| ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2 |
| ; RUN: llc -mtriple=amdgcn-- -verify-machineinstrs < %s | FileCheck -check-prefix=GFX6 %s |
| ; RUN: llc -mtriple=amdgcn-- -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GFX8 %s |
| ; RUN: llc -mtriple=r600-- -mcpu=cypress < %s | FileCheck -check-prefix=EG %s |
| ; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX12 %s |
| |
| define amdgpu_kernel void @constant_load_i1(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_load_i1: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_and_b32_e32 v0, 1, v0 |
| ; GFX6-NEXT: buffer_store_byte v0, off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_load_i1: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_and_b32_e32 v2, 1, v2 |
| ; GFX8-NEXT: flat_store_byte v[0:1], v2 |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_load_i1: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 11, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT MSKOR T0.XW, T1.X |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: AND_INT T0.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: AND_INT * T1.W, T0.X, 1, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.x, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: LSHL T0.X, T1.W, PV.W, |
| ; EG-NEXT: LSHL * T0.W, literal.x, PV.W, |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) |
| ; EG-NEXT: MOV T0.Y, 0.0, |
| ; EG-NEXT: MOV * T0.Z, 0.0, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_load_i1: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_and_b32 s2, s2, 1 |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2 |
| ; GFX12-NEXT: global_store_b8 v0, v1, s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load i1, ptr addrspace(4) %in |
| store i1 %load, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_load_v2i1(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_load_v2i1: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: buffer_store_byte v0, off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_load_v2i1: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: flat_store_byte v[0:1], v2 |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_load_v2i1: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 11, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT MSKOR T0.XW, T1.X |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: AND_INT T0.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: AND_INT * T1.W, T0.X, literal.x, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.x, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: LSHL T0.X, T1.W, PV.W, |
| ; EG-NEXT: LSHL * T0.W, literal.x, PV.W, |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) |
| ; EG-NEXT: MOV T0.Y, 0.0, |
| ; EG-NEXT: MOV * T0.Z, 0.0, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_load_v2i1: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v0, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v1, v0, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: global_store_b8 v0, v1, s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <2 x i1>, ptr addrspace(4) %in |
| store <2 x i1> %load, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_load_v3i1(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_load_v3i1: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: buffer_store_byte v0, off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_load_v3i1: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: flat_store_byte v[0:1], v2 |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_load_v3i1: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 10, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT MSKOR T0.XW, T1.X |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: AND_INT * T0.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.x, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: LSHL T0.X, T0.X, PV.W, |
| ; EG-NEXT: LSHL * T0.W, literal.x, PV.W, |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) |
| ; EG-NEXT: MOV T0.Y, 0.0, |
| ; EG-NEXT: MOV * T0.Z, 0.0, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_load_v3i1: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v0, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v1, v0, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: global_store_b8 v0, v1, s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <3 x i1>, ptr addrspace(4) %in |
| store <3 x i1> %load, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_load_v4i1(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_load_v4i1: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: buffer_store_byte v0, off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_load_v4i1: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: flat_store_byte v[0:1], v2 |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_load_v4i1: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 11, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT MSKOR T0.XW, T1.X |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: AND_INT T0.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: AND_INT * T1.W, T0.X, literal.y, |
| ; EG-NEXT: 3(4.203895e-45), 15(2.101948e-44) |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.x, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: LSHL T0.X, T1.W, PV.W, |
| ; EG-NEXT: LSHL * T0.W, literal.x, PV.W, |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) |
| ; EG-NEXT: MOV T0.Y, 0.0, |
| ; EG-NEXT: MOV * T0.Z, 0.0, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_load_v4i1: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v0, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v1, v0, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: global_store_b8 v0, v1, s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <4 x i1>, ptr addrspace(4) %in |
| store <4 x i1> %load, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_load_v8i1(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_load_v8i1: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: buffer_store_byte v0, off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_load_v8i1: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: flat_store_byte v[0:1], v2 |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_load_v8i1: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 11, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT MSKOR T0.XW, T1.X |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: AND_INT T0.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: AND_INT * T1.W, T0.X, literal.y, |
| ; EG-NEXT: 3(4.203895e-45), 255(3.573311e-43) |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.x, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: LSHL T0.X, T1.W, PV.W, |
| ; EG-NEXT: LSHL * T0.W, literal.x, PV.W, |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) |
| ; EG-NEXT: MOV T0.Y, 0.0, |
| ; EG-NEXT: MOV * T0.Z, 0.0, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_load_v8i1: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v0, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v1, v0, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: global_store_b8 v0, v1, s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <8 x i1>, ptr addrspace(4) %in |
| store <8 x i1> %load, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_load_v16i1(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_load_v16i1: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ushort v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: buffer_store_short v0, off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_load_v16i1: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ushort v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: flat_store_short v[0:1], v2 |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_load_v16i1: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 11, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT MSKOR T0.XW, T1.X |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: AND_INT T0.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: AND_INT * T1.W, T0.X, literal.y, |
| ; EG-NEXT: 3(4.203895e-45), 65535(9.183409e-41) |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.x, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: LSHL T0.X, T1.W, PV.W, |
| ; EG-NEXT: LSHL * T0.W, literal.x, PV.W, |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) |
| ; EG-NEXT: MOV T0.Y, 0.0, |
| ; EG-NEXT: MOV * T0.Z, 0.0, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_load_v16i1: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v0, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u16 v1, v0, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: global_store_b16 v0, v1, s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <16 x i1>, ptr addrspace(4) %in |
| store <16 x i1> %load, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_load_v32i1(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_load_v32i1: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_load_dword s4, s[2:3], 0x0 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX6-NEXT: buffer_store_dword v0, off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_load_v32i1: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_load_dword s2, s[2:3], 0x0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s2 |
| ; GFX8-NEXT: flat_store_dword v[0:1], v2 |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_load_v32i1: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 1, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_load_v32i1: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_b32 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2 |
| ; GFX12-NEXT: global_store_b32 v0, v1, s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <32 x i1>, ptr addrspace(4) %in |
| store <32 x i1> %load, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_load_v64i1(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_load_v64i1: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s5 |
| ; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_load_v64i1: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s3 |
| ; GFX8-NEXT: flat_store_dwordx2 v[0:1], v[2:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_load_v64i1: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 1, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_64 T0.XY, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_load_v64i1: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x0 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3 |
| ; GFX12-NEXT: global_store_b64 v2, v[0:1], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <64 x i1>, ptr addrspace(4) %in |
| store <64 x i1> %load, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_i1_to_i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_i1_to_i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_i1_to_i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: flat_store_dword v[0:1], v2 |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_i1_to_i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 1, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_i1_to_i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2 |
| ; GFX12-NEXT: global_store_b32 v0, v1, s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %a = load i1, ptr addrspace(4) %in |
| %ext = zext i1 %a to i32 |
| store i32 %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_i1_to_i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_i1_to_i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_i1_to_i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX8-NEXT: flat_store_dword v[0:1], v2 |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_i1_to_i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 2, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_INT T0.X, T0.X, 0.0, 1, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_i1_to_i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_i32 s2, s2, 0x10000 |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2 |
| ; GFX12-NEXT: global_store_b32 v0, v1, s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %a = load i1, ptr addrspace(4) %in |
| %ext = sext i1 %a to i32 |
| store i32 %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v1i1_to_v1i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v1i1_to_v1i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v1i1_to_v1i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: flat_store_dword v[0:1], v2 |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v1i1_to_v1i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 1, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v1i1_to_v1i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2 |
| ; GFX12-NEXT: global_store_b32 v0, v1, s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <1 x i1>, ptr addrspace(4) %in |
| %ext = zext <1 x i1> %load to <1 x i32> |
| store <1 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v1i1_to_v1i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v1i1_to_v1i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v1i1_to_v1i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX8-NEXT: flat_store_dword v[0:1], v2 |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v1i1_to_v1i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 2, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_INT T0.X, T0.X, 0.0, 1, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v1i1_to_v1i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_i32 s2, s2, 0x10000 |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2 |
| ; GFX12-NEXT: global_store_b32 v0, v1, s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <1 x i1>, ptr addrspace(4) %in |
| %ext = sext <1 x i1> %load to <1 x i32> |
| store <1 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v2i1_to_v2i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v2i1_to_v2i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v1, 1, v0 |
| ; GFX6-NEXT: v_and_b32_e32 v0, 1, v0 |
| ; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v2i1_to_v2i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, 1 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 |
| ; GFX8-NEXT: v_and_b32_e32 v2, 1, v2 |
| ; GFX8-NEXT: flat_store_dwordx2 v[0:1], v[2:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v2i1_to_v2i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 3, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_UINT * T0.Y, T0.X, 1, 1, |
| ; EG-NEXT: AND_INT T0.X, T0.X, 1, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v2i1_to_v2i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v0, v2, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_and_b32_e32 v1, 0xffff, v0 |
| ; GFX12-NEXT: v_and_b32_e32 v0, 1, v0 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v1, 1, v1 |
| ; GFX12-NEXT: global_store_b64 v2, v[0:1], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <2 x i1>, ptr addrspace(4) %in |
| %ext = zext <2 x i1> %load to <2 x i32> |
| store <2 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v2i1_to_v2i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v2i1_to_v2i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_i32 v1, v0, 1, 1 |
| ; GFX6-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v2i1_to_v2i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_bfe_i32 v3, v2, 1, 1 |
| ; GFX8-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX8-NEXT: flat_store_dwordx2 v[0:1], v[2:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v2i1_to_v2i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 4, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T1.XY, T0.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_INT T1.X, T0.X, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T0.X, 1, |
| ; EG-NEXT: LSHR * T0.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT * T1.Y, PV.W, 0.0, 1, |
| ; |
| ; GFX12-LABEL: constant_sextload_v2i1_to_v2i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_i32 s3, s2, 0x10000 |
| ; GFX12-NEXT: s_bfe_i32 s2, s2, 0x10001 |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v1, s2 |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s3 |
| ; GFX12-NEXT: global_store_b64 v2, v[0:1], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <2 x i1>, ptr addrspace(4) %in |
| %ext = sext <2 x i1> %load to <2 x i32> |
| store <2 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v3i1_to_v3i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v3i1_to_v3i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v2, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_and_b32_e32 v0, 1, v2 |
| ; GFX6-NEXT: v_bfe_u32 v1, v2, 1, 1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v2, 2, v2 |
| ; GFX6-NEXT: buffer_store_dword v2, off, s[4:7], 0 offset:8 |
| ; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v3i1_to_v3i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v1, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, 2 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 |
| ; GFX8-NEXT: v_and_b32_e32 v0, 1, v1 |
| ; GFX8-NEXT: v_bfe_u32 v1, v1, 1, 1 |
| ; GFX8-NEXT: flat_store_dwordx3 v[3:4], v[0:2] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v3i1_to_v3i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 8, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T3.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T1.XY, T2.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_UINT * T1.Y, T0.X, 1, 1, |
| ; EG-NEXT: AND_INT T1.X, T0.X, 1, |
| ; EG-NEXT: LSHR * T2.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T0.X, T0.X, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 8(1.121039e-44) |
| ; EG-NEXT: LSHR * T3.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v3i1_to_v3i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v3, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v1, v3, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_and_b32_e32 v0, 0xffff, v1 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v2, 2, v0 |
| ; GFX12-NEXT: v_and_b32_e32 v0, 1, v1 |
| ; GFX12-NEXT: v_bfe_u32 v1, v1, 1, 1 |
| ; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX12-NEXT: global_store_b96 v3, v[0:2], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <3 x i1>, ptr addrspace(4) %in |
| %ext = zext <3 x i1> %load to <3 x i32> |
| store <3 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v3i1_to_v3i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v3i1_to_v3i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v2, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_i32 v1, v2, 1, 1 |
| ; GFX6-NEXT: v_bfe_i32 v0, v2, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v2, v2, 2, 1 |
| ; GFX6-NEXT: buffer_store_dword v2, off, s[4:7], 0 offset:8 |
| ; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v3i1_to_v3i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v0, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_bfe_i32 v2, v0, 2, 1 |
| ; GFX8-NEXT: v_bfe_i32 v1, v0, 1, 1 |
| ; GFX8-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX8-NEXT: flat_store_dwordx3 v[3:4], v[0:2] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v3i1_to_v3i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 10, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T2.X, T0.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T3.XY, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: LSHR T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT * T2.X, PV.W, 0.0, 1, |
| ; EG-NEXT: BFE_INT T3.X, T0.X, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T0.X, 1, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T0.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT * T3.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v3i1_to_v3i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_i32 s3, s2, 0x10002 |
| ; GFX12-NEXT: s_bfe_i32 s4, s2, 0x10000 |
| ; GFX12-NEXT: s_bfe_i32 s2, s2, 0x10001 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, 0 :: v_dual_mov_b32 v0, s4 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s2 :: v_dual_mov_b32 v2, s3 |
| ; GFX12-NEXT: global_store_b96 v3, v[0:2], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <3 x i1>, ptr addrspace(4) %in |
| %ext = sext <3 x i1> %load to <3 x i32> |
| store <3 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v4i1_to_v4i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v4i1_to_v4i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v1, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v3, 3, v1 |
| ; GFX6-NEXT: v_and_b32_e32 v0, 1, v1 |
| ; GFX6-NEXT: v_bfe_u32 v2, v1, 2, 1 |
| ; GFX6-NEXT: v_bfe_u32 v1, v1, 1, 1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v4i1_to_v4i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v1, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, 3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 |
| ; GFX8-NEXT: v_and_b32_e32 v0, 1, v1 |
| ; GFX8-NEXT: v_bfe_u32 v2, v1, 2, 1 |
