRUE16-NEXT: .LBB59_2: ; %cmp.true |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s47, s57, s47 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s46, s56, s46 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s29, s29, s45 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s28, s28, s44 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s27, s27, s43 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s26, s26, s42 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s25, s25, s41 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s24, s24, s40 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s23, s15 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s22, s14 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s21, s13 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s20, s12 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s19, s11 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s18, s10 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s17, s9 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s16, s8 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s6 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s4 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s7 |
| ; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s5 |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, s47 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, s46 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, s29 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, s28 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s26 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, s25 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, s24 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, s15 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, s14 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s13 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, s12 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, s11 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, s10 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s9 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, s8 op_sel_hi:[0,1] |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v0 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v1 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v3 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v4 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v6 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v8 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v9 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v10 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v11 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v13 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v14 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v15 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 16, v16 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 16, v18 |
| ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 16, v19 |
| ; GFX11-TRUE16-NEXT: s_branch .LBB59_5 |
| ; GFX11-TRUE16-NEXT: .LBB59_3: |
| ; GFX11-TRUE16-NEXT: s_branch .LBB59_2 |
| ; GFX11-TRUE16-NEXT: .LBB59_4: |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, s57 :: v_dual_mov_b32 v18, s56 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, s29 :: v_dual_mov_b32 v16, s28 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v14, s26 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v13, s25 :: v_dual_mov_b32 v12, s24 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, s23 :: v_dual_mov_b32 v10, s22 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, s21 :: v_dual_mov_b32 v8, s20 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, s19 :: v_dual_mov_b32 v6, s18 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v4, s16 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v2, s2 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v0, s0 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s47 :: v_dual_mov_b32 v21, s46 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v23, s44 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v24, s43 :: v_dual_mov_b32 v25, s42 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v27, s40 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, s13 :: v_dual_mov_b32 v31, s12 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v32, s11 :: v_dual_mov_b32 v33, s10 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v34, s9 :: v_dual_mov_b32 v35, s8 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v36, s6 :: v_dual_mov_b32 v37, s4 |
| ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v38, s5 :: v_dual_mov_b32 v39, s7 |
| ; GFX11-TRUE16-NEXT: .LBB59_5: ; %end |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v39.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v38.l |
| ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v37.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v36.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v35.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v34.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v33.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v32.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v31.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v30.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v29.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v28.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v27.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v26.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v25.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v24.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v23.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v22.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v21.l |
| ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v20.l |
| ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] |
| ; |
| ; GFX11-FAKE16-LABEL: bitcast_v40f16_to_v40i16_scalar: |
| ; GFX11-FAKE16: ; %bb.0: |
| ; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s57, v1 |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s56, v0 |
| ; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s58, v2 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s29, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s28, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s27, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s26, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s25, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s40, s24, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s23, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s22, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s21, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s20, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s19, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s18, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s17, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s16, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s3, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s2, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s1, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s0, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s47, s57, 16 |
| ; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s56, 16 |
| ; GFX11-FAKE16-NEXT: s_cmp_lg_u32 s58, 0 |
| ; GFX11-FAKE16-NEXT: s_mov_b32 s58, 0 |
| ; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB59_3 |
| ; GFX11-FAKE16-NEXT: ; %bb.1: ; %Flow |
| ; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s58 |
| ; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB59_4 |
| ; GFX11-FAKE16-NEXT: .LBB59_2: ; %cmp.true |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s47, s57, s47 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s46, s56, s46 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s29, s29, s45 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s28, s28, s44 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s27, s27, s43 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s26, s26, s42 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s25, s25, s41 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s24, s24, s40 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s15, s23, s15 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s14, s22, s14 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s13, s21, s13 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s12, s20, s12 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s11, s19, s11 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s10, s18, s10 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s9, s17, s9 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s16, s8 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s6 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s4 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s0, s7 |
| ; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s1, s5 |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, s47 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, s46 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, s29 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v18, 0x200, s28 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v19, 0x200, s27 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, s26 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, s25 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, s24 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, s15 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, s13 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, s12 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, s11 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, s10 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v1, 0x200, s3 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, s8 op_sel_hi:[0,1] |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v1 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v0 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v6 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v11 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v10 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v24, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v16 |
| ; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v15 |
| ; GFX11-FAKE16-NEXT: s_branch .LBB59_5 |
| ; GFX11-FAKE16-NEXT: .LBB59_3: |
| ; GFX11-FAKE16-NEXT: s_branch .LBB59_2 |
| ; GFX11-FAKE16-NEXT: .LBB59_4: |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v15, s57 :: v_dual_mov_b32 v16, s56 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v17, s29 :: v_dual_mov_b32 v18, s28 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v19, s27 :: v_dual_mov_b32 v10, s26 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s25 :: v_dual_mov_b32 v12, s24 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v13, s23 :: v_dual_mov_b32 v14, s22 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, s21 :: v_dual_mov_b32 v6, s20 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v7, s19 :: v_dual_mov_b32 v8, s18 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v9, s17 :: v_dual_mov_b32 v0, s16 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v1, s3 :: v_dual_mov_b32 v2, s2 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v4, s0 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v20, s47 :: v_dual_mov_b32 v21, s46 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v23, s44 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s43 :: v_dual_mov_b32 v25, s42 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v27, s40 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v30, s13 :: v_dual_mov_b32 v31, s12 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v32, s11 :: v_dual_mov_b32 v33, s10 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v34, s9 :: v_dual_mov_b32 v35, s8 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v36, s6 :: v_dual_mov_b32 v37, s4 |
| ; GFX11-FAKE16-NEXT: v_dual_mov_b32 v38, s5 :: v_dual_mov_b32 v39, s7 |
| ; GFX11-FAKE16-NEXT: .LBB59_5: ; %end |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v1 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v0 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v39, 16, v4 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v38, 16, v3 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v36, 16, v48 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v35, 16, v49 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v6 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v36, 0xffff, v5 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v34, 16, v9 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v33, 16, v8 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v31, 16, v35 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v30, 16, v36 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v30, 0xffff, v11 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xffff, v10 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v29, 16, v14 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v28, 16, v13 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v26, 16, v30 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v25, 16, v31 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v18 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v16 |
| ; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v15 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v37, 16, v2 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v32, 16, v7 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v27, 16, v12 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v24, 16, v19 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v23, 16, v18 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v22, 16, v17 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v21, 16, v25 |
| ; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v20, 16, v26 |
| ; GFX11-FAKE16-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 <40 x half> %a, splat (half 0xH0200) |
| %a2 = bitcast <40 x half> %a1 to <40 x i16> |
| br label %end |
| |
| cmp.false: |
| %a3 = bitcast <40 x half> %a to <40 x i16> |
| br label %end |
| |
| end: |
| %phi = phi <40 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ] |
| ret <40 x i16> %phi |
| } |