| ; GFX8-NEXT: v_bfe_u32 v1, v1, 1, 1 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v4i1_to_v4i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 7, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_UINT * T0.W, T0.X, literal.x, 1, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT * T0.Z, T0.X, literal.x, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT * T0.Y, T0.X, 1, 1, |
| ; EG-NEXT: AND_INT T0.X, T0.X, 1, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v4i1_to_v4i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v4, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v1, v4, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v1 |
| ; GFX12-NEXT: v_and_b32_e32 v0, 1, v1 |
| ; GFX12-NEXT: v_bfe_u32 v2, v1, 2, 1 |
| ; GFX12-NEXT: v_bfe_u32 v1, v1, 1, 1 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v3, 3, v3 |
| ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <4 x i1>, ptr addrspace(4) %in |
| %ext = zext <4 x i1> %load to <4 x i32> |
| store <4 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v4i1_to_v4i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v4i1_to_v4i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_i32 v3, v0, 3, 1 |
| ; GFX6-NEXT: v_bfe_i32 v2, v0, 2, 1 |
| ; GFX6-NEXT: v_bfe_i32 v1, v0, 1, 1 |
| ; GFX6-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v4i1_to_v4i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v0, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_bfe_i32 v3, v0, 3, 1 |
| ; GFX8-NEXT: v_bfe_i32 v2, v0, 2, 1 |
| ; GFX8-NEXT: v_bfe_i32 v1, v0, 1, 1 |
| ; GFX8-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v4i1_to_v4i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 10, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T1.XYZW, T0.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_INT T1.X, T0.X, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.x, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T1.Z, PS, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T0.X, 1, |
| ; EG-NEXT: LSHR T0.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: BFE_INT * T1.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v4i1_to_v4i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_i32 s3, s2, 0x10003 |
| ; GFX12-NEXT: s_bfe_i32 s4, s2, 0x10002 |
| ; GFX12-NEXT: s_bfe_i32 s5, s2, 0x10000 |
| ; GFX12-NEXT: s_bfe_i32 s2, s2, 0x10001 |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v1, s2 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v3, s3 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <4 x i1>, ptr addrspace(4) %in |
| %ext = sext <4 x i1> %load to <4 x i32> |
| store <4 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v8i1_to_v8i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v8i1_to_v8i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v4, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_u32 v3, v4, 3, 1 |
| ; GFX6-NEXT: v_bfe_u32 v1, v4, 1, 1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v7, 7, v4 |
| ; GFX6-NEXT: v_bfe_u32 v5, v4, 5, 1 |
| ; GFX6-NEXT: v_and_b32_e32 v0, 1, v4 |
| ; GFX6-NEXT: v_bfe_u32 v2, v4, 2, 1 |
| ; GFX6-NEXT: v_bfe_u32 v6, v4, 6, 1 |
| ; GFX6-NEXT: v_bfe_u32 v4, v4, 4, 1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[4:7], off, s[4:7], 0 offset:16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v8i1_to_v8i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v0, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v9, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, s0 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_readfirstlane_b32 s2, v0 |
| ; GFX8-NEXT: s_bfe_u32 s3, s2, 0x10003 |
| ; GFX8-NEXT: s_bfe_u32 s4, s2, 0x10001 |
| ; GFX8-NEXT: s_bfe_u32 s5, s2, 0x10005 |
| ; GFX8-NEXT: s_and_b32 s6, s2, 1 |
| ; GFX8-NEXT: s_bfe_u32 s7, s2, 0x10002 |
| ; GFX8-NEXT: s_bfe_u32 s2, s2, 0x10004 |
| ; GFX8-NEXT: s_add_u32 s0, s0, 16 |
| ; GFX8-NEXT: s_addc_u32 s1, s1, 0 |
| ; GFX8-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX8-NEXT: v_mov_b32_e32 v11, s1 |
| ; GFX8-NEXT: v_lshrrev_b32_e32 v3, 7, v0 |
| ; GFX8-NEXT: v_bfe_u32 v2, v0, 6, 1 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s5 |
| ; GFX8-NEXT: v_mov_b32_e32 v10, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v6, s7 |
| ; GFX8-NEXT: v_mov_b32_e32 v7, s3 |
| ; GFX8-NEXT: flat_store_dwordx4 v[10:11], v[0:3] |
| ; GFX8-NEXT: flat_store_dwordx4 v[8:9], v[4:7] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v8i1_to_v8i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 17, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.XYZW, T8.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T6.XYZW, T7.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T5.X, T5.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T5.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_UINT * T6.W, T5.X, literal.x, 1, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT * T6.Z, T5.X, literal.x, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T6.Y, T5.X, 1, 1, |
| ; EG-NEXT: BFE_UINT * T5.W, T5.X, literal.x, 1, |
| ; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00) |
| ; EG-NEXT: AND_INT T6.X, T5.X, 1, |
| ; EG-NEXT: BFE_UINT T5.Z, T5.X, literal.x, 1, |
| ; EG-NEXT: LSHR * T7.X, KC0[2].Y, literal.y, |
| ; EG-NEXT: 6(8.407791e-45), 2(2.802597e-45) |
| ; EG-NEXT: BFE_UINT * T5.Y, T5.X, literal.x, 1, |
| ; EG-NEXT: 5(7.006492e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T5.X, T5.X, literal.x, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 4(5.605194e-45), 16(2.242078e-44) |
| ; EG-NEXT: LSHR * T8.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v8i1_to_v8i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v8, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v0, v8, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_readfirstlane_b32 s2, v0 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10001 |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v5, s4 :: v_dual_and_b32 v0, 0xffff, v0 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10003 |
| ; GFX12-NEXT: s_bfe_u32 s5, s2, 0x10005 |
| ; GFX12-NEXT: s_and_b32 s6, s2, 1 |
| ; GFX12-NEXT: s_bfe_u32 s7, s2, 0x10002 |
| ; GFX12-NEXT: s_bfe_u32 s2, s2, 0x10004 |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v3, 7, v0 |
| ; GFX12-NEXT: v_bfe_u32 v2, v0, 6, 1 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v7, s3 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s5 :: v_dual_mov_b32 v4, s6 |
| ; GFX12-NEXT: v_mov_b32_e32 v6, s7 |
| ; GFX12-NEXT: s_clause 0x1 |
| ; GFX12-NEXT: global_store_b128 v8, v[0:3], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v8, v[4:7], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <8 x i1>, ptr addrspace(4) %in |
| %ext = zext <8 x i1> %load to <8 x i32> |
| store <8 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v8i1_to_v8i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v8i1_to_v8i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v4, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_i32 v3, v4, 3, 1 |
| ; GFX6-NEXT: v_bfe_i32 v2, v4, 2, 1 |
| ; GFX6-NEXT: v_bfe_i32 v1, v4, 1, 1 |
| ; GFX6-NEXT: v_bfe_i32 v0, v4, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v7, v4, 7, 1 |
| ; GFX6-NEXT: v_bfe_i32 v6, v4, 6, 1 |
| ; GFX6-NEXT: v_bfe_i32 v5, v4, 5, 1 |
| ; GFX6-NEXT: v_bfe_i32 v4, v4, 4, 1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[4:7], off, s[4:7], 0 offset:16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v8i1_to_v8i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v4, v[0:1] |
| ; GFX8-NEXT: s_add_u32 s2, s0, 16 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v11, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v9, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v10, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, s0 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_bfe_i32 v3, v4, 3, 1 |
| ; GFX8-NEXT: v_bfe_i32 v2, v4, 2, 1 |
| ; GFX8-NEXT: v_bfe_i32 v1, v4, 1, 1 |
| ; GFX8-NEXT: v_bfe_i32 v0, v4, 0, 1 |
| ; GFX8-NEXT: v_bfe_i32 v7, v4, 7, 1 |
| ; GFX8-NEXT: v_bfe_i32 v6, v4, 6, 1 |
| ; GFX8-NEXT: v_bfe_i32 v5, v4, 5, 1 |
| ; GFX8-NEXT: v_bfe_i32 v4, v4, 4, 1 |
| ; GFX8-NEXT: flat_store_dwordx4 v[10:11], v[4:7] |
| ; GFX8-NEXT: flat_store_dwordx4 v[8:9], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v8i1_to_v8i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 23, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T6.XYZW, T5.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T7.XYZW, T8.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T5.X, T5.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T5.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: LSHR * T0.W, T5.X, literal.x, |
| ; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T6.W, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T5.X, literal.x, |
| ; EG-NEXT: 6(8.407791e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T7.X, T5.X, 0.0, 1, |
| ; EG-NEXT: BFE_INT T6.Z, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T5.X, literal.x, |
| ; EG-NEXT: LSHR * T1.W, T5.X, literal.y, |
| ; EG-NEXT: 3(4.203895e-45), 5(7.006492e-45) |
| ; EG-NEXT: LSHR T8.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: BFE_INT T6.Y, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T5.X, literal.x, |
| ; EG-NEXT: BFE_INT T7.W, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T5.X, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 4(5.605194e-45) |
| ; EG-NEXT: BFE_INT T6.X, PS, 0.0, 1, |
| ; EG-NEXT: BFE_INT T7.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T5.X, 1, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T5.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT * T7.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v8i1_to_v8i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_i32 s3, s2, 0x10003 |
| ; GFX12-NEXT: s_bfe_i32 s4, s2, 0x10002 |
| ; GFX12-NEXT: s_bfe_i32 s5, s2, 0x10001 |
| ; GFX12-NEXT: s_bfe_i32 s6, s2, 0x10000 |
| ; GFX12-NEXT: s_bfe_i32 s7, s2, 0x10007 |
| ; GFX12-NEXT: s_bfe_i32 s8, s2, 0x10006 |
| ; GFX12-NEXT: s_bfe_i32 s9, s2, 0x10004 |
| ; GFX12-NEXT: s_bfe_i32 s2, s2, 0x10005 |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v8, 0 :: v_dual_mov_b32 v1, s2 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s9 :: v_dual_mov_b32 v3, s7 |
| ; GFX12-NEXT: v_dual_mov_b32 v2, s8 :: v_dual_mov_b32 v5, s5 |
| ; GFX12-NEXT: v_dual_mov_b32 v4, s6 :: v_dual_mov_b32 v7, s3 |
| ; GFX12-NEXT: v_mov_b32_e32 v6, s4 |
| ; GFX12-NEXT: s_clause 0x1 |
| ; GFX12-NEXT: global_store_b128 v8, v[0:3], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v8, v[4:7], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <8 x i1>, ptr addrspace(4) %in |
| %ext = sext <8 x i1> %load to <8 x i32> |
| store <8 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v16i1_to_v16i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v16i1_to_v16i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s2 |
| ; GFX6-NEXT: s_mov_b32 s11, s3 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s6 |
| ; GFX6-NEXT: s_mov_b32 s9, s7 |
| ; GFX6-NEXT: buffer_load_ushort v12, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s0, s4 |
| ; GFX6-NEXT: s_mov_b32 s1, s5 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_u32 v3, v12, 3, 1 |
| ; GFX6-NEXT: v_bfe_u32 v1, v12, 1, 1 |
| ; GFX6-NEXT: v_bfe_u32 v7, v12, 7, 1 |
| ; GFX6-NEXT: v_bfe_u32 v5, v12, 5, 1 |
| ; GFX6-NEXT: v_bfe_u32 v11, v12, 11, 1 |
| ; GFX6-NEXT: v_bfe_u32 v9, v12, 9, 1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v15, 15, v12 |
| ; GFX6-NEXT: v_bfe_u32 v13, v12, 13, 1 |
| ; GFX6-NEXT: v_and_b32_e32 v0, 1, v12 |
| ; GFX6-NEXT: v_bfe_u32 v2, v12, 2, 1 |
| ; GFX6-NEXT: v_bfe_u32 v6, v12, 6, 1 |
| ; GFX6-NEXT: v_bfe_u32 v4, v12, 4, 1 |
| ; GFX6-NEXT: v_bfe_u32 v10, v12, 10, 1 |
| ; GFX6-NEXT: v_bfe_u32 v8, v12, 8, 1 |
| ; GFX6-NEXT: v_bfe_u32 v14, v12, 14, 1 |
| ; GFX6-NEXT: v_bfe_u32 v12, v12, 12, 1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v16i1_to_v16i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ushort v0, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v17, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v16, s0 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_readfirstlane_b32 s2, v0 |
| ; GFX8-NEXT: s_and_b32 s6, 0xffff, s2 |
| ; GFX8-NEXT: s_bfe_u32 s3, s2, 0x10003 |
| ; GFX8-NEXT: s_bfe_u32 s4, s2, 0x10001 |
| ; GFX8-NEXT: s_bfe_u32 s5, s2, 0x10007 |
| ; GFX8-NEXT: s_bfe_u32 s7, s2, 0x10009 |
| ; GFX8-NEXT: s_bfe_u32 s8, s2, 0x1000d |
| ; GFX8-NEXT: s_and_b32 s9, s2, 1 |
| ; GFX8-NEXT: s_bfe_u32 s10, s2, 0x1000a |
| ; GFX8-NEXT: s_bfe_u32 s2, s2, 0x1000c |
| ; GFX8-NEXT: s_bfe_u32 s11, s6, 0x10005 |
| ; GFX8-NEXT: s_bfe_u32 s12, s6, 0x1000b |
| ; GFX8-NEXT: s_lshr_b32 s13, s6, 15 |
| ; GFX8-NEXT: s_bfe_u32 s14, s6, 0x10002 |
| ; GFX8-NEXT: s_bfe_u32 s15, s6, 0x10006 |
| ; GFX8-NEXT: s_bfe_u32 s16, s6, 0x10004 |
| ; GFX8-NEXT: s_bfe_u32 s17, s6, 0x10008 |
| ; GFX8-NEXT: s_bfe_u32 s6, s6, 0x1000e |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 48 |
| ; GFX8-NEXT: v_mov_b32_e32 v15, s3 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v19, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s8 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s13 |
| ; GFX8-NEXT: v_mov_b32_e32 v18, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 32 |
| ; GFX8-NEXT: flat_store_dwordx4 v[18:19], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: v_mov_b32_e32 v6, s10 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s17 |
| ; GFX8-NEXT: v_mov_b32_e32 v7, s12 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: s_add_u32 s0, s0, 16 |
| ; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[4:7] |
| ; GFX8-NEXT: s_addc_u32 s1, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v11, s5 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, s16 |
| ; GFX8-NEXT: v_mov_b32_e32 v9, s11 |
| ; GFX8-NEXT: v_mov_b32_e32 v10, s15 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v12, s9 |
| ; GFX8-NEXT: v_mov_b32_e32 v13, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v14, s14 |
| ; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[8:11] |
| ; GFX8-NEXT: flat_store_dwordx4 v[16:17], v[12:15] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v16i1_to_v16i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @10, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @8 |
| ; EG-NEXT: ALU 36, @11, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T7.XYZW, T14.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T11.XYZW, T13.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T9.XYZW, T12.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T8.XYZW, T10.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 8: |
| ; EG-NEXT: VTX_READ_16 T7.X, T7.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 10: |
| ; EG-NEXT: MOV * T7.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 11: |
| ; EG-NEXT: BFE_UINT * T8.W, T7.X, literal.x, 1, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT * T8.Z, T7.X, literal.x, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T8.Y, T7.X, 1, 1, |
| ; EG-NEXT: BFE_UINT * T9.W, T7.X, literal.x, 1, |
| ; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00) |
| ; EG-NEXT: AND_INT T8.X, T7.X, 1, |
| ; EG-NEXT: BFE_UINT T9.Z, T7.X, literal.x, 1, |
| ; EG-NEXT: LSHR * T10.X, KC0[2].Y, literal.y, |
| ; EG-NEXT: 6(8.407791e-45), 2(2.802597e-45) |
| ; EG-NEXT: BFE_UINT T9.Y, T7.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT * T11.W, T7.X, literal.y, 1, |
| ; EG-NEXT: 5(7.006492e-45), 11(1.541428e-44) |
| ; EG-NEXT: BFE_UINT T9.X, T7.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T11.Z, T7.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 4(5.605194e-45), 10(1.401298e-44) |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T12.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T11.Y, T7.X, literal.y, 1, |
| ; EG-NEXT: LSHR * T7.W, T7.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 9(1.261169e-44) |
| ; EG-NEXT: 15(2.101948e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T11.X, T7.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T7.Z, T7.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 8(1.121039e-44), 14(1.961818e-44) |
| ; EG-NEXT: 32(4.484155e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T13.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT * T7.Y, T7.X, literal.y, 1, |
| ; EG-NEXT: 2(2.802597e-45), 13(1.821688e-44) |
| ; EG-NEXT: BFE_UINT T7.X, T7.X, literal.x, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 12(1.681558e-44), 48(6.726233e-44) |
| ; EG-NEXT: LSHR * T14.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v16i1_to_v16i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v16, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u16 v0, v16, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_readfirstlane_b32 s2, v0 |
| ; GFX12-NEXT: s_and_b32 s6, 0xffff, s2 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10003 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10001 |
| ; GFX12-NEXT: s_bfe_u32 s5, s2, 0x10007 |
| ; GFX12-NEXT: s_bfe_u32 s7, s2, 0x10009 |
| ; GFX12-NEXT: s_bfe_u32 s8, s2, 0x1000d |
| ; GFX12-NEXT: s_and_b32 s9, s2, 1 |
| ; GFX12-NEXT: v_mov_b32_e32 v1, s8 |
| ; GFX12-NEXT: s_bfe_u32 s10, s2, 0x1000a |
| ; GFX12-NEXT: s_bfe_u32 s2, s2, 0x1000c |
| ; GFX12-NEXT: s_bfe_u32 s11, s6, 0x10005 |
| ; GFX12-NEXT: s_bfe_u32 s12, s6, 0x1000b |
| ; GFX12-NEXT: s_lshr_b32 s13, s6, 15 |
| ; GFX12-NEXT: s_bfe_u32 s14, s6, 0x10002 |
| ; GFX12-NEXT: s_bfe_u32 s15, s6, 0x10006 |
| ; GFX12-NEXT: s_bfe_u32 s16, s6, 0x10004 |
| ; GFX12-NEXT: s_bfe_u32 s17, s6, 0x10008 |
| ; GFX12-NEXT: s_bfe_u32 s6, s6, 0x1000e |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v5, s7 |
| ; GFX12-NEXT: v_dual_mov_b32 v15, s3 :: v_dual_mov_b32 v2, s6 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s13 :: v_dual_mov_b32 v4, s17 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s10 :: v_dual_mov_b32 v11, s5 |
| ; GFX12-NEXT: v_dual_mov_b32 v7, s12 :: v_dual_mov_b32 v8, s16 |
| ; GFX12-NEXT: v_dual_mov_b32 v9, s11 :: v_dual_mov_b32 v10, s15 |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v14, s14 |
| ; GFX12-NEXT: s_clause 0x3 |
| ; GFX12-NEXT: global_store_b128 v16, v[0:3], s[0:1] offset:48 |
| ; GFX12-NEXT: global_store_b128 v16, v[4:7], s[0:1] offset:32 |
| ; GFX12-NEXT: global_store_b128 v16, v[8:11], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v16, v[12:15], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <16 x i1>, ptr addrspace(4) %in |
| %ext = zext <16 x i1> %load to <16 x i32> |
| store <16 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v16i1_to_v16i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v16i1_to_v16i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s2 |
| ; GFX6-NEXT: s_mov_b32 s11, s3 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s6 |
| ; GFX6-NEXT: s_mov_b32 s9, s7 |
| ; GFX6-NEXT: buffer_load_ushort v12, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s0, s4 |
| ; GFX6-NEXT: s_mov_b32 s1, s5 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_i32 v3, v12, 3, 1 |
| ; GFX6-NEXT: v_bfe_i32 v2, v12, 2, 1 |
| ; GFX6-NEXT: v_bfe_i32 v1, v12, 1, 1 |
| ; GFX6-NEXT: v_bfe_i32 v0, v12, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v7, v12, 7, 1 |
| ; GFX6-NEXT: v_bfe_i32 v6, v12, 6, 1 |
| ; GFX6-NEXT: v_bfe_i32 v5, v12, 5, 1 |
| ; GFX6-NEXT: v_bfe_i32 v4, v12, 4, 1 |
| ; GFX6-NEXT: v_bfe_i32 v11, v12, 11, 1 |
| ; GFX6-NEXT: v_bfe_i32 v10, v12, 10, 1 |
| ; GFX6-NEXT: v_bfe_i32 v9, v12, 9, 1 |
| ; GFX6-NEXT: v_bfe_i32 v8, v12, 8, 1 |
| ; GFX6-NEXT: v_bfe_i32 v15, v12, 15, 1 |
| ; GFX6-NEXT: v_bfe_i32 v14, v12, 14, 1 |
| ; GFX6-NEXT: v_bfe_i32 v13, v12, 13, 1 |
| ; GFX6-NEXT: v_bfe_i32 v12, v12, 12, 1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v16i1_to_v16i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ushort v18, v[0:1] |
| ; GFX8-NEXT: s_add_u32 s2, s0, 48 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v9, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 32 |
| ; GFX8-NEXT: v_mov_b32_e32 v13, s1 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v12, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s0, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v15, s3 |
| ; GFX8-NEXT: s_addc_u32 s1, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v14, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v17, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v16, s0 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_bfe_i32 v7, v18, 15, 1 |
| ; GFX8-NEXT: v_bfe_i32 v6, v18, 14, 1 |
| ; GFX8-NEXT: v_bfe_i32 v5, v18, 13, 1 |
| ; GFX8-NEXT: v_bfe_i32 v4, v18, 12, 1 |
| ; GFX8-NEXT: flat_store_dwordx4 v[8:9], v[4:7] |
| ; GFX8-NEXT: v_bfe_i32 v11, v18, 11, 1 |
| ; GFX8-NEXT: v_bfe_i32 v10, v18, 10, 1 |
| ; GFX8-NEXT: v_bfe_i32 v9, v18, 9, 1 |
| ; GFX8-NEXT: v_bfe_i32 v8, v18, 8, 1 |
| ; GFX8-NEXT: v_bfe_i32 v3, v18, 3, 1 |
| ; GFX8-NEXT: v_bfe_i32 v2, v18, 2, 1 |
| ; GFX8-NEXT: v_bfe_i32 v1, v18, 1, 1 |
| ; GFX8-NEXT: v_bfe_i32 v0, v18, 0, 1 |
| ; GFX8-NEXT: v_bfe_i32 v7, v18, 7, 1 |
| ; GFX8-NEXT: v_bfe_i32 v6, v18, 6, 1 |
| ; GFX8-NEXT: v_bfe_i32 v5, v18, 5, 1 |
| ; GFX8-NEXT: v_bfe_i32 v4, v18, 4, 1 |
| ; GFX8-NEXT: flat_store_dwordx4 v[14:15], v[8:11] |
| ; GFX8-NEXT: flat_store_dwordx4 v[16:17], v[4:7] |
| ; GFX8-NEXT: flat_store_dwordx4 v[12:13], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v16i1_to_v16i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @10, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @8 |
| ; EG-NEXT: ALU 51, @11, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T12.XYZW, T7.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T10.XYZW, T14.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T8.XYZW, T11.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T13.XYZW, T9.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 8: |
| ; EG-NEXT: VTX_READ_16 T7.X, T7.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 10: |
| ; EG-NEXT: MOV * T7.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 11: |
| ; EG-NEXT: LSHR * T0.W, T7.X, literal.x, |
| ; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T8.W, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T7.X, literal.x, |
| ; EG-NEXT: 6(8.407791e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T8.Z, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T7.X, literal.x, |
| ; EG-NEXT: LSHR * T1.W, T7.X, literal.y, |
| ; EG-NEXT: 11(1.541428e-44), 5(7.006492e-45) |
| ; EG-NEXT: LSHR T9.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: BFE_INT T8.Y, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T7.X, literal.y, |
| ; EG-NEXT: BFE_INT T10.W, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T7.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 10(1.401298e-44) |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T8.X, PS, 0.0, 1, |
| ; EG-NEXT: BFE_INT T10.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T7.X, literal.x, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 9(1.261169e-44), 16(2.242078e-44) |
| ; EG-NEXT: LSHR T11.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T10.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T7.X, literal.y, |
| ; EG-NEXT: LSHR * T1.W, T7.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 15(2.101948e-44) |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T10.X, PS, 0.0, 1, |
| ; EG-NEXT: BFE_INT T12.W, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T7.X, literal.x, |
| ; EG-NEXT: 14(1.961818e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T13.X, T7.X, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T7.X, literal.x, |
| ; EG-NEXT: BFE_INT T12.Z, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T7.X, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 3(4.203895e-45), 13(1.821688e-44) |
| ; EG-NEXT: 32(4.484155e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T14.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T12.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T7.X, literal.x, |
| ; EG-NEXT: BFE_INT T13.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T7.X, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 12(1.681558e-44) |
| ; EG-NEXT: BFE_INT T12.X, PS, 0.0, 1, |
| ; EG-NEXT: BFE_INT T13.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T7.X, 1, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 48(6.726233e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T7.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT * T13.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v16i1_to_v16i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u16 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_i32 s3, s2, 0x10003 |
| ; GFX12-NEXT: s_bfe_i32 s4, s2, 0x10002 |
| ; GFX12-NEXT: s_bfe_i32 s5, s2, 0x10001 |
| ; GFX12-NEXT: s_bfe_i32 s6, s2, 0x10000 |
| ; GFX12-NEXT: s_bfe_i32 s7, s2, 0x10007 |
| ; GFX12-NEXT: s_bfe_i32 s8, s2, 0x10006 |
| ; GFX12-NEXT: s_bfe_i32 s9, s2, 0x10005 |
| ; GFX12-NEXT: s_bfe_i32 s10, s2, 0x10004 |
| ; GFX12-NEXT: s_bfe_i32 s11, s2, 0x1000b |
| ; GFX12-NEXT: s_bfe_i32 s12, s2, 0x1000a |
| ; GFX12-NEXT: s_bfe_i32 s13, s2, 0x10009 |
| ; GFX12-NEXT: s_bfe_i32 s14, s2, 0x10008 |
| ; GFX12-NEXT: s_bfe_i32 s15, s2, 0x1000f |
| ; GFX12-NEXT: s_bfe_i32 s16, s2, 0x1000e |
| ; GFX12-NEXT: s_bfe_i32 s17, s2, 0x1000c |
| ; GFX12-NEXT: s_bfe_i32 s2, s2, 0x1000d |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v16, 0 :: v_dual_mov_b32 v1, s2 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s17 :: v_dual_mov_b32 v3, s15 |
| ; GFX12-NEXT: v_dual_mov_b32 v2, s16 :: v_dual_mov_b32 v5, s13 |
| ; GFX12-NEXT: v_dual_mov_b32 v4, s14 :: v_dual_mov_b32 v7, s11 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s12 :: v_dual_mov_b32 v9, s9 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s10 :: v_dual_mov_b32 v11, s7 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s8 :: v_dual_mov_b32 v13, s5 |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s6 :: v_dual_mov_b32 v15, s3 |
| ; GFX12-NEXT: v_mov_b32_e32 v14, s4 |
| ; GFX12-NEXT: s_clause 0x3 |
| ; GFX12-NEXT: global_store_b128 v16, v[0:3], s[0:1] offset:48 |
| ; GFX12-NEXT: global_store_b128 v16, v[4:7], s[0:1] offset:32 |
| ; GFX12-NEXT: global_store_b128 v16, v[8:11], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v16, v[12:15], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <16 x i1>, ptr addrspace(4) %in |
| %ext = sext <16 x i1> %load to <16 x i32> |
| store <16 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v32i1_to_v32i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v32i1_to_v32i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_load_dword s4, s[2:3], 0x0 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_bfe_u32 s5, s4, 0x10003 |
| ; GFX6-NEXT: s_bfe_u32 s6, s4, 0x10001 |
| ; GFX6-NEXT: s_bfe_u32 s7, s4, 0x10007 |
| ; GFX6-NEXT: s_bfe_u32 s8, s4, 0x10005 |
| ; GFX6-NEXT: s_bfe_u32 s9, s4, 0x1000b |
| ; GFX6-NEXT: s_bfe_u32 s10, s4, 0x10009 |
| ; GFX6-NEXT: s_bfe_u32 s11, s4, 0x1000f |
| ; GFX6-NEXT: s_bfe_u32 s12, s4, 0x1000d |
| ; GFX6-NEXT: s_bfe_u32 s13, s4, 0x10013 |
| ; GFX6-NEXT: s_bfe_u32 s14, s4, 0x10011 |
| ; GFX6-NEXT: s_bfe_u32 s15, s4, 0x10017 |
| ; GFX6-NEXT: s_bfe_u32 s16, s4, 0x10015 |
| ; GFX6-NEXT: s_bfe_u32 s17, s4, 0x1001b |
| ; GFX6-NEXT: s_bfe_u32 s18, s4, 0x10019 |
| ; GFX6-NEXT: s_lshr_b32 s19, s4, 31 |
| ; GFX6-NEXT: s_bfe_u32 s20, s4, 0x1001d |
| ; GFX6-NEXT: s_and_b32 s21, s4, 1 |
| ; GFX6-NEXT: s_bfe_u32 s22, s4, 0x10002 |
| ; GFX6-NEXT: s_bfe_u32 s23, s4, 0x10006 |
| ; GFX6-NEXT: s_bfe_u32 s24, s4, 0x10004 |
| ; GFX6-NEXT: s_bfe_u32 s25, s4, 0x1000a |
| ; GFX6-NEXT: s_bfe_u32 s26, s4, 0x10008 |
| ; GFX6-NEXT: s_bfe_u32 s27, s4, 0x1000e |
| ; GFX6-NEXT: s_bfe_u32 s28, s4, 0x1000c |
| ; GFX6-NEXT: s_bfe_u32 s29, s4, 0x10012 |
| ; GFX6-NEXT: s_bfe_u32 s30, s4, 0x10010 |
| ; GFX6-NEXT: s_bfe_u32 s31, s4, 0x10016 |
| ; GFX6-NEXT: s_bfe_u32 s33, s4, 0x10014 |
| ; GFX6-NEXT: s_bfe_u32 s34, s4, 0x1001a |
| ; GFX6-NEXT: s_bfe_u32 s35, s4, 0x1001e |
| ; GFX6-NEXT: s_bfe_u32 s36, s4, 0x1001c |
| ; GFX6-NEXT: s_bfe_u32 s4, s4, 0x10018 |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s20 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s35 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:112 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s18 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s34 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s17 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:96 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s33 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s16 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s31 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s15 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s30 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s14 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s29 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s13 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s28 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s12 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s27 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s11 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s26 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s10 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s25 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s9 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s24 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s8 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s23 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s7 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s21 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s6 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s22 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s5 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v32i1_to_v32i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_load_dword s2, s[2:3], 0x0 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_bfe_u32 s4, s2, 0x10003 |
| ; GFX8-NEXT: s_bfe_u32 s5, s2, 0x10001 |
| ; GFX8-NEXT: s_bfe_u32 s6, s2, 0x10007 |
| ; GFX8-NEXT: s_bfe_u32 s7, s2, 0x10005 |
| ; GFX8-NEXT: s_bfe_u32 s8, s2, 0x1000b |
| ; GFX8-NEXT: s_bfe_u32 s9, s2, 0x10009 |
| ; GFX8-NEXT: s_bfe_u32 s10, s2, 0x1000f |
| ; GFX8-NEXT: s_bfe_u32 s11, s2, 0x1000d |
| ; GFX8-NEXT: s_bfe_u32 s12, s2, 0x10013 |
| ; GFX8-NEXT: s_bfe_u32 s13, s2, 0x10011 |
| ; GFX8-NEXT: s_bfe_u32 s14, s2, 0x10017 |
| ; GFX8-NEXT: s_bfe_u32 s15, s2, 0x1001b |
| ; GFX8-NEXT: s_bfe_u32 s16, s2, 0x10019 |
| ; GFX8-NEXT: s_lshr_b32 s3, s2, 31 |
| ; GFX8-NEXT: s_bfe_u32 s17, s2, 0x1001d |
| ; GFX8-NEXT: s_and_b32 s18, s2, 1 |
| ; GFX8-NEXT: s_bfe_u32 s19, s2, 0x10002 |
| ; GFX8-NEXT: s_bfe_u32 s20, s2, 0x10006 |
| ; GFX8-NEXT: s_bfe_u32 s21, s2, 0x10004 |
| ; GFX8-NEXT: s_bfe_u32 s22, s2, 0x1000a |
| ; GFX8-NEXT: s_bfe_u32 s23, s2, 0x10008 |
| ; GFX8-NEXT: s_bfe_u32 s24, s2, 0x1000e |
| ; GFX8-NEXT: s_bfe_u32 s25, s2, 0x1000c |
| ; GFX8-NEXT: s_bfe_u32 s26, s2, 0x10012 |
| ; GFX8-NEXT: s_bfe_u32 s27, s2, 0x10010 |
| ; GFX8-NEXT: s_bfe_u32 s28, s2, 0x10016 |
| ; GFX8-NEXT: s_bfe_u32 s29, s2, 0x10015 |
| ; GFX8-NEXT: s_bfe_u32 s30, s2, 0x10014 |
| ; GFX8-NEXT: s_bfe_u32 s31, s2, 0x1001a |
| ; GFX8-NEXT: s_bfe_u32 s33, s2, 0x10018 |
| ; GFX8-NEXT: s_bfe_u32 s34, s2, 0x1001e |
| ; GFX8-NEXT: s_bfe_u32 s2, s2, 0x1001c |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 0x70 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s3 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 0x60 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s34 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 0x50 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s33 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s16 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s31 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s15 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 64 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s30 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s29 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s28 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s14 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 48 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s27 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s13 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s12 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 32 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s25 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s11 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s24 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s10 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s23 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s9 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s22 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s8 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s21 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s7 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s20 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s18 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s5 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s19 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v32i1_to_v32i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @14, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @12 |
| ; EG-NEXT: ALU 76, @15, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T11.XYZW, T26.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T23.XYZW, T25.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T21.XYZW, T24.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T22.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T17.XYZW, T20.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T15.XYZW, T18.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T13.XYZW, T16.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T12.XYZW, T14.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 12: |
| ; EG-NEXT: VTX_READ_32 T11.X, T11.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 14: |
| ; EG-NEXT: MOV * T11.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 15: |
| ; EG-NEXT: BFE_UINT * T12.W, T11.X, literal.x, 1, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT * T12.Z, T11.X, literal.x, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T12.Y, T11.X, 1, 1, |
| ; EG-NEXT: BFE_UINT * T13.W, T11.X, literal.x, 1, |
| ; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00) |
| ; EG-NEXT: AND_INT T12.X, T11.X, 1, |
| ; EG-NEXT: BFE_UINT T13.Z, T11.X, literal.x, 1, |
| ; EG-NEXT: LSHR * T14.X, KC0[2].Y, literal.y, |
| ; EG-NEXT: 6(8.407791e-45), 2(2.802597e-45) |
| ; EG-NEXT: BFE_UINT T13.Y, T11.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT * T15.W, T11.X, literal.y, 1, |
| ; EG-NEXT: 5(7.006492e-45), 11(1.541428e-44) |
| ; EG-NEXT: BFE_UINT T13.X, T11.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T15.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 4(5.605194e-45), 10(1.401298e-44) |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T16.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T15.Y, T11.X, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T17.W, T11.X, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 9(1.261169e-44) |
| ; EG-NEXT: 15(2.101948e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T15.X, T11.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T17.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 8(1.121039e-44), 14(1.961818e-44) |
| ; EG-NEXT: 32(4.484155e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T18.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T17.Y, T11.X, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T19.W, T11.X, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 13(1.821688e-44) |
| ; EG-NEXT: 19(2.662467e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T17.X, T11.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T19.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 12(1.681558e-44), 18(2.522337e-44) |
| ; EG-NEXT: 48(6.726233e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T20.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T19.Y, T11.X, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T21.W, T11.X, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 17(2.382207e-44) |
| ; EG-NEXT: 23(3.222986e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T19.X, T11.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T21.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 16(2.242078e-44), 22(3.082857e-44) |
| ; EG-NEXT: 64(8.968310e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T22.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T21.Y, T11.X, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T23.W, T11.X, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 21(2.942727e-44) |
| ; EG-NEXT: 27(3.783506e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T21.X, T11.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T23.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 20(2.802597e-44), 26(3.643376e-44) |
| ; EG-NEXT: 80(1.121039e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T24.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T23.Y, T11.X, literal.y, 1, |
| ; EG-NEXT: LSHR * T11.W, T11.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 25(3.503246e-44) |
| ; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T23.X, T11.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T11.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 24(3.363116e-44), 30(4.203895e-44) |
| ; EG-NEXT: 96(1.345247e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T25.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT * T11.Y, T11.X, literal.y, 1, |
| ; EG-NEXT: 2(2.802597e-45), 29(4.063766e-44) |
| ; EG-NEXT: BFE_UINT T11.X, T11.X, literal.x, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 28(3.923636e-44), 112(1.569454e-43) |
| ; EG-NEXT: LSHR * T26.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v32i1_to_v32i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_b32 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10003 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10001 |
| ; GFX12-NEXT: s_bfe_u32 s5, s2, 0x10007 |
| ; GFX12-NEXT: s_bfe_u32 s6, s2, 0x10005 |
| ; GFX12-NEXT: s_bfe_u32 s7, s2, 0x1000b |
| ; GFX12-NEXT: s_bfe_u32 s8, s2, 0x10009 |
| ; GFX12-NEXT: s_bfe_u32 s9, s2, 0x1000f |
| ; GFX12-NEXT: s_bfe_u32 s10, s2, 0x1000d |
| ; GFX12-NEXT: s_bfe_u32 s11, s2, 0x10013 |
| ; GFX12-NEXT: s_bfe_u32 s12, s2, 0x10011 |
| ; GFX12-NEXT: s_bfe_u32 s13, s2, 0x10017 |
| ; GFX12-NEXT: s_bfe_u32 s14, s2, 0x1001b |
| ; GFX12-NEXT: s_bfe_u32 s15, s2, 0x10019 |
| ; GFX12-NEXT: s_lshr_b32 s16, s2, 31 |
| ; GFX12-NEXT: s_bfe_u32 s17, s2, 0x1001d |
| ; GFX12-NEXT: s_and_b32 s18, s2, 1 |
| ; GFX12-NEXT: s_bfe_u32 s19, s2, 0x10002 |
| ; GFX12-NEXT: s_bfe_u32 s20, s2, 0x10006 |
| ; GFX12-NEXT: s_bfe_u32 s21, s2, 0x10004 |
| ; GFX12-NEXT: s_bfe_u32 s22, s2, 0x1000a |
| ; GFX12-NEXT: s_bfe_u32 s23, s2, 0x10008 |
| ; GFX12-NEXT: s_bfe_u32 s24, s2, 0x1000e |
| ; GFX12-NEXT: s_bfe_u32 s25, s2, 0x1000c |
| ; GFX12-NEXT: s_bfe_u32 s26, s2, 0x10012 |
| ; GFX12-NEXT: s_bfe_u32 s27, s2, 0x10010 |
| ; GFX12-NEXT: s_bfe_u32 s28, s2, 0x10016 |
| ; GFX12-NEXT: s_bfe_u32 s29, s2, 0x10015 |
| ; GFX12-NEXT: s_bfe_u32 s30, s2, 0x10014 |
| ; GFX12-NEXT: s_bfe_u32 s31, s2, 0x1001a |
| ; GFX12-NEXT: s_bfe_u32 s33, s2, 0x10018 |
| ; GFX12-NEXT: s_bfe_u32 s34, s2, 0x1001c |
| ; GFX12-NEXT: s_bfe_u32 s2, s2, 0x1001e |
| ; GFX12-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s17 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s34 :: v_dual_mov_b32 v3, s16 |
| ; GFX12-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s15 |
| ; GFX12-NEXT: v_dual_mov_b32 v4, s33 :: v_dual_mov_b32 v7, s14 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s31 :: v_dual_mov_b32 v9, s29 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s30 :: v_dual_mov_b32 v11, s13 |
| ; GFX12-NEXT: v_mov_b32_e32 v10, s28 |
| ; GFX12-NEXT: s_clause 0x1 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:112 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:96 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s12 :: v_dual_mov_b32 v0, s27 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s11 :: v_dual_mov_b32 v2, s26 |
| ; GFX12-NEXT: v_dual_mov_b32 v5, s10 :: v_dual_mov_b32 v4, s25 |
| ; GFX12-NEXT: v_dual_mov_b32 v7, s9 :: v_dual_mov_b32 v6, s24 |
| ; GFX12-NEXT: v_dual_mov_b32 v13, s8 :: v_dual_mov_b32 v12, s23 |
| ; GFX12-NEXT: v_dual_mov_b32 v15, s7 :: v_dual_mov_b32 v14, s22 |
| ; GFX12-NEXT: v_dual_mov_b32 v17, s6 :: v_dual_mov_b32 v16, s21 |
| ; GFX12-NEXT: v_dual_mov_b32 v19, s5 :: v_dual_mov_b32 v18, s20 |
| ; GFX12-NEXT: v_dual_mov_b32 v21, s4 :: v_dual_mov_b32 v20, s18 |
| ; GFX12-NEXT: v_dual_mov_b32 v23, s3 :: v_dual_mov_b32 v22, s19 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v24, v[8:11], s[0:1] offset:80 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:64 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:48 |
| ; GFX12-NEXT: global_store_b128 v24, v[12:15], s[0:1] offset:32 |
| ; GFX12-NEXT: global_store_b128 v24, v[16:19], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v24, v[20:23], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <32 x i1>, ptr addrspace(4) %in |
| %ext = zext <32 x i1> %load to <32 x i32> |
| store <32 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v32i1_to_v32i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v32i1_to_v32i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_load_dword s4, s[2:3], 0x0 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_bfe_i32 s5, s4, 0x10003 |
| ; GFX6-NEXT: s_bfe_i32 s6, s4, 0x10002 |
| ; GFX6-NEXT: s_bfe_i32 s7, s4, 0x10001 |
| ; GFX6-NEXT: s_bfe_i32 s8, s4, 0x10000 |
| ; GFX6-NEXT: s_bfe_i32 s9, s4, 0x10007 |
| ; GFX6-NEXT: s_bfe_i32 s10, s4, 0x10006 |
| ; GFX6-NEXT: s_bfe_i32 s11, s4, 0x10005 |
| ; GFX6-NEXT: s_bfe_i32 s12, s4, 0x10004 |
| ; GFX6-NEXT: s_bfe_i32 s13, s4, 0x1000b |
| ; GFX6-NEXT: s_bfe_i32 s14, s4, 0x1000a |
| ; GFX6-NEXT: s_bfe_i32 s15, s4, 0x10009 |
| ; GFX6-NEXT: s_bfe_i32 s16, s4, 0x10008 |
| ; GFX6-NEXT: s_bfe_i32 s17, s4, 0x1000f |
| ; GFX6-NEXT: s_bfe_i32 s18, s4, 0x1000e |
| ; GFX6-NEXT: s_bfe_i32 s19, s4, 0x1000d |
| ; GFX6-NEXT: s_bfe_i32 s20, s4, 0x1000c |
| ; GFX6-NEXT: s_bfe_i32 s21, s4, 0x10013 |
| ; GFX6-NEXT: s_bfe_i32 s22, s4, 0x10012 |
| ; GFX6-NEXT: s_bfe_i32 s23, s4, 0x10011 |
| ; GFX6-NEXT: s_bfe_i32 s24, s4, 0x10010 |
| ; GFX6-NEXT: s_bfe_i32 s25, s4, 0x10017 |
| ; GFX6-NEXT: s_bfe_i32 s26, s4, 0x10016 |
| ; GFX6-NEXT: s_bfe_i32 s27, s4, 0x10015 |
| ; GFX6-NEXT: s_bfe_i32 s28, s4, 0x10014 |
| ; GFX6-NEXT: s_bfe_i32 s29, s4, 0x1001b |
| ; GFX6-NEXT: s_bfe_i32 s30, s4, 0x1001a |
| ; GFX6-NEXT: s_bfe_i32 s31, s4, 0x10019 |
| ; GFX6-NEXT: s_ashr_i32 s33, s4, 31 |
| ; GFX6-NEXT: s_bfe_i32 s34, s4, 0x1001e |
| ; GFX6-NEXT: s_bfe_i32 s35, s4, 0x1001d |
| ; GFX6-NEXT: s_bfe_i32 s36, s4, 0x1001c |
| ; GFX6-NEXT: s_bfe_i32 s4, s4, 0x10018 |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s35 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s34 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s33 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:112 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s31 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s30 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s29 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:96 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s28 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s27 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s26 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s25 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s24 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s23 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s22 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s21 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s20 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s19 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s17 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s15 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s14 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s13 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s12 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s11 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s10 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s9 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s8 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s7 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s6 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s5 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v32i1_to_v32i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_load_dword s2, s[2:3], 0x0 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_bfe_i32 s4, s2, 0x10003 |
| ; GFX8-NEXT: s_bfe_i32 s5, s2, 0x10002 |
| ; GFX8-NEXT: s_bfe_i32 s6, s2, 0x10001 |
| ; GFX8-NEXT: s_bfe_i32 s7, s2, 0x10000 |
| ; GFX8-NEXT: s_bfe_i32 s8, s2, 0x10007 |
| ; GFX8-NEXT: s_bfe_i32 s9, s2, 0x10006 |
| ; GFX8-NEXT: s_bfe_i32 s10, s2, 0x10005 |
| ; GFX8-NEXT: s_bfe_i32 s11, s2, 0x10004 |
| ; GFX8-NEXT: s_bfe_i32 s12, s2, 0x1000b |
| ; GFX8-NEXT: s_bfe_i32 s13, s2, 0x1000a |
| ; GFX8-NEXT: s_bfe_i32 s14, s2, 0x10009 |
| ; GFX8-NEXT: s_bfe_i32 s15, s2, 0x10008 |
| ; GFX8-NEXT: s_bfe_i32 s16, s2, 0x1000f |
| ; GFX8-NEXT: s_bfe_i32 s17, s2, 0x1000e |
| ; GFX8-NEXT: s_bfe_i32 s18, s2, 0x1000d |
| ; GFX8-NEXT: s_bfe_i32 s19, s2, 0x1000c |
| ; GFX8-NEXT: s_bfe_i32 s20, s2, 0x10013 |
| ; GFX8-NEXT: s_bfe_i32 s21, s2, 0x10012 |
| ; GFX8-NEXT: s_bfe_i32 s22, s2, 0x10011 |
| ; GFX8-NEXT: s_bfe_i32 s23, s2, 0x10010 |
| ; GFX8-NEXT: s_bfe_i32 s24, s2, 0x10017 |
| ; GFX8-NEXT: s_bfe_i32 s25, s2, 0x10016 |
| ; GFX8-NEXT: s_bfe_i32 s26, s2, 0x10015 |
| ; GFX8-NEXT: s_bfe_i32 s27, s2, 0x10014 |
| ; GFX8-NEXT: s_bfe_i32 s28, s2, 0x1001b |
| ; GFX8-NEXT: s_bfe_i32 s29, s2, 0x1001a |
| ; GFX8-NEXT: s_bfe_i32 s30, s2, 0x10019 |
| ; GFX8-NEXT: s_bfe_i32 s31, s2, 0x10018 |
| ; GFX8-NEXT: s_ashr_i32 s3, s2, 31 |
| ; GFX8-NEXT: s_bfe_i32 s33, s2, 0x1001e |
| ; GFX8-NEXT: s_bfe_i32 s34, s2, 0x1001d |
| ; GFX8-NEXT: s_bfe_i32 s2, s2, 0x1001c |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 0x70 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s3 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 0x60 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s34 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s33 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 0x50 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s31 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s30 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s29 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s28 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 64 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s27 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s25 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s24 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 48 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s23 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s22 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s21 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s20 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 32 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s19 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s18 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s17 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s16 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s15 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s14 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s13 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s12 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s11 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s10 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s9 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s8 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s7 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s5 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v32i1_to_v32i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @16, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @14 |
| ; EG-NEXT: ALU 99, @17, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: ALU 5, @117, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T24.XYZW, T11.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T22.XYZW, T26.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T23.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T18.XYZW, T21.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T16.XYZW, T19.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T14.XYZW, T17.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T12.XYZW, T15.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T25.XYZW, T13.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 14: |
| ; EG-NEXT: VTX_READ_32 T11.X, T11.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 16: |
| ; EG-NEXT: MOV * T11.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 17: |
| ; EG-NEXT: LSHR * T0.W, T11.X, literal.x, |
| ; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T12.W, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T11.X, literal.x, |
| ; EG-NEXT: 6(8.407791e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T12.Z, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T11.X, literal.x, |
| ; EG-NEXT: LSHR * T1.W, T11.X, literal.y, |
| ; EG-NEXT: 11(1.541428e-44), 5(7.006492e-45) |
| ; EG-NEXT: LSHR T13.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: BFE_INT T12.Y, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T11.X, literal.y, |
| ; EG-NEXT: BFE_INT T14.W, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T11.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 10(1.401298e-44) |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T12.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T11.X, literal.x, |
| ; EG-NEXT: BFE_INT T14.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T11.X, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 15(2.101948e-44), 9(1.261169e-44) |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T15.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T14.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T11.X, literal.y, |
| ; EG-NEXT: BFE_INT T16.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T11.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 14(1.961818e-44) |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T14.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T11.X, literal.x, |
| ; EG-NEXT: BFE_INT T16.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T11.X, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 19(2.662467e-44), 13(1.821688e-44) |
| ; EG-NEXT: 32(4.484155e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T17.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T16.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T11.X, literal.y, |
| ; EG-NEXT: BFE_INT T18.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T11.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 18(2.522337e-44) |
| ; EG-NEXT: 12(1.681558e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T16.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T11.X, literal.x, |
| ; EG-NEXT: BFE_INT T18.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T11.X, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 23(3.222986e-44), 17(2.382207e-44) |
| ; EG-NEXT: 48(6.726233e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T19.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T18.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T11.X, literal.y, |
| ; EG-NEXT: BFE_INT T20.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T11.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 22(3.082857e-44) |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T18.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T11.X, literal.x, |
| ; EG-NEXT: BFE_INT T20.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T11.X, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 27(3.783506e-44), 21(2.942727e-44) |
| ; EG-NEXT: 64(8.968310e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T21.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T20.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T11.X, literal.y, |
| ; EG-NEXT: BFE_INT T22.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T11.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 26(3.643376e-44) |
| ; EG-NEXT: 20(2.802597e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T20.X, PS, 0.0, 1, |
| ; EG-NEXT: BFE_INT T22.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T11.X, literal.x, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 25(3.503246e-44), 80(1.121039e-43) |
| ; EG-NEXT: LSHR T23.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T22.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T11.X, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 24(3.363116e-44) |
| ; EG-NEXT: BFE_INT T22.X, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T11.X, literal.x, |
| ; EG-NEXT: ASHR * T24.W, T11.X, literal.y, |
| ; EG-NEXT: 30(4.203895e-44), 31(4.344025e-44) |
| ; EG-NEXT: BFE_INT T25.X, T11.X, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T11.X, literal.x, |
| ; EG-NEXT: BFE_INT T24.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T11.X, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 3(4.203895e-45), 29(4.063766e-44) |
| ; EG-NEXT: 96(1.345247e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T26.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T24.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T11.X, literal.x, |
| ; EG-NEXT: BFE_INT T25.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T11.X, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 28(3.923636e-44) |
| ; EG-NEXT: BFE_INT T24.X, PS, 0.0, 1, |
| ; EG-NEXT: BFE_INT * T25.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: ALU clause starting at 117: |
| ; EG-NEXT: LSHR T0.W, T11.X, 1, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 112(1.569454e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T11.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT * T25.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v32i1_to_v32i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_b32 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_i32 s3, s2, 0x10003 |
| ; GFX12-NEXT: s_bfe_i32 s4, s2, 0x10002 |
| ; GFX12-NEXT: s_bfe_i32 s5, s2, 0x10001 |
| ; GFX12-NEXT: s_bfe_i32 s6, s2, 0x10000 |
| ; GFX12-NEXT: s_bfe_i32 s7, s2, 0x10007 |
| ; GFX12-NEXT: s_bfe_i32 s8, s2, 0x10006 |
| ; GFX12-NEXT: s_bfe_i32 s9, s2, 0x10005 |
| ; GFX12-NEXT: s_bfe_i32 s10, s2, 0x10004 |
| ; GFX12-NEXT: s_bfe_i32 s11, s2, 0x1000b |
| ; GFX12-NEXT: s_bfe_i32 s12, s2, 0x1000a |
| ; GFX12-NEXT: s_bfe_i32 s13, s2, 0x10009 |
| ; GFX12-NEXT: s_bfe_i32 s14, s2, 0x10008 |
| ; GFX12-NEXT: s_bfe_i32 s15, s2, 0x1000f |
| ; GFX12-NEXT: s_bfe_i32 s16, s2, 0x1000e |
| ; GFX12-NEXT: s_bfe_i32 s17, s2, 0x1000d |
| ; GFX12-NEXT: s_bfe_i32 s18, s2, 0x1000c |
| ; GFX12-NEXT: s_bfe_i32 s19, s2, 0x10013 |
| ; GFX12-NEXT: s_bfe_i32 s20, s2, 0x10012 |
| ; GFX12-NEXT: s_bfe_i32 s21, s2, 0x10011 |
| ; GFX12-NEXT: s_bfe_i32 s22, s2, 0x10010 |
| ; GFX12-NEXT: s_bfe_i32 s23, s2, 0x10017 |
| ; GFX12-NEXT: s_bfe_i32 s24, s2, 0x10016 |
| ; GFX12-NEXT: s_bfe_i32 s25, s2, 0x10015 |
| ; GFX12-NEXT: s_bfe_i32 s26, s2, 0x10014 |
| ; GFX12-NEXT: s_bfe_i32 s27, s2, 0x1001b |
| ; GFX12-NEXT: s_bfe_i32 s28, s2, 0x1001a |
| ; GFX12-NEXT: s_bfe_i32 s29, s2, 0x10019 |
| ; GFX12-NEXT: s_bfe_i32 s30, s2, 0x10018 |
| ; GFX12-NEXT: s_ashr_i32 s31, s2, 31 |
| ; GFX12-NEXT: s_bfe_i32 s33, s2, 0x1001e |
| ; GFX12-NEXT: s_bfe_i32 s34, s2, 0x1001c |
| ; GFX12-NEXT: s_bfe_i32 s2, s2, 0x1001d |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s2 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s34 :: v_dual_mov_b32 v3, s31 |
| ; GFX12-NEXT: v_dual_mov_b32 v2, s33 :: v_dual_mov_b32 v5, s29 |
| ; GFX12-NEXT: v_dual_mov_b32 v4, s30 :: v_dual_mov_b32 v7, s27 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s28 :: v_dual_mov_b32 v9, s25 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s26 :: v_dual_mov_b32 v11, s23 |
| ; GFX12-NEXT: v_mov_b32_e32 v10, s24 |
| ; GFX12-NEXT: s_clause 0x1 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:112 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:96 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s21 :: v_dual_mov_b32 v0, s22 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s19 :: v_dual_mov_b32 v2, s20 |
| ; GFX12-NEXT: v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v4, s18 |
| ; GFX12-NEXT: v_dual_mov_b32 v7, s15 :: v_dual_mov_b32 v6, s16 |
| ; GFX12-NEXT: v_dual_mov_b32 v13, s13 :: v_dual_mov_b32 v12, s14 |
| ; GFX12-NEXT: v_dual_mov_b32 v15, s11 :: v_dual_mov_b32 v14, s12 |
| ; GFX12-NEXT: v_dual_mov_b32 v17, s9 :: v_dual_mov_b32 v16, s10 |
| ; GFX12-NEXT: v_dual_mov_b32 v19, s7 :: v_dual_mov_b32 v18, s8 |
| ; GFX12-NEXT: v_dual_mov_b32 v21, s5 :: v_dual_mov_b32 v20, s6 |
| ; GFX12-NEXT: v_dual_mov_b32 v23, s3 :: v_dual_mov_b32 v22, s4 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v24, v[8:11], s[0:1] offset:80 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:64 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:48 |
| ; GFX12-NEXT: global_store_b128 v24, v[12:15], s[0:1] offset:32 |
| ; GFX12-NEXT: global_store_b128 v24, v[16:19], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v24, v[20:23], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <32 x i1>, ptr addrspace(4) %in |
| %ext = sext <32 x i1> %load to <32 x i32> |
| store <32 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v64i1_to_v64i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v64i1_to_v64i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_bfe_u32 s4, s2, 0x10003 |
| ; GFX6-NEXT: s_bfe_u32 s5, s2, 0x10001 |
| ; GFX6-NEXT: s_bfe_u32 s6, s2, 0x10007 |
| ; GFX6-NEXT: s_bfe_u32 s7, s2, 0x10005 |
| ; GFX6-NEXT: s_bfe_u32 s8, s2, 0x1000b |
| ; GFX6-NEXT: s_bfe_u32 s9, s2, 0x10009 |
| ; GFX6-NEXT: s_bfe_u32 s10, s2, 0x1000f |
| ; GFX6-NEXT: s_bfe_u32 s13, s2, 0x1000d |
| ; GFX6-NEXT: s_bfe_u32 s14, s2, 0x10013 |
| ; GFX6-NEXT: s_bfe_u32 s15, s2, 0x10011 |
| ; GFX6-NEXT: s_bfe_u32 s16, s2, 0x10017 |
| ; GFX6-NEXT: s_bfe_u32 s17, s2, 0x10015 |
| ; GFX6-NEXT: s_bfe_u32 s18, s2, 0x1001b |
| ; GFX6-NEXT: s_bfe_u32 s19, s2, 0x10019 |
| ; GFX6-NEXT: s_lshr_b32 s20, s2, 31 |
| ; GFX6-NEXT: s_bfe_u32 s21, s2, 0x1001d |
| ; GFX6-NEXT: s_bfe_u32 s22, s3, 0x10003 |
| ; GFX6-NEXT: s_bfe_u32 s23, s3, 0x10001 |
| ; GFX6-NEXT: s_bfe_u32 s24, s3, 0x10007 |
| ; GFX6-NEXT: s_bfe_u32 s25, s3, 0x10005 |
| ; GFX6-NEXT: s_bfe_u32 s26, s3, 0x1000b |
| ; GFX6-NEXT: s_bfe_u32 s27, s3, 0x10009 |
| ; GFX6-NEXT: s_bfe_u32 s28, s3, 0x1000f |
| ; GFX6-NEXT: s_bfe_u32 s29, s3, 0x1000d |
| ; GFX6-NEXT: s_bfe_u32 s30, s3, 0x10013 |
| ; GFX6-NEXT: s_bfe_u32 s31, s3, 0x10011 |
| ; GFX6-NEXT: s_bfe_u32 s33, s3, 0x10017 |
| ; GFX6-NEXT: s_bfe_u32 s34, s3, 0x10015 |
| ; GFX6-NEXT: s_bfe_u32 s35, s3, 0x1001b |
| ; GFX6-NEXT: s_bfe_u32 s36, s3, 0x10019 |
| ; GFX6-NEXT: s_lshr_b32 s37, s3, 31 |
| ; GFX6-NEXT: s_bfe_u32 s38, s3, 0x1001d |
| ; GFX6-NEXT: s_and_b32 s12, s2, 1 |
| ; GFX6-NEXT: s_bfe_u32 s11, s2, 0x10002 |
| ; GFX6-NEXT: s_bfe_u32 s39, s2, 0x10006 |
| ; GFX6-NEXT: s_bfe_u32 s40, s2, 0x10004 |
| ; GFX6-NEXT: s_bfe_u32 s41, s2, 0x1000a |
| ; GFX6-NEXT: s_bfe_u32 s42, s2, 0x10008 |
| ; GFX6-NEXT: s_bfe_u32 s43, s2, 0x1000e |
| ; GFX6-NEXT: s_bfe_u32 s44, s2, 0x1000c |
| ; GFX6-NEXT: s_bfe_u32 s45, s2, 0x10012 |
| ; GFX6-NEXT: s_bfe_u32 s46, s2, 0x10010 |
| ; GFX6-NEXT: s_bfe_u32 s47, s2, 0x10016 |
| ; GFX6-NEXT: s_bfe_u32 s48, s2, 0x10014 |
| ; GFX6-NEXT: s_bfe_u32 s49, s2, 0x1001a |
| ; GFX6-NEXT: s_bfe_u32 s50, s2, 0x10018 |
| ; GFX6-NEXT: s_bfe_u32 s51, s2, 0x1001e |
| ; GFX6-NEXT: s_bfe_u32 s52, s2, 0x1001c |
| ; GFX6-NEXT: s_and_b32 s53, s3, 1 |
| ; GFX6-NEXT: s_bfe_u32 s54, s3, 0x10002 |
| ; GFX6-NEXT: s_bfe_u32 s55, s3, 0x10006 |
| ; GFX6-NEXT: s_bfe_u32 s56, s3, 0x10004 |
| ; GFX6-NEXT: s_bfe_u32 s57, s3, 0x10008 |
| ; GFX6-NEXT: s_bfe_u32 s58, s3, 0x1000e |
| ; GFX6-NEXT: s_bfe_u32 s59, s3, 0x1000c |
| ; GFX6-NEXT: s_bfe_u32 s60, s3, 0x10012 |
| ; GFX6-NEXT: s_bfe_u32 s61, s3, 0x10010 |
| ; GFX6-NEXT: s_bfe_u32 s62, s3, 0x10016 |
| ; GFX6-NEXT: s_bfe_u32 s63, s3, 0x10014 |
| ; GFX6-NEXT: s_bfe_u32 s64, s3, 0x1001a |
| ; GFX6-NEXT: s_bfe_u32 s65, s3, 0x10018 |
| ; GFX6-NEXT: s_bfe_u32 s66, s3, 0x1001e |
| ; GFX6-NEXT: s_bfe_u32 s67, s3, 0x1001c |
| ; GFX6-NEXT: s_bfe_u32 s68, s3, 0x1000a |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s67 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s38 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s66 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s37 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s65 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s36 |
| ; GFX6-NEXT: v_mov_b32_e32 v6, s64 |
| ; GFX6-NEXT: v_mov_b32_e32 v7, s35 |
| ; GFX6-NEXT: v_mov_b32_e32 v8, s63 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, s34 |
| ; GFX6-NEXT: v_mov_b32_e32 v10, s62 |
| ; GFX6-NEXT: v_mov_b32_e32 v11, s33 |
| ; GFX6-NEXT: v_mov_b32_e32 v12, s61 |
| ; GFX6-NEXT: v_mov_b32_e32 v13, s31 |
| ; GFX6-NEXT: v_mov_b32_e32 v14, s60 |
| ; GFX6-NEXT: v_mov_b32_e32 v15, s30 |
| ; GFX6-NEXT: v_mov_b32_e32 v16, s59 |
| ; GFX6-NEXT: v_mov_b32_e32 v17, s29 |
| ; GFX6-NEXT: v_mov_b32_e32 v18, s58 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:240 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s57 |
| ; GFX6-NEXT: v_mov_b32_e32 v19, s28 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s27 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s68 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s26 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:224 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:208 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:192 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:176 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:160 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s56 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s25 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s55 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s24 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:144 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s53 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s23 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s54 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s22 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:128 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s52 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s21 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s51 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s20 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:112 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s50 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s19 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s49 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s18 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:96 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s48 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s47 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s46 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s15 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s45 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s14 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s44 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s13 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s43 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s10 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s42 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s9 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s41 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s8 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s40 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s7 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s39 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s6 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s12 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s5 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s11 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s4 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v64i1_to_v64i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_load_dwordx2 s[26:27], s[2:3], 0x0 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_bfe_u32 s2, s26, 0x10003 |
| ; GFX8-NEXT: s_bfe_u32 s3, s26, 0x10001 |
| ; GFX8-NEXT: s_bfe_u32 s4, s26, 0x10007 |
| ; GFX8-NEXT: s_bfe_u32 s5, s26, 0x10005 |
| ; GFX8-NEXT: s_bfe_u32 s6, s26, 0x1000b |
| ; GFX8-NEXT: s_bfe_u32 s9, s26, 0x10009 |
| ; GFX8-NEXT: s_bfe_u32 s11, s26, 0x1000f |
| ; GFX8-NEXT: s_bfe_u32 s13, s26, 0x1000d |
| ; GFX8-NEXT: s_bfe_u32 s15, s26, 0x10013 |
| ; GFX8-NEXT: s_bfe_u32 s17, s26, 0x10011 |
| ; GFX8-NEXT: s_bfe_u32 s19, s26, 0x10017 |
| ; GFX8-NEXT: s_bfe_u32 s21, s26, 0x1001b |
| ; GFX8-NEXT: s_bfe_u32 s23, s26, 0x10019 |
| ; GFX8-NEXT: s_lshr_b32 s25, s26, 31 |
| ; GFX8-NEXT: s_bfe_u32 s28, s26, 0x1001d |
| ; GFX8-NEXT: s_bfe_u32 s29, s27, 0x10003 |
| ; GFX8-NEXT: s_bfe_u32 s30, s27, 0x10001 |
| ; GFX8-NEXT: s_bfe_u32 s31, s27, 0x10007 |
| ; GFX8-NEXT: s_bfe_u32 s33, s27, 0x10005 |
| ; GFX8-NEXT: s_bfe_u32 s34, s27, 0x1000b |
| ; GFX8-NEXT: s_bfe_u32 s35, s27, 0x10009 |
| ; GFX8-NEXT: s_bfe_u32 s36, s27, 0x1000f |
| ; GFX8-NEXT: s_bfe_u32 s37, s27, 0x1000d |
| ; GFX8-NEXT: s_bfe_u32 s38, s27, 0x10013 |
| ; GFX8-NEXT: s_bfe_u32 s39, s27, 0x10011 |
| ; GFX8-NEXT: s_bfe_u32 s40, s27, 0x10017 |
| ; GFX8-NEXT: s_bfe_u32 s41, s27, 0x1001b |
| ; GFX8-NEXT: s_bfe_u32 s42, s27, 0x10019 |
| ; GFX8-NEXT: s_lshr_b32 s43, s27, 31 |
| ; GFX8-NEXT: s_bfe_u32 s44, s27, 0x1001d |
| ; GFX8-NEXT: s_and_b32 s8, s26, 1 |
| ; GFX8-NEXT: s_bfe_u32 s7, s26, 0x10002 |
| ; GFX8-NEXT: s_bfe_u32 s10, s26, 0x10006 |
| ; GFX8-NEXT: s_bfe_u32 s12, s26, 0x10004 |
| ; GFX8-NEXT: s_bfe_u32 s14, s26, 0x1000a |
| ; GFX8-NEXT: s_bfe_u32 s16, s26, 0x10008 |
| ; GFX8-NEXT: s_bfe_u32 s18, s26, 0x1000e |
| ; GFX8-NEXT: s_bfe_u32 s20, s26, 0x1000c |
| ; GFX8-NEXT: s_bfe_u32 s22, s26, 0x10012 |
| ; GFX8-NEXT: s_bfe_u32 s24, s26, 0x10010 |
| ; GFX8-NEXT: s_bfe_u32 s45, s26, 0x10016 |
| ; GFX8-NEXT: s_bfe_u32 s46, s26, 0x10015 |
| ; GFX8-NEXT: s_bfe_u32 s47, s26, 0x10014 |
| ; GFX8-NEXT: s_bfe_u32 s48, s26, 0x1001a |
| ; GFX8-NEXT: s_bfe_u32 s49, s26, 0x10018 |
| ; GFX8-NEXT: s_bfe_u32 s50, s26, 0x1001e |
| ; GFX8-NEXT: s_bfe_u32 s51, s26, 0x1001c |
| ; GFX8-NEXT: s_and_b32 s52, s27, 1 |
| ; GFX8-NEXT: s_bfe_u32 s53, s27, 0x10002 |
| ; GFX8-NEXT: s_bfe_u32 s54, s27, 0x10006 |
| ; GFX8-NEXT: s_bfe_u32 s55, s27, 0x10004 |
| ; GFX8-NEXT: s_bfe_u32 s56, s27, 0x1000a |
| ; GFX8-NEXT: s_bfe_u32 s57, s27, 0x10008 |
| ; GFX8-NEXT: s_bfe_u32 s58, s27, 0x1000e |
| ; GFX8-NEXT: s_bfe_u32 s59, s27, 0x1000c |
| ; GFX8-NEXT: s_bfe_u32 s60, s27, 0x10012 |
| ; GFX8-NEXT: s_bfe_u32 s61, s27, 0x10010 |
| ; GFX8-NEXT: s_bfe_u32 s62, s27, 0x10016 |
| ; GFX8-NEXT: s_bfe_u32 s63, s27, 0x10015 |
| ; GFX8-NEXT: s_bfe_u32 s64, s27, 0x10014 |
| ; GFX8-NEXT: s_bfe_u32 s65, s27, 0x1001a |
| ; GFX8-NEXT: s_bfe_u32 s66, s27, 0x10018 |
| ; GFX8-NEXT: s_bfe_u32 s26, s27, 0x1001e |
| ; GFX8-NEXT: s_bfe_u32 s27, s27, 0x1001c |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s26 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xf0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s27 |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s44 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s43 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xe0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s66 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s65 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s41 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xd0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s64 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s63 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s62 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s40 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xc0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s61 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s39 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s60 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s38 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xb0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s59 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s58 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s36 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xa0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s57 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s35 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s56 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s34 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0x90 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s55 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s33 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s54 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s31 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0x80 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s52 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s30 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s53 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s29 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0x70 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s51 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s28 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s50 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s25 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0x60 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s49 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s23 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s48 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s21 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0x50 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s47 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s46 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s45 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s22 |
| ; GFX8-NEXT: s_add_u32 s22, s0, 64 |
| ; GFX8-NEXT: s_addc_u32 s23, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s22 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s24 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s15 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s23 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX8-NEXT: s_add_u32 s18, s0, 48 |
| ; GFX8-NEXT: s_addc_u32 s19, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s18 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s20 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s13 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s11 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s19 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s14 |
| ; GFX8-NEXT: s_add_u32 s14, s0, 32 |
| ; GFX8-NEXT: s_addc_u32 s15, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s14 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s9 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s15 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s4 |
| ; GFX8-NEXT: s_add_u32 s4, s0, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s5 |
| ; GFX8-NEXT: s_addc_u32 s5, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s12 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s10 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s5 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s8 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s7 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v64i1_to_v64i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @24, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @22 |
| ; EG-NEXT: ALU 96, @25, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: ALU 57, @122, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T48.XYZW, T50.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T46.XYZW, T49.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T44.XYZW, T47.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T42.XYZW, T45.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T40.XYZW, T43.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T38.XYZW, T41.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T36.XYZW, T39.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T35.XYZW, T37.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T33.XYZW, T21.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T31.XYZW, T34.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T29.XYZW, T32.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T27.XYZW, T30.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T25.XYZW, T28.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T23.XYZW, T26.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T24.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T22.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 22: |
| ; EG-NEXT: VTX_READ_64 T21.XY, T19.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 24: |
| ; EG-NEXT: MOV * T19.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 25: |
| ; EG-NEXT: BFE_UINT * T19.W, T21.X, literal.x, 1, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT * T19.Z, T21.X, literal.x, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T19.Y, T21.X, 1, 1, |
| ; EG-NEXT: BFE_UINT * T20.W, T21.X, literal.x, 1, |
| ; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00) |
| ; EG-NEXT: AND_INT T19.X, T21.X, 1, |
| ; EG-NEXT: BFE_UINT T20.Z, T21.X, literal.x, 1, |
| ; EG-NEXT: LSHR * T22.X, KC0[2].Y, literal.y, |
| ; EG-NEXT: 6(8.407791e-45), 2(2.802597e-45) |
| ; EG-NEXT: BFE_UINT T20.Y, T21.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT * T23.W, T21.X, literal.y, 1, |
| ; EG-NEXT: 5(7.006492e-45), 11(1.541428e-44) |
| ; EG-NEXT: BFE_UINT T20.X, T21.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T23.Z, T21.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 4(5.605194e-45), 10(1.401298e-44) |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T24.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T23.Y, T21.X, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T25.W, T21.X, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 9(1.261169e-44) |
| ; EG-NEXT: 15(2.101948e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T23.X, T21.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T25.Z, T21.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 8(1.121039e-44), 14(1.961818e-44) |
| ; EG-NEXT: 32(4.484155e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T26.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T25.Y, T21.X, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T27.W, T21.X, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 13(1.821688e-44) |
| ; EG-NEXT: 19(2.662467e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T25.X, T21.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T27.Z, T21.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 12(1.681558e-44), 18(2.522337e-44) |
| ; EG-NEXT: 48(6.726233e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T28.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T27.Y, T21.X, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T29.W, T21.X, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 17(2.382207e-44) |
| ; EG-NEXT: 23(3.222986e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T27.X, T21.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T29.Z, T21.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 16(2.242078e-44), 22(3.082857e-44) |
| ; EG-NEXT: 64(8.968310e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T30.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T29.Y, T21.X, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T31.W, T21.X, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 21(2.942727e-44) |
| ; EG-NEXT: 27(3.783506e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T29.X, T21.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T31.Z, T21.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 20(2.802597e-44), 26(3.643376e-44) |
| ; EG-NEXT: 80(1.121039e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T32.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T31.Y, T21.X, literal.y, 1, |
| ; EG-NEXT: LSHR * T33.W, T21.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 25(3.503246e-44) |
| ; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T31.X, T21.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T33.Z, T21.X, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 24(3.363116e-44), 30(4.203895e-44) |
| ; EG-NEXT: 96(1.345247e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T34.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T33.Y, T21.X, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T35.W, T21.Y, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 29(4.063766e-44) |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T33.X, T21.X, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T35.Z, T21.Y, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 28(3.923636e-44), 2(2.802597e-45) |
| ; EG-NEXT: 112(1.569454e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T21.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T35.Y, T21.Y, 1, 1, |
| ; EG-NEXT: BFE_UINT T36.W, T21.Y, literal.y, 1, |
| ; EG-NEXT: AND_INT * T35.X, T21.Y, 1, |
| ; EG-NEXT: 2(2.802597e-45), 7(9.809089e-45) |
| ; EG-NEXT: BFE_UINT T36.Z, T21.Y, literal.x, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 6(8.407791e-45), 128(1.793662e-43) |
| ; EG-NEXT: LSHR T37.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T36.Y, T21.Y, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T38.W, T21.Y, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 5(7.006492e-45) |
| ; EG-NEXT: 11(1.541428e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T36.X, T21.Y, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T38.Z, T21.Y, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 4(5.605194e-45), 10(1.401298e-44) |
| ; EG-NEXT: 144(2.017870e-43), 0(0.000000e+00) |
| ; EG-NEXT: ALU clause starting at 122: |
| ; EG-NEXT: LSHR T39.X, T0.W, literal.x, |
| ; EG-NEXT: BFE_UINT T38.Y, T21.Y, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T40.W, T21.Y, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 9(1.261169e-44) |
| ; EG-NEXT: 15(2.101948e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T38.X, T21.Y, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T40.Z, T21.Y, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 8(1.121039e-44), 14(1.961818e-44) |
| ; EG-NEXT: 160(2.242078e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T41.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T40.Y, T21.Y, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T42.W, T21.Y, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 13(1.821688e-44) |
| ; EG-NEXT: 19(2.662467e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T40.X, T21.Y, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T42.Z, T21.Y, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 12(1.681558e-44), 18(2.522337e-44) |
| ; EG-NEXT: 176(2.466285e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T43.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T42.Y, T21.Y, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T44.W, T21.Y, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 17(2.382207e-44) |
| ; EG-NEXT: 23(3.222986e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T42.X, T21.Y, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T44.Z, T21.Y, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 16(2.242078e-44), 22(3.082857e-44) |
| ; EG-NEXT: 192(2.690493e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T45.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T44.Y, T21.Y, literal.y, 1, |
| ; EG-NEXT: BFE_UINT * T46.W, T21.Y, literal.z, 1, |
| ; EG-NEXT: 2(2.802597e-45), 21(2.942727e-44) |
| ; EG-NEXT: 27(3.783506e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T44.X, T21.Y, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T46.Z, T21.Y, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 20(2.802597e-44), 26(3.643376e-44) |
| ; EG-NEXT: 208(2.914701e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T47.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT T46.Y, T21.Y, literal.y, 1, |
| ; EG-NEXT: LSHR * T48.W, T21.Y, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 25(3.503246e-44) |
| ; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T46.X, T21.Y, literal.x, 1, |
| ; EG-NEXT: BFE_UINT T48.Z, T21.Y, literal.y, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 24(3.363116e-44), 30(4.203895e-44) |
| ; EG-NEXT: 224(3.138909e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T49.X, PV.W, literal.x, |
| ; EG-NEXT: BFE_UINT * T48.Y, T21.Y, literal.y, 1, |
| ; EG-NEXT: 2(2.802597e-45), 29(4.063766e-44) |
| ; EG-NEXT: BFE_UINT T48.X, T21.Y, literal.x, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 28(3.923636e-44), 240(3.363116e-43) |
| ; EG-NEXT: LSHR * T50.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v64i1_to_v64i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_lshr_b32 s33, s3, 31 |
| ; GFX12-NEXT: s_bfe_u32 s34, s3, 0x1001d |
| ; GFX12-NEXT: s_bfe_u32 s65, s3, 0x1001c |
| ; GFX12-NEXT: s_bfe_u32 s66, s3, 0x1001e |
| ; GFX12-NEXT: s_bfe_u32 s30, s3, 0x1001b |
| ; GFX12-NEXT: s_bfe_u32 s31, s3, 0x10019 |
| ; GFX12-NEXT: s_bfe_u32 s63, s3, 0x1001a |
| ; GFX12-NEXT: s_bfe_u32 s64, s3, 0x10018 |
| ; GFX12-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s34 |
| ; GFX12-NEXT: s_bfe_u32 s29, s3, 0x10017 |
| ; GFX12-NEXT: s_bfe_u32 s60, s3, 0x10016 |
| ; GFX12-NEXT: s_bfe_u32 s61, s3, 0x10015 |
| ; GFX12-NEXT: s_bfe_u32 s62, s3, 0x10014 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s65 :: v_dual_mov_b32 v3, s33 |
| ; GFX12-NEXT: v_dual_mov_b32 v2, s66 :: v_dual_mov_b32 v5, s31 |
| ; GFX12-NEXT: s_bfe_u32 s27, s3, 0x10013 |
| ; GFX12-NEXT: s_bfe_u32 s28, s3, 0x10011 |
| ; GFX12-NEXT: s_bfe_u32 s58, s3, 0x10012 |
| ; GFX12-NEXT: s_bfe_u32 s59, s3, 0x10010 |
| ; GFX12-NEXT: v_dual_mov_b32 v4, s64 :: v_dual_mov_b32 v7, s30 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s63 :: v_dual_mov_b32 v9, s61 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s62 :: v_dual_mov_b32 v11, s29 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s60 :: v_dual_mov_b32 v13, s28 |
| ; GFX12-NEXT: s_bfe_u32 s19, s3, 0x10003 |
| ; GFX12-NEXT: s_bfe_u32 s20, s3, 0x10001 |
| ; GFX12-NEXT: s_bfe_u32 s21, s3, 0x10007 |
| ; GFX12-NEXT: s_bfe_u32 s22, s3, 0x10005 |
| ; GFX12-NEXT: s_bfe_u32 s23, s3, 0x1000b |
| ; GFX12-NEXT: s_bfe_u32 s24, s3, 0x10009 |
| ; GFX12-NEXT: s_bfe_u32 s25, s3, 0x1000f |
| ; GFX12-NEXT: s_bfe_u32 s26, s3, 0x1000d |
| ; GFX12-NEXT: s_and_b32 s51, s3, 1 |
| ; GFX12-NEXT: s_bfe_u32 s52, s3, 0x10002 |
| ; GFX12-NEXT: s_bfe_u32 s53, s3, 0x10006 |
| ; GFX12-NEXT: s_bfe_u32 s54, s3, 0x10004 |
| ; GFX12-NEXT: s_bfe_u32 s55, s3, 0x1000a |
| ; GFX12-NEXT: s_bfe_u32 s56, s3, 0x10008 |
| ; GFX12-NEXT: s_bfe_u32 s57, s3, 0x1000e |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s59 :: v_dual_mov_b32 v15, s27 |
| ; GFX12-NEXT: v_mov_b32_e32 v14, s58 |
| ; GFX12-NEXT: s_bfe_u32 s3, s3, 0x1000c |
| ; GFX12-NEXT: s_clause 0x3 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:240 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:224 |
| ; GFX12-NEXT: global_store_b128 v24, v[8:11], s[0:1] offset:208 |
| ; GFX12-NEXT: global_store_b128 v24, v[12:15], s[0:1] offset:192 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s26 :: v_dual_mov_b32 v0, s3 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s25 :: v_dual_mov_b32 v2, s57 |
| ; GFX12-NEXT: v_dual_mov_b32 v5, s24 :: v_dual_mov_b32 v4, s56 |
| ; GFX12-NEXT: v_dual_mov_b32 v7, s23 :: v_dual_mov_b32 v6, s55 |
| ; GFX12-NEXT: v_mov_b32_e32 v9, s22 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10003 |
| ; GFX12-NEXT: s_bfe_u32 s5, s2, 0x10001 |
| ; GFX12-NEXT: s_bfe_u32 s6, s2, 0x10007 |
| ; GFX12-NEXT: s_bfe_u32 s7, s2, 0x10005 |
| ; GFX12-NEXT: s_bfe_u32 s8, s2, 0x1000b |
| ; GFX12-NEXT: s_bfe_u32 s9, s2, 0x10009 |
| ; GFX12-NEXT: s_bfe_u32 s10, s2, 0x1000f |
| ; GFX12-NEXT: s_bfe_u32 s11, s2, 0x1000d |
| ; GFX12-NEXT: s_bfe_u32 s12, s2, 0x10013 |
| ; GFX12-NEXT: s_bfe_u32 s13, s2, 0x10011 |
| ; GFX12-NEXT: s_bfe_u32 s14, s2, 0x10017 |
| ; GFX12-NEXT: s_bfe_u32 s15, s2, 0x1001b |
| ; GFX12-NEXT: s_bfe_u32 s16, s2, 0x10019 |
| ; GFX12-NEXT: s_lshr_b32 s17, s2, 31 |
| ; GFX12-NEXT: s_bfe_u32 s18, s2, 0x1001d |
| ; GFX12-NEXT: s_and_b32 s35, s2, 1 |
| ; GFX12-NEXT: s_bfe_u32 s36, s2, 0x10002 |
| ; GFX12-NEXT: s_bfe_u32 s37, s2, 0x10006 |
| ; GFX12-NEXT: s_bfe_u32 s38, s2, 0x10004 |
| ; GFX12-NEXT: s_bfe_u32 s39, s2, 0x1000a |
| ; GFX12-NEXT: s_bfe_u32 s40, s2, 0x10008 |
| ; GFX12-NEXT: s_bfe_u32 s41, s2, 0x1000e |
| ; GFX12-NEXT: s_bfe_u32 s42, s2, 0x1000c |
| ; GFX12-NEXT: s_bfe_u32 s43, s2, 0x10012 |
| ; GFX12-NEXT: s_bfe_u32 s44, s2, 0x10010 |
| ; GFX12-NEXT: s_bfe_u32 s45, s2, 0x10016 |
| ; GFX12-NEXT: s_bfe_u32 s46, s2, 0x10015 |
| ; GFX12-NEXT: s_bfe_u32 s47, s2, 0x10014 |
| ; GFX12-NEXT: s_bfe_u32 s48, s2, 0x1001a |
| ; GFX12-NEXT: s_bfe_u32 s49, s2, 0x10018 |
| ; GFX12-NEXT: s_bfe_u32 s50, s2, 0x1001e |
| ; GFX12-NEXT: s_bfe_u32 s2, s2, 0x1001c |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s54 :: v_dual_mov_b32 v11, s21 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s53 :: v_dual_mov_b32 v13, s20 |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s51 :: v_dual_mov_b32 v15, s19 |
| ; GFX12-NEXT: v_dual_mov_b32 v14, s52 :: v_dual_mov_b32 v17, s18 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_dual_mov_b32 v16, s2 :: v_dual_mov_b32 v19, s17 |
| ; GFX12-NEXT: v_dual_mov_b32 v18, s50 :: v_dual_mov_b32 v21, s16 |
| ; GFX12-NEXT: v_dual_mov_b32 v20, s49 :: v_dual_mov_b32 v23, s15 |
| ; GFX12-NEXT: v_mov_b32_e32 v22, s48 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:176 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:160 |
| ; GFX12-NEXT: global_store_b128 v24, v[8:11], s[0:1] offset:144 |
| ; GFX12-NEXT: global_store_b128 v24, v[12:15], s[0:1] offset:128 |
| ; GFX12-NEXT: global_store_b128 v24, v[16:19], s[0:1] offset:112 |
| ; GFX12-NEXT: global_store_b128 v24, v[20:23], s[0:1] offset:96 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s46 :: v_dual_mov_b32 v0, s47 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s14 :: v_dual_mov_b32 v2, s45 |
| ; GFX12-NEXT: v_dual_mov_b32 v5, s13 :: v_dual_mov_b32 v4, s44 |
| ; GFX12-NEXT: v_dual_mov_b32 v7, s12 :: v_dual_mov_b32 v6, s43 |
| ; GFX12-NEXT: v_dual_mov_b32 v9, s11 :: v_dual_mov_b32 v8, s42 |
| ; GFX12-NEXT: v_dual_mov_b32 v11, s10 :: v_dual_mov_b32 v10, s41 |
| ; GFX12-NEXT: v_dual_mov_b32 v13, s9 :: v_dual_mov_b32 v12, s40 |
| ; GFX12-NEXT: v_dual_mov_b32 v15, s8 :: v_dual_mov_b32 v14, s39 |
| ; GFX12-NEXT: v_dual_mov_b32 v17, s7 :: v_dual_mov_b32 v16, s38 |
| ; GFX12-NEXT: v_dual_mov_b32 v19, s6 :: v_dual_mov_b32 v18, s37 |
| ; GFX12-NEXT: v_dual_mov_b32 v21, s5 :: v_dual_mov_b32 v20, s35 |
| ; GFX12-NEXT: v_dual_mov_b32 v23, s4 :: v_dual_mov_b32 v22, s36 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:80 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:64 |
| ; GFX12-NEXT: global_store_b128 v24, v[8:11], s[0:1] offset:48 |
| ; GFX12-NEXT: global_store_b128 v24, v[12:15], s[0:1] offset:32 |
| ; GFX12-NEXT: global_store_b128 v24, v[16:19], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v24, v[20:23], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <64 x i1>, ptr addrspace(4) %in |
| %ext = zext <64 x i1> %load to <64 x i32> |
| store <64 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v64i1_to_v64i32(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v64i1_to_v64i32: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_bfe_i32 s4, s2, 0x10003 |
| ; GFX6-NEXT: s_bfe_i32 s5, s2, 0x10002 |
| ; GFX6-NEXT: s_bfe_i32 s6, s2, 0x10001 |
| ; GFX6-NEXT: s_bfe_i32 s7, s2, 0x10000 |
| ; GFX6-NEXT: s_bfe_i32 s8, s2, 0x10007 |
| ; GFX6-NEXT: s_bfe_i32 s9, s2, 0x10006 |
| ; GFX6-NEXT: s_bfe_i32 s10, s2, 0x10005 |
| ; GFX6-NEXT: s_bfe_i32 s11, s2, 0x10004 |
| ; GFX6-NEXT: s_bfe_i32 s12, s2, 0x1000b |
| ; GFX6-NEXT: s_bfe_i32 s13, s2, 0x1000a |
| ; GFX6-NEXT: s_bfe_i32 s14, s2, 0x10009 |
| ; GFX6-NEXT: s_bfe_i32 s15, s2, 0x10008 |
| ; GFX6-NEXT: s_bfe_i32 s16, s2, 0x1000f |
| ; GFX6-NEXT: s_bfe_i32 s17, s2, 0x1000e |
| ; GFX6-NEXT: s_bfe_i32 s18, s2, 0x1000d |
| ; GFX6-NEXT: s_bfe_i32 s19, s2, 0x1000c |
| ; GFX6-NEXT: s_bfe_i32 s20, s2, 0x10013 |
| ; GFX6-NEXT: s_bfe_i32 s21, s2, 0x10012 |
| ; GFX6-NEXT: s_bfe_i32 s22, s2, 0x10011 |
| ; GFX6-NEXT: s_bfe_i32 s23, s2, 0x10010 |
| ; GFX6-NEXT: s_bfe_i32 s24, s2, 0x10017 |
| ; GFX6-NEXT: s_bfe_i32 s25, s2, 0x10016 |
| ; GFX6-NEXT: s_bfe_i32 s26, s2, 0x10015 |
| ; GFX6-NEXT: s_bfe_i32 s27, s2, 0x10014 |
| ; GFX6-NEXT: s_bfe_i32 s28, s2, 0x1001b |
| ; GFX6-NEXT: s_bfe_i32 s29, s2, 0x1001a |
| ; GFX6-NEXT: s_bfe_i32 s30, s2, 0x10019 |
| ; GFX6-NEXT: s_bfe_i32 s31, s2, 0x10018 |
| ; GFX6-NEXT: s_ashr_i32 s33, s2, 31 |
| ; GFX6-NEXT: s_bfe_i32 s34, s2, 0x1001e |
| ; GFX6-NEXT: s_bfe_i32 s35, s2, 0x1001d |
| ; GFX6-NEXT: s_bfe_i32 s36, s2, 0x1001c |
| ; GFX6-NEXT: s_bfe_i32 s37, s3, 0x10003 |
| ; GFX6-NEXT: s_bfe_i32 s38, s3, 0x10002 |
| ; GFX6-NEXT: s_bfe_i32 s39, s3, 0x10001 |
| ; GFX6-NEXT: s_bfe_i32 s40, s3, 0x10000 |
| ; GFX6-NEXT: s_bfe_i32 s41, s3, 0x10007 |
| ; GFX6-NEXT: s_bfe_i32 s42, s3, 0x10006 |
| ; GFX6-NEXT: s_bfe_i32 s43, s3, 0x10005 |
| ; GFX6-NEXT: s_bfe_i32 s44, s3, 0x10004 |
| ; GFX6-NEXT: s_bfe_i32 s45, s3, 0x1000b |
| ; GFX6-NEXT: s_bfe_i32 s46, s3, 0x1000a |
| ; GFX6-NEXT: s_bfe_i32 s47, s3, 0x10009 |
| ; GFX6-NEXT: s_bfe_i32 s48, s3, 0x10008 |
| ; GFX6-NEXT: s_bfe_i32 s49, s3, 0x1000e |
| ; GFX6-NEXT: s_bfe_i32 s50, s3, 0x1000d |
| ; GFX6-NEXT: s_bfe_i32 s51, s3, 0x1000c |
| ; GFX6-NEXT: s_bfe_i32 s52, s3, 0x10013 |
| ; GFX6-NEXT: s_bfe_i32 s53, s3, 0x10012 |
| ; GFX6-NEXT: s_bfe_i32 s54, s3, 0x10011 |
| ; GFX6-NEXT: s_bfe_i32 s55, s3, 0x10010 |
| ; GFX6-NEXT: s_bfe_i32 s56, s3, 0x10017 |
| ; GFX6-NEXT: s_bfe_i32 s57, s3, 0x10016 |
| ; GFX6-NEXT: s_bfe_i32 s58, s3, 0x10015 |
| ; GFX6-NEXT: s_bfe_i32 s59, s3, 0x10014 |
| ; GFX6-NEXT: s_bfe_i32 s60, s3, 0x1001b |
| ; GFX6-NEXT: s_bfe_i32 s61, s3, 0x1001a |
| ; GFX6-NEXT: s_bfe_i32 s62, s3, 0x10019 |
| ; GFX6-NEXT: s_bfe_i32 s63, s3, 0x10018 |
| ; GFX6-NEXT: s_ashr_i32 s64, s3, 31 |
| ; GFX6-NEXT: s_bfe_i32 s65, s3, 0x1001e |
| ; GFX6-NEXT: s_bfe_i32 s66, s3, 0x1001d |
| ; GFX6-NEXT: s_bfe_i32 s67, s3, 0x1001c |
| ; GFX6-NEXT: s_bfe_i32 s68, s3, 0x1000f |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s67 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s66 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s65 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s64 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s63 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s62 |
| ; GFX6-NEXT: v_mov_b32_e32 v6, s61 |
| ; GFX6-NEXT: v_mov_b32_e32 v7, s60 |
| ; GFX6-NEXT: v_mov_b32_e32 v8, s59 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, s58 |
| ; GFX6-NEXT: v_mov_b32_e32 v10, s57 |
| ; GFX6-NEXT: v_mov_b32_e32 v11, s56 |
| ; GFX6-NEXT: v_mov_b32_e32 v12, s55 |
| ; GFX6-NEXT: v_mov_b32_e32 v13, s54 |
| ; GFX6-NEXT: v_mov_b32_e32 v14, s53 |
| ; GFX6-NEXT: v_mov_b32_e32 v15, s52 |
| ; GFX6-NEXT: v_mov_b32_e32 v16, s51 |
| ; GFX6-NEXT: v_mov_b32_e32 v17, s50 |
| ; GFX6-NEXT: v_mov_b32_e32 v18, s49 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:240 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s48 |
| ; GFX6-NEXT: v_mov_b32_e32 v19, s68 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s47 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s46 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s45 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:224 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:208 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:192 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:176 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:160 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s44 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s43 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s42 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s41 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:144 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s40 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s39 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s37 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:128 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s35 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s34 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s33 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:112 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s31 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s30 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s29 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s28 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:96 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s27 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s26 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s25 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s24 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s23 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s22 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s21 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s20 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s19 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s18 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s17 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s15 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s14 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s13 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s12 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s11 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s10 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s9 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s8 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s7 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s6 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s5 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s4 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v64i1_to_v64i32: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_load_dwordx2 s[26:27], s[2:3], 0x0 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_bfe_i32 s2, s26, 0x10003 |
| ; GFX8-NEXT: s_bfe_i32 s3, s26, 0x10002 |
| ; GFX8-NEXT: s_bfe_i32 s4, s26, 0x10001 |
| ; GFX8-NEXT: s_bfe_i32 s5, s26, 0x10000 |
| ; GFX8-NEXT: s_bfe_i32 s6, s26, 0x10007 |
| ; GFX8-NEXT: s_bfe_i32 s7, s26, 0x10006 |
| ; GFX8-NEXT: s_bfe_i32 s8, s26, 0x10005 |
| ; GFX8-NEXT: s_bfe_i32 s9, s26, 0x10004 |
| ; GFX8-NEXT: s_bfe_i32 s10, s26, 0x1000b |
| ; GFX8-NEXT: s_bfe_i32 s11, s26, 0x1000a |
| ; GFX8-NEXT: s_bfe_i32 s12, s26, 0x10009 |
| ; GFX8-NEXT: s_bfe_i32 s13, s26, 0x10008 |
| ; GFX8-NEXT: s_bfe_i32 s14, s26, 0x1000f |
| ; GFX8-NEXT: s_bfe_i32 s15, s26, 0x1000e |
| ; GFX8-NEXT: s_bfe_i32 s16, s26, 0x1000d |
| ; GFX8-NEXT: s_bfe_i32 s17, s26, 0x1000c |
| ; GFX8-NEXT: s_bfe_i32 s18, s26, 0x10013 |
| ; GFX8-NEXT: s_bfe_i32 s19, s26, 0x10012 |
| ; GFX8-NEXT: s_bfe_i32 s20, s26, 0x10011 |
| ; GFX8-NEXT: s_bfe_i32 s21, s26, 0x10010 |
| ; GFX8-NEXT: s_bfe_i32 s22, s26, 0x10017 |
| ; GFX8-NEXT: s_bfe_i32 s23, s26, 0x10016 |
| ; GFX8-NEXT: s_bfe_i32 s24, s26, 0x10015 |
| ; GFX8-NEXT: s_bfe_i32 s25, s26, 0x10014 |
| ; GFX8-NEXT: s_bfe_i32 s28, s26, 0x1001b |
| ; GFX8-NEXT: s_bfe_i32 s29, s26, 0x1001a |
| ; GFX8-NEXT: s_bfe_i32 s30, s26, 0x10019 |
| ; GFX8-NEXT: s_bfe_i32 s31, s26, 0x10018 |
| ; GFX8-NEXT: s_ashr_i32 s33, s26, 31 |
| ; GFX8-NEXT: s_bfe_i32 s34, s26, 0x1001e |
| ; GFX8-NEXT: s_bfe_i32 s35, s26, 0x1001d |
| ; GFX8-NEXT: s_bfe_i32 s36, s26, 0x1001c |
| ; GFX8-NEXT: s_bfe_i32 s37, s27, 0x10003 |
| ; GFX8-NEXT: s_bfe_i32 s38, s27, 0x10002 |
| ; GFX8-NEXT: s_bfe_i32 s39, s27, 0x10001 |
| ; GFX8-NEXT: s_bfe_i32 s40, s27, 0x10000 |
| ; GFX8-NEXT: s_bfe_i32 s41, s27, 0x10007 |
| ; GFX8-NEXT: s_bfe_i32 s42, s27, 0x10006 |
| ; GFX8-NEXT: s_bfe_i32 s43, s27, 0x10005 |
| ; GFX8-NEXT: s_bfe_i32 s44, s27, 0x10004 |
| ; GFX8-NEXT: s_bfe_i32 s45, s27, 0x1000b |
| ; GFX8-NEXT: s_bfe_i32 s46, s27, 0x1000a |
| ; GFX8-NEXT: s_bfe_i32 s47, s27, 0x10009 |
| ; GFX8-NEXT: s_bfe_i32 s48, s27, 0x10008 |
| ; GFX8-NEXT: s_bfe_i32 s49, s27, 0x1000f |
| ; GFX8-NEXT: s_bfe_i32 s50, s27, 0x1000e |
| ; GFX8-NEXT: s_bfe_i32 s51, s27, 0x1000d |
| ; GFX8-NEXT: s_bfe_i32 s52, s27, 0x1000c |
| ; GFX8-NEXT: s_bfe_i32 s53, s27, 0x10013 |
| ; GFX8-NEXT: s_bfe_i32 s54, s27, 0x10012 |
| ; GFX8-NEXT: s_bfe_i32 s55, s27, 0x10011 |
| ; GFX8-NEXT: s_bfe_i32 s56, s27, 0x10010 |
| ; GFX8-NEXT: s_bfe_i32 s57, s27, 0x10017 |
| ; GFX8-NEXT: s_bfe_i32 s58, s27, 0x10016 |
| ; GFX8-NEXT: s_bfe_i32 s59, s27, 0x10015 |
| ; GFX8-NEXT: s_bfe_i32 s60, s27, 0x10014 |
| ; GFX8-NEXT: s_bfe_i32 s61, s27, 0x1001b |
| ; GFX8-NEXT: s_bfe_i32 s62, s27, 0x1001a |
| ; GFX8-NEXT: s_bfe_i32 s63, s27, 0x10019 |
| ; GFX8-NEXT: s_bfe_i32 s64, s27, 0x10018 |
| ; GFX8-NEXT: s_ashr_i32 s26, s27, 31 |
| ; GFX8-NEXT: s_bfe_i32 s65, s27, 0x1001e |
| ; GFX8-NEXT: s_bfe_i32 s66, s27, 0x1001d |
| ; GFX8-NEXT: s_bfe_i32 s27, s27, 0x1001c |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s26 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xf0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s27 |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s66 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s65 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xe0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s64 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s63 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s62 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s61 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xd0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s60 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s59 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s58 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s57 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xc0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s56 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s55 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s54 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s53 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xb0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s52 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s51 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s50 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s49 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xa0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s48 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s47 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s46 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s45 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0x90 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s44 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s43 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s41 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0x80 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s40 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s39 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s37 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0x70 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s35 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s34 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s33 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0x60 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s31 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s30 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s29 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s28 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s22 |
| ; GFX8-NEXT: s_add_u32 s22, s0, 0x50 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s23 |
| ; GFX8-NEXT: s_addc_u32 s23, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s22 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s25 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s24 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s23 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s18 |
| ; GFX8-NEXT: s_add_u32 s18, s0, 64 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s19 |
| ; GFX8-NEXT: s_addc_u32 s19, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s18 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s21 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s20 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s19 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s14 |
| ; GFX8-NEXT: s_add_u32 s14, s0, 48 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s15 |
| ; GFX8-NEXT: s_addc_u32 s15, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s14 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s17 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s16 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s15 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s10 |
| ; GFX8-NEXT: s_add_u32 s10, s0, 32 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s11 |
| ; GFX8-NEXT: s_addc_u32 s11, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s10 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s13 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s12 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s11 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s6 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s7 |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s9 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s8 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v64i1_to_v64i32: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @24, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @22 |
| ; EG-NEXT: ALU 99, @25, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: ALU 98, @125, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: ALU 13, @224, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T48.XYZW, T50.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T45.XYZW, T49.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T43.XYZW, T46.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T41.XYZW, T44.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T39.XYZW, T42.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T37.XYZW, T40.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T34.XYZW, T38.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T36.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T32.XYZW, T35.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T30.XYZW, T33.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T28.XYZW, T31.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T26.XYZW, T29.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T24.XYZW, T27.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T22.XYZW, T25.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T23.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T47.XYZW, T21.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 22: |
| ; EG-NEXT: VTX_READ_64 T19.XY, T19.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 24: |
| ; EG-NEXT: MOV * T19.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 25: |
| ; EG-NEXT: LSHR * T0.W, T19.X, literal.x, |
| ; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T20.W, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T19.X, literal.x, |
| ; EG-NEXT: 6(8.407791e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T20.Z, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.X, literal.x, |
| ; EG-NEXT: LSHR * T1.W, T19.X, literal.y, |
| ; EG-NEXT: 11(1.541428e-44), 5(7.006492e-45) |
| ; EG-NEXT: LSHR T21.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: BFE_INT T20.Y, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.X, literal.y, |
| ; EG-NEXT: BFE_INT T22.W, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T19.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 10(1.401298e-44) |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T20.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T19.X, literal.x, |
| ; EG-NEXT: BFE_INT T22.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.X, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 15(2.101948e-44), 9(1.261169e-44) |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T23.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T22.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.X, literal.y, |
| ; EG-NEXT: BFE_INT T24.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T19.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 14(1.961818e-44) |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T22.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T19.X, literal.x, |
| ; EG-NEXT: BFE_INT T24.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.X, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 19(2.662467e-44), 13(1.821688e-44) |
| ; EG-NEXT: 32(4.484155e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T25.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T24.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.X, literal.y, |
| ; EG-NEXT: BFE_INT T26.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T19.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 18(2.522337e-44) |
| ; EG-NEXT: 12(1.681558e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T24.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T19.X, literal.x, |
| ; EG-NEXT: BFE_INT T26.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.X, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 23(3.222986e-44), 17(2.382207e-44) |
| ; EG-NEXT: 48(6.726233e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T27.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T26.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.X, literal.y, |
| ; EG-NEXT: BFE_INT T28.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T19.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 22(3.082857e-44) |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T26.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T19.X, literal.x, |
| ; EG-NEXT: BFE_INT T28.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.X, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 27(3.783506e-44), 21(2.942727e-44) |
| ; EG-NEXT: 64(8.968310e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T29.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T28.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.X, literal.y, |
| ; EG-NEXT: BFE_INT T30.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T19.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 26(3.643376e-44) |
| ; EG-NEXT: 20(2.802597e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T28.X, PS, 0.0, 1, |
| ; EG-NEXT: BFE_INT T30.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.X, literal.x, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 25(3.503246e-44), 80(1.121039e-43) |
| ; EG-NEXT: LSHR T31.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T30.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.X, literal.y, |
| ; EG-NEXT: LSHR T0.W, T19.X, literal.z, |
| ; EG-NEXT: ASHR * T32.W, T19.X, literal.w, |
| ; EG-NEXT: 2(2.802597e-45), 30(4.203895e-44) |
| ; EG-NEXT: 24(3.363116e-44), 31(4.344025e-44) |
| ; EG-NEXT: BFE_INT T30.X, PV.W, 0.0, 1, |
| ; EG-NEXT: BFE_INT T32.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.X, literal.x, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 29(4.063766e-44), 96(1.345247e-43) |
| ; EG-NEXT: LSHR T33.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T32.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.Y, literal.y, |
| ; EG-NEXT: LSHR * T1.W, T19.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 7(9.809089e-45) |
| ; EG-NEXT: 28(3.923636e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T32.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.Y, literal.x, |
| ; EG-NEXT: BFE_INT T34.W, PV.W, 0.0, 1, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 6(8.407791e-45), 112(1.569454e-43) |
| ; EG-NEXT: ALU clause starting at 125: |
| ; EG-NEXT: LSHR T35.X, T0.W, literal.x, |
| ; EG-NEXT: LSHR T0.Y, T19.Y, literal.y, |
| ; EG-NEXT: BFE_INT T34.Z, T0.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.Y, literal.z, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.w, |
| ; EG-NEXT: 2(2.802597e-45), 11(1.541428e-44) |
| ; EG-NEXT: 5(7.006492e-45), 128(1.793662e-43) |
| ; EG-NEXT: LSHR T36.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T34.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.Y, literal.y, |
| ; EG-NEXT: BFE_INT T37.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T19.Y, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 10(1.401298e-44) |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T34.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T19.Y, literal.x, |
| ; EG-NEXT: BFE_INT T37.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.Y, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 15(2.101948e-44), 9(1.261169e-44) |
| ; EG-NEXT: 144(2.017870e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T38.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T37.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.Y, literal.y, |
| ; EG-NEXT: BFE_INT T39.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T19.Y, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 14(1.961818e-44) |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T37.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T19.Y, literal.x, |
| ; EG-NEXT: BFE_INT T39.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.Y, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 19(2.662467e-44), 13(1.821688e-44) |
| ; EG-NEXT: 160(2.242078e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T40.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T39.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.Y, literal.y, |
| ; EG-NEXT: BFE_INT T41.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T19.Y, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 18(2.522337e-44) |
| ; EG-NEXT: 12(1.681558e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T39.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T19.Y, literal.x, |
| ; EG-NEXT: BFE_INT T41.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.Y, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 23(3.222986e-44), 17(2.382207e-44) |
| ; EG-NEXT: 176(2.466285e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T42.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T41.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.Y, literal.y, |
| ; EG-NEXT: BFE_INT T43.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T19.Y, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 22(3.082857e-44) |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T41.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T19.Y, literal.x, |
| ; EG-NEXT: BFE_INT T43.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.Y, literal.y, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z, |
| ; EG-NEXT: 27(3.783506e-44), 21(2.942727e-44) |
| ; EG-NEXT: 192(2.690493e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T44.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T43.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.Y, literal.y, |
| ; EG-NEXT: BFE_INT T45.W, PV.Y, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T19.Y, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 26(3.643376e-44) |
| ; EG-NEXT: 20(2.802597e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T43.X, PS, 0.0, 1, |
| ; EG-NEXT: BFE_INT T45.Z, PV.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.Y, literal.x, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 25(3.503246e-44), 208(2.914701e-43) |
| ; EG-NEXT: LSHR T46.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT T45.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T19.Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 24(3.363116e-44) |
| ; EG-NEXT: BFE_INT T45.X, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Z, T19.Y, literal.x, |
| ; EG-NEXT: LSHR T0.W, T19.X, 1, |
| ; EG-NEXT: LSHR * T1.W, T19.Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 3(4.203895e-45) |
| ; EG-NEXT: BFE_INT T47.X, T19.X, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T19.X, literal.x, |
| ; EG-NEXT: LSHR T1.Z, T19.X, literal.y, |
| ; EG-NEXT: LSHR T2.W, T19.Y, literal.z, |
| ; EG-NEXT: ASHR * T48.W, T19.Y, literal.w, |
| ; EG-NEXT: 2(2.802597e-45), 3(4.203895e-45) |
| ; EG-NEXT: 30(4.203895e-44), 31(4.344025e-44) |
| ; EG-NEXT: BFE_INT T19.X, T19.Y, 0.0, 1, |
| ; EG-NEXT: LSHR T1.Y, T19.Y, literal.x, |
| ; EG-NEXT: BFE_INT T48.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: BFE_INT T47.W, PV.Z, 0.0, 1, |
| ; EG-NEXT: ADD_INT * T2.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 29(4.063766e-44), 224(3.138909e-43) |
| ; EG-NEXT: LSHR * T49.X, PS, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: ALU clause starting at 224: |
| ; EG-NEXT: BFE_INT T48.Y, T1.Y, 0.0, 1, |
| ; EG-NEXT: BFE_INT T47.Z, T0.Y, 0.0, 1, BS:VEC_120/SCL_212 |
| ; EG-NEXT: BFE_INT T19.W, T1.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T1.W, T19.Y, literal.x, |
| ; EG-NEXT: 28(3.923636e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T48.X, PS, 0.0, 1, |
| ; EG-NEXT: BFE_INT T47.Y, T0.W, 0.0, 1, |
| ; EG-NEXT: BFE_INT T19.Z, T0.Z, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T19.Y, 1, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 240(3.363116e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T50.X, PS, literal.x, |
| ; EG-NEXT: BFE_INT * T19.Y, PV.W, 0.0, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v64i1_to_v64i32: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_ashr_i32 s63, s3, 31 |
| ; GFX12-NEXT: s_bfe_i32 s64, s3, 0x1001e |
| ; GFX12-NEXT: s_bfe_i32 s65, s3, 0x1001c |
| ; GFX12-NEXT: s_bfe_i32 s66, s3, 0x1001d |
| ; GFX12-NEXT: s_bfe_i32 s59, s3, 0x1001b |
| ; GFX12-NEXT: s_bfe_i32 s60, s3, 0x1001a |
| ; GFX12-NEXT: s_bfe_i32 s61, s3, 0x10019 |
| ; GFX12-NEXT: s_bfe_i32 s62, s3, 0x10018 |
| ; GFX12-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s66 |
| ; GFX12-NEXT: s_bfe_i32 s55, s3, 0x10017 |
| ; GFX12-NEXT: s_bfe_i32 s56, s3, 0x10016 |
| ; GFX12-NEXT: s_bfe_i32 s57, s3, 0x10015 |
| ; GFX12-NEXT: s_bfe_i32 s58, s3, 0x10014 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s65 :: v_dual_mov_b32 v3, s63 |
| ; GFX12-NEXT: v_dual_mov_b32 v2, s64 :: v_dual_mov_b32 v5, s61 |
| ; GFX12-NEXT: s_bfe_i32 s51, s3, 0x10013 |
| ; GFX12-NEXT: s_bfe_i32 s52, s3, 0x10012 |
| ; GFX12-NEXT: s_bfe_i32 s53, s3, 0x10011 |
| ; GFX12-NEXT: s_bfe_i32 s54, s3, 0x10010 |
| ; GFX12-NEXT: v_dual_mov_b32 v4, s62 :: v_dual_mov_b32 v7, s59 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s60 :: v_dual_mov_b32 v9, s57 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s58 :: v_dual_mov_b32 v11, s55 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s56 :: v_dual_mov_b32 v13, s53 |
| ; GFX12-NEXT: s_bfe_i32 s36, s3, 0x10003 |
| ; GFX12-NEXT: s_bfe_i32 s37, s3, 0x10002 |
| ; GFX12-NEXT: s_bfe_i32 s38, s3, 0x10001 |
| ; GFX12-NEXT: s_bfe_i32 s39, s3, 0x10000 |
| ; GFX12-NEXT: s_bfe_i32 s40, s3, 0x10007 |
| ; GFX12-NEXT: s_bfe_i32 s41, s3, 0x10006 |
| ; GFX12-NEXT: s_bfe_i32 s42, s3, 0x10005 |
| ; GFX12-NEXT: s_bfe_i32 s43, s3, 0x10004 |
| ; GFX12-NEXT: s_bfe_i32 s44, s3, 0x1000b |
| ; GFX12-NEXT: s_bfe_i32 s45, s3, 0x1000a |
| ; GFX12-NEXT: s_bfe_i32 s46, s3, 0x10009 |
| ; GFX12-NEXT: s_bfe_i32 s47, s3, 0x10008 |
| ; GFX12-NEXT: s_bfe_i32 s48, s3, 0x1000f |
| ; GFX12-NEXT: s_bfe_i32 s49, s3, 0x1000e |
| ; GFX12-NEXT: s_bfe_i32 s50, s3, 0x1000d |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s54 :: v_dual_mov_b32 v15, s51 |
| ; GFX12-NEXT: v_mov_b32_e32 v14, s52 |
| ; GFX12-NEXT: s_bfe_i32 s3, s3, 0x1000c |
| ; GFX12-NEXT: s_clause 0x3 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:240 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:224 |
| ; GFX12-NEXT: global_store_b128 v24, v[8:11], s[0:1] offset:208 |
| ; GFX12-NEXT: global_store_b128 v24, v[12:15], s[0:1] offset:192 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s50 :: v_dual_mov_b32 v0, s3 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s48 :: v_dual_mov_b32 v2, s49 |
| ; GFX12-NEXT: v_dual_mov_b32 v5, s46 :: v_dual_mov_b32 v4, s47 |
| ; GFX12-NEXT: v_dual_mov_b32 v7, s44 :: v_dual_mov_b32 v6, s45 |
| ; GFX12-NEXT: v_mov_b32_e32 v9, s42 |
| ; GFX12-NEXT: s_bfe_i32 s4, s2, 0x10003 |
| ; GFX12-NEXT: s_bfe_i32 s5, s2, 0x10002 |
| ; GFX12-NEXT: s_bfe_i32 s6, s2, 0x10001 |
| ; GFX12-NEXT: s_bfe_i32 s7, s2, 0x10000 |
| ; GFX12-NEXT: s_bfe_i32 s8, s2, 0x10007 |
| ; GFX12-NEXT: s_bfe_i32 s9, s2, 0x10006 |
| ; GFX12-NEXT: s_bfe_i32 s10, s2, 0x10005 |
| ; GFX12-NEXT: s_bfe_i32 s11, s2, 0x10004 |
| ; GFX12-NEXT: s_bfe_i32 s12, s2, 0x1000b |
| ; GFX12-NEXT: s_bfe_i32 s13, s2, 0x1000a |
| ; GFX12-NEXT: s_bfe_i32 s14, s2, 0x10009 |
| ; GFX12-NEXT: s_bfe_i32 s15, s2, 0x10008 |
| ; GFX12-NEXT: s_bfe_i32 s16, s2, 0x1000f |
| ; GFX12-NEXT: s_bfe_i32 s17, s2, 0x1000e |
| ; GFX12-NEXT: s_bfe_i32 s18, s2, 0x1000d |
| ; GFX12-NEXT: s_bfe_i32 s19, s2, 0x1000c |
| ; GFX12-NEXT: s_bfe_i32 s20, s2, 0x10013 |
| ; GFX12-NEXT: s_bfe_i32 s21, s2, 0x10012 |
| ; GFX12-NEXT: s_bfe_i32 s22, s2, 0x10011 |
| ; GFX12-NEXT: s_bfe_i32 s23, s2, 0x10010 |
| ; GFX12-NEXT: s_bfe_i32 s24, s2, 0x10017 |
| ; GFX12-NEXT: s_bfe_i32 s25, s2, 0x10016 |
| ; GFX12-NEXT: s_bfe_i32 s26, s2, 0x10015 |
| ; GFX12-NEXT: s_bfe_i32 s27, s2, 0x10014 |
| ; GFX12-NEXT: s_bfe_i32 s28, s2, 0x1001b |
| ; GFX12-NEXT: s_bfe_i32 s29, s2, 0x1001a |
| ; GFX12-NEXT: s_bfe_i32 s30, s2, 0x10019 |
| ; GFX12-NEXT: s_bfe_i32 s31, s2, 0x10018 |
| ; GFX12-NEXT: s_ashr_i32 s33, s2, 31 |
| ; GFX12-NEXT: s_bfe_i32 s34, s2, 0x1001e |
| ; GFX12-NEXT: s_bfe_i32 s35, s2, 0x1001d |
| ; GFX12-NEXT: s_bfe_i32 s2, s2, 0x1001c |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s43 :: v_dual_mov_b32 v11, s40 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s41 :: v_dual_mov_b32 v13, s38 |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s39 :: v_dual_mov_b32 v15, s36 |
| ; GFX12-NEXT: v_dual_mov_b32 v14, s37 :: v_dual_mov_b32 v17, s35 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_dual_mov_b32 v16, s2 :: v_dual_mov_b32 v19, s33 |
| ; GFX12-NEXT: v_dual_mov_b32 v18, s34 :: v_dual_mov_b32 v21, s30 |
| ; GFX12-NEXT: v_dual_mov_b32 v20, s31 :: v_dual_mov_b32 v23, s28 |
| ; GFX12-NEXT: v_mov_b32_e32 v22, s29 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:176 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:160 |
| ; GFX12-NEXT: global_store_b128 v24, v[8:11], s[0:1] offset:144 |
| ; GFX12-NEXT: global_store_b128 v24, v[12:15], s[0:1] offset:128 |
| ; GFX12-NEXT: global_store_b128 v24, v[16:19], s[0:1] offset:112 |
| ; GFX12-NEXT: global_store_b128 v24, v[20:23], s[0:1] offset:96 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s26 :: v_dual_mov_b32 v0, s27 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s24 :: v_dual_mov_b32 v2, s25 |
| ; GFX12-NEXT: v_dual_mov_b32 v5, s22 :: v_dual_mov_b32 v4, s23 |
| ; GFX12-NEXT: v_dual_mov_b32 v7, s20 :: v_dual_mov_b32 v6, s21 |
| ; GFX12-NEXT: v_dual_mov_b32 v9, s18 :: v_dual_mov_b32 v8, s19 |
| ; GFX12-NEXT: v_dual_mov_b32 v11, s16 :: v_dual_mov_b32 v10, s17 |
| ; GFX12-NEXT: v_dual_mov_b32 v13, s14 :: v_dual_mov_b32 v12, s15 |
| ; GFX12-NEXT: v_dual_mov_b32 v15, s12 :: v_dual_mov_b32 v14, s13 |
| ; GFX12-NEXT: v_dual_mov_b32 v17, s10 :: v_dual_mov_b32 v16, s11 |
| ; GFX12-NEXT: v_dual_mov_b32 v19, s8 :: v_dual_mov_b32 v18, s9 |
| ; GFX12-NEXT: v_dual_mov_b32 v21, s6 :: v_dual_mov_b32 v20, s7 |
| ; GFX12-NEXT: v_dual_mov_b32 v23, s4 :: v_dual_mov_b32 v22, s5 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:80 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:64 |
| ; GFX12-NEXT: global_store_b128 v24, v[8:11], s[0:1] offset:48 |
| ; GFX12-NEXT: global_store_b128 v24, v[12:15], s[0:1] offset:32 |
| ; GFX12-NEXT: global_store_b128 v24, v[16:19], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v24, v[20:23], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <64 x i1>, ptr addrspace(4) %in |
| %ext = sext <64 x i1> %load to <64 x i32> |
| store <64 x i32> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_i1_to_i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_i1_to_i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_and_b32_e32 v0, 1, v0 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_i1_to_i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, 0 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_and_b32_e32 v2, 1, v2 |
| ; GFX8-NEXT: flat_store_dwordx2 v[0:1], v[2:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_i1_to_i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 2, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: MOV * T0.Y, 0.0, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_i1_to_i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_and_b32 s2, s2, 1 |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2 |
| ; GFX12-NEXT: global_store_b64 v1, v[0:1], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %a = load i1, ptr addrspace(4) %in |
| %ext = zext i1 %a to i64 |
| store i64 %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_i1_to_i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_i1_to_i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_i1_to_i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX8-NEXT: flat_store_dwordx2 v[0:1], v[2:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_i1_to_i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 3, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_INT T0.X, T0.X, 0.0, 1, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: MOV * T0.Y, PV.X, |
| ; |
| ; GFX12-LABEL: constant_sextload_i1_to_i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x10000 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, 0 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3 |
| ; GFX12-NEXT: global_store_b64 v2, v[0:1], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %a = load i1, ptr addrspace(4) %in |
| %ext = sext i1 %a to i64 |
| store i64 %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v1i1_to_v1i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v1i1_to_v1i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_and_b32_e32 v0, 1, v0 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v1i1_to_v1i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, 0 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_and_b32_e32 v2, 1, v2 |
| ; GFX8-NEXT: flat_store_dwordx2 v[0:1], v[2:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v1i1_to_v1i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 2, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: MOV * T0.Y, 0.0, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v1i1_to_v1i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_and_b32 s2, s2, 1 |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2 |
| ; GFX12-NEXT: global_store_b64 v1, v[0:1], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <1 x i1>, ptr addrspace(4) %in |
| %ext = zext <1 x i1> %load to <1 x i64> |
| store <1 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v1i1_to_v1i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v1i1_to_v1i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v1i1_to_v1i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v2, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX8-NEXT: flat_store_dwordx2 v[0:1], v[2:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v1i1_to_v1i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 3, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_INT T0.X, T0.X, 0.0, 1, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: MOV * T0.Y, PV.X, |
| ; |
| ; GFX12-LABEL: constant_sextload_v1i1_to_v1i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_u8 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x10000 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, 0 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3 |
| ; GFX12-NEXT: global_store_b64 v2, v[0:1], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <1 x i1>, ptr addrspace(4) %in |
| %ext = sext <1 x i1> %load to <1 x i64> |
| store <1 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v2i1_to_v2i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v2i1_to_v2i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v2, 1, v0 |
| ; GFX6-NEXT: v_and_b32_e32 v0, 1, v0 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, v1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v2i1_to_v2i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, 1 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v0, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, v1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 |
| ; GFX8-NEXT: v_and_b32_e32 v0, 1, v0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v2i1_to_v2i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 5, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T1.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_UINT * T0.Z, T0.X, 1, 1, |
| ; EG-NEXT: AND_INT T0.X, T0.X, 1, |
| ; EG-NEXT: MOV T0.Y, 0.0, |
| ; EG-NEXT: MOV T0.W, 0.0, |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v2i1_to_v2i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v0, v1, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v0 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_and_b32 v0, 1, v0 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v2, 1, v2 |
| ; GFX12-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <2 x i1>, ptr addrspace(4) %in |
| %ext = zext <2 x i1> %load to <2 x i64> |
| store <2 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v2i1_to_v2i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v2i1_to_v2i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v2, 1, v0 |
| ; GFX6-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX6-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v2i1_to_v2i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v0, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_lshrrev_b32_e32 v2, 1, v0 |
| ; GFX8-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX8-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v2i1_to_v2i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 6, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T1.XYZW, T0.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_INT T1.X, T0.X, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T0.X, 1, |
| ; EG-NEXT: BFE_INT * T1.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: MOV * T1.Y, T1.X, |
| ; EG-NEXT: LSHR T0.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: MOV * T1.W, T1.Z, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v2i1_to_v2i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v4, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v0, v4, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v1, 1, v0 |
| ; GFX12-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) |
| ; GFX12-NEXT: v_bfe_i32 v2, v1, 0, 1 |
| ; GFX12-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX12-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <2 x i1>, ptr addrspace(4) %in |
| %ext = sext <2 x i1> %load to <2 x i64> |
| store <2 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v3i1_to_v3i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v3i1_to_v3i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v4, off, s[8:11], 0 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, 0 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, v5 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, v5 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_and_b32_e32 v0, 1, v4 |
| ; GFX6-NEXT: v_bfe_u32 v2, v4, 1, 1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v4, 2, v4 |
| ; GFX6-NEXT: buffer_store_dwordx2 v[4:5], off, s[4:7], 0 offset:16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v3i1_to_v3i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: v_mov_b32_e32 v10, 2 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, v5 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v4, v[0:1] |
| ; GFX8-NEXT: s_add_u32 s2, s0, 16 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v9, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v7, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, v5 |
| ; GFX8-NEXT: v_mov_b32_e32 v6, s0 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_and_b32_e32 v0, 1, v4 |
| ; GFX8-NEXT: v_bfe_u32 v2, v4, 1, 1 |
| ; GFX8-NEXT: v_lshrrev_b32_sdwa v4, v10, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 |
| ; GFX8-NEXT: flat_store_dwordx2 v[8:9], v[4:5] |
| ; GFX8-NEXT: flat_store_dwordx4 v[6:7], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v3i1_to_v3i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 11, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T3.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T1.XYZW, T2.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_UINT * T1.Z, T0.X, 1, 1, |
| ; EG-NEXT: AND_INT T1.X, T0.X, 1, |
| ; EG-NEXT: MOV T1.Y, 0.0, |
| ; EG-NEXT: LSHR * T0.X, T0.X, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: MOV T0.Y, 0.0, |
| ; EG-NEXT: MOV * T1.W, 0.0, |
| ; EG-NEXT: LSHR T2.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44) |
| ; EG-NEXT: LSHR * T3.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v3i1_to_v3i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v5, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v0, v5, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_and_b32_e32 v1, 0xffff, v0 |
| ; GFX12-NEXT: v_bfe_u32 v2, v0, 1, 1 |
| ; GFX12-NEXT: v_and_b32_e32 v0, 1, v0 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v4, 2, v1 |
| ; GFX12-NEXT: v_mov_b32_e32 v3, v5 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, v5 :: v_dual_and_b32 v2, 0xffff, v2 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX12-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX12-NEXT: v_and_b32_e32 v4, 0xffff, v4 |
| ; GFX12-NEXT: s_clause 0x1 |
| ; GFX12-NEXT: global_store_b64 v5, v[4:5], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v5, v[0:3], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <3 x i1>, ptr addrspace(4) %in |
| %ext = zext <3 x i1> %load to <3 x i64> |
| store <3 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v3i1_to_v3i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v3i1_to_v3i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v3, 2, v0 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v2, 1, v0 |
| ; GFX6-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX6-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v4, v3, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v5, 31, v4 |
| ; GFX6-NEXT: buffer_store_dwordx2 v[4:5], off, s[4:7], 0 offset:16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v3i1_to_v3i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v0, v[0:1] |
| ; GFX8-NEXT: s_add_u32 s2, s0, 16 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v7, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v6, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_lshrrev_b32_e32 v3, 2, v0 |
| ; GFX8-NEXT: v_lshrrev_b32_e32 v2, 1, v0 |
| ; GFX8-NEXT: v_bfe_i32 v8, v3, 0, 1 |
| ; GFX8-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX8-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v9, 31, v8 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX8-NEXT: flat_store_dwordx2 v[6:7], v[8:9] |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v3i1_to_v3i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 14, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T3.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T1.XYZW, T2.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_INT T1.X, T0.X, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T0.X, 1, |
| ; EG-NEXT: LSHR * T2.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T1.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T0.X, PV.W, 0.0, 1, |
| ; EG-NEXT: MOV T1.Y, T1.X, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T3.X, PV.W, literal.x, |
| ; EG-NEXT: MOV T0.Y, PV.X, |
| ; EG-NEXT: MOV * T1.W, T1.Z, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v3i1_to_v3i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v6, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v0, v6, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v1, 2, v0 |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v2, 1, v0 |
| ; GFX12-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX12-NEXT: v_bfe_i32 v4, v1, 0, 1 |
| ; GFX12-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) |
| ; GFX12-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX12-NEXT: v_ashrrev_i32_e32 v5, 31, v4 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) |
| ; GFX12-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX12-NEXT: s_clause 0x1 |
| ; GFX12-NEXT: global_store_b64 v6, v[4:5], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v6, v[0:3], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <3 x i1>, ptr addrspace(4) %in |
| %ext = sext <3 x i1> %load to <3 x i64> |
| store <3 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v4i1_to_v4i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v4i1_to_v4i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v7, v1 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_and_b32_e32 v4, 1, v0 |
| ; GFX6-NEXT: v_bfe_u32 v6, v0, 1, 1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v2, 3, v0 |
| ; GFX6-NEXT: v_bfe_u32 v0, v0, 2, 1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 offset:16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[4:7], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v4i1_to_v4i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, 3 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v0, v[0:1] |
| ; GFX8-NEXT: s_add_u32 s2, s0, 16 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v11, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v9, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v10, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v7, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, s0 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_and_b32_e32 v4, 1, v0 |
| ; GFX8-NEXT: v_bfe_u32 v6, v0, 1, 1 |
| ; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 |
| ; GFX8-NEXT: v_bfe_u32 v0, v0, 2, 1 |
| ; GFX8-NEXT: flat_store_dwordx4 v[10:11], v[0:3] |
| ; GFX8-NEXT: flat_store_dwordx4 v[8:9], v[4:7] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v4i1_to_v4i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 14, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T1.XYZW, T3.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T2.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: BFE_UINT * T1.Z, T0.X, literal.x, 1, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T1.X, T0.X, literal.x, 1, |
| ; EG-NEXT: MOV T1.Y, 0.0, |
| ; EG-NEXT: BFE_UINT T0.Z, T0.X, 1, 1, |
| ; EG-NEXT: AND_INT * T0.X, T0.X, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: MOV T0.Y, 0.0, |
| ; EG-NEXT: MOV T1.W, 0.0, |
| ; EG-NEXT: MOV * T0.W, 0.0, |
| ; EG-NEXT: LSHR T2.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: ADD_INT * T2.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44) |
| ; EG-NEXT: LSHR * T3.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v4i1_to_v4i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v0, v1, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_readfirstlane_b32 s2, v0 |
| ; GFX12-NEXT: v_and_b32_e32 v0, 0xffff, v0 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10002 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v2, 3, v0 |
| ; GFX12-NEXT: s_and_b32 s3, 0xffff, s3 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v0, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10001 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) |
| ; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2 |
| ; GFX12-NEXT: s_and_b32 s2, s2, 1 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: s_and_b32 s3, 0xffff, s3 |
| ; GFX12-NEXT: s_and_b32 s2, 0xffff, s2 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:16 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <4 x i1>, ptr addrspace(4) %in |
| %ext = zext <4 x i1> %load to <4 x i64> |
| store <4 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v4i1_to_v4i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v4i1_to_v4i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s7, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s6, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s6 |
| ; GFX6-NEXT: s_mov_b32 s11, s7 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s2 |
| ; GFX6-NEXT: s_mov_b32 s9, s3 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s4, s0 |
| ; GFX6-NEXT: s_mov_b32 s5, s1 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v3, 2, v0 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v4, 3, v0 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v2, 1, v0 |
| ; GFX6-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX6-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v6, v4, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v4, v3, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v7, 31, v6 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v5, 31, v4 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[4:7], off, s[4:7], 0 offset:16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v4i1_to_v4i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v0, v[0:1] |
| ; GFX8-NEXT: s_add_u32 s2, s0, 16 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v11, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v9, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v10, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, s0 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_lshrrev_b32_e32 v3, 2, v0 |
| ; GFX8-NEXT: v_lshrrev_b32_e32 v4, 3, v0 |
| ; GFX8-NEXT: v_lshrrev_b32_e32 v2, 1, v0 |
| ; GFX8-NEXT: v_bfe_i32 v6, v4, 0, 1 |
| ; GFX8-NEXT: v_bfe_i32 v4, v3, 0, 1 |
| ; GFX8-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX8-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v7, 31, v6 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v5, 31, v4 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX8-NEXT: flat_store_dwordx4 v[10:11], v[4:7] |
| ; GFX8-NEXT: flat_store_dwordx4 v[8:9], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v4i1_to_v4i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @6 |
| ; EG-NEXT: ALU 17, @9, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T2.XYZW, T3.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T1.XYZW, T0.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 6: |
| ; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 8: |
| ; EG-NEXT: MOV * T0.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 9: |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.x, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T1.X, T0.X, 0.0, 1, |
| ; EG-NEXT: BFE_INT T2.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T2.X, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T0.X, 1, |
| ; EG-NEXT: MOV T2.Y, PV.X, |
| ; EG-NEXT: BFE_INT * T1.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: MOV T1.Y, T1.X, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44) |
| ; EG-NEXT: LSHR T3.X, PV.W, literal.x, |
| ; EG-NEXT: MOV T1.W, T1.Z, |
| ; EG-NEXT: MOV * T2.W, T2.Z, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v4i1_to_v4i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v8, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v0, v8, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v1, 3, v0 |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v2, 2, v0 |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v3, 1, v0 |
| ; GFX12-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX12-NEXT: v_bfe_i32 v6, v1, 0, 1 |
| ; GFX12-NEXT: v_bfe_i32 v4, v2, 0, 1 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX12-NEXT: v_bfe_i32 v2, v3, 0, 1 |
| ; GFX12-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) |
| ; GFX12-NEXT: v_ashrrev_i32_e32 v7, 31, v6 |
| ; GFX12-NEXT: v_ashrrev_i32_e32 v5, 31, v4 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX12-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX12-NEXT: s_clause 0x1 |
| ; GFX12-NEXT: global_store_b128 v8, v[4:7], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v8, v[0:3], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <4 x i1>, ptr addrspace(4) %in |
| %ext = sext <4 x i1> %load to <4 x i64> |
| store <4 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v8i1_to_v8i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v8i1_to_v8i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s2 |
| ; GFX6-NEXT: s_mov_b32 s11, s3 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s6 |
| ; GFX6-NEXT: s_mov_b32 s9, s7 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v7, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v11, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v13, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v15, v1 |
| ; GFX6-NEXT: s_mov_b32 s0, s4 |
| ; GFX6-NEXT: s_mov_b32 s1, s5 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_u32 v14, v0, 1, 1 |
| ; GFX6-NEXT: v_bfe_u32 v10, v0, 3, 1 |
| ; GFX6-NEXT: v_bfe_u32 v6, v0, 5, 1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v2, 7, v0 |
| ; GFX6-NEXT: v_and_b32_e32 v12, 1, v0 |
| ; GFX6-NEXT: v_bfe_u32 v8, v0, 2, 1 |
| ; GFX6-NEXT: v_bfe_u32 v4, v0, 4, 1 |
| ; GFX6-NEXT: v_bfe_u32 v0, v0, 6, 1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v8i1_to_v8i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v0, v[0:1] |
| ; GFX8-NEXT: s_add_u32 s2, s0, 48 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: s_add_u32 s4, s0, 32 |
| ; GFX8-NEXT: s_addc_u32 s5, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v16, s5 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v7, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v15, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v10, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v12, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v14, v1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_bfe_u32 v6, v0, 5, 1 |
| ; GFX8-NEXT: v_bfe_u32 v4, v0, 4, 1 |
| ; GFX8-NEXT: flat_store_dwordx4 v[15:16], v[4:7] |
| ; GFX8-NEXT: v_mov_b32_e32 v16, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s0, 16 |
| ; GFX8-NEXT: s_addc_u32 s1, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v18, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v17, s0 |
| ; GFX8-NEXT: v_and_b32_e32 v6, 0xffff, v0 |
| ; GFX8-NEXT: v_bfe_u32 v9, v0, 3, 1 |
| ; GFX8-NEXT: v_bfe_u32 v7, v0, 2, 1 |
| ; GFX8-NEXT: v_mov_b32_e32 v15, s2 |
| ; GFX8-NEXT: v_bfe_u32 v13, v0, 1, 1 |
| ; GFX8-NEXT: v_and_b32_e32 v11, 1, v0 |
| ; GFX8-NEXT: v_lshrrev_b32_e32 v2, 7, v6 |
| ; GFX8-NEXT: v_bfe_u32 v0, v6, 6, 1 |
| ; GFX8-NEXT: flat_store_dwordx4 v[17:18], v[7:10] |
| ; GFX8-NEXT: flat_store_dwordx4 v[15:16], v[0:3] |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[11:14] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v8i1_to_v8i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @10, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @8 |
| ; EG-NEXT: ALU 30, @11, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T6.XYZW, T12.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T7.XYZW, T11.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T8.XYZW, T10.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.XYZW, T9.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 8: |
| ; EG-NEXT: VTX_READ_8 T5.X, T5.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 10: |
| ; EG-NEXT: MOV * T5.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 11: |
| ; EG-NEXT: BFE_UINT * T6.Z, T5.X, literal.x, 1, |
| ; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T6.X, T5.X, literal.x, 1, |
| ; EG-NEXT: MOV T6.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T7.Z, T5.X, literal.y, 1, |
| ; EG-NEXT: 6(8.407791e-45), 5(7.006492e-45) |
| ; EG-NEXT: BFE_UINT T7.X, T5.X, literal.x, 1, |
| ; EG-NEXT: MOV T7.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T8.Z, T5.X, literal.y, 1, |
| ; EG-NEXT: 4(5.605194e-45), 3(4.203895e-45) |
| ; EG-NEXT: BFE_UINT T8.X, T5.X, literal.x, 1, |
| ; EG-NEXT: MOV T8.Y, 0.0, |
| ; EG-NEXT: BFE_UINT T5.Z, T5.X, 1, 1, |
| ; EG-NEXT: AND_INT * T5.X, T5.X, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: MOV T5.Y, 0.0, |
| ; EG-NEXT: MOV T6.W, 0.0, |
| ; EG-NEXT: MOV * T7.W, 0.0, |
| ; EG-NEXT: MOV T8.W, 0.0, |
| ; EG-NEXT: MOV * T5.W, 0.0, |
| ; EG-NEXT: LSHR T9.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44) |
| ; EG-NEXT: LSHR T10.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44) |
| ; EG-NEXT: LSHR T11.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44) |
| ; EG-NEXT: LSHR * T12.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v8i1_to_v8i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v12, v1, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_and_b32 v0, 0xffff, v12 |
| ; GFX12-NEXT: v_mov_b32_e32 v5, v1 |
| ; GFX12-NEXT: v_mov_b32_e32 v7, v1 |
| ; GFX12-NEXT: v_bfe_u32 v6, v12, 5, 1 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v2, 7, v0 |
| ; GFX12-NEXT: v_bfe_u32 v0, v0, 6, 1 |
| ; GFX12-NEXT: v_bfe_u32 v4, v12, 4, 1 |
| ; GFX12-NEXT: v_mov_b32_e32 v9, v1 |
| ; GFX12-NEXT: v_mov_b32_e32 v11, v1 |
| ; GFX12-NEXT: v_bfe_u32 v10, v12, 3, 1 |
| ; GFX12-NEXT: v_bfe_u32 v8, v12, 2, 1 |
| ; GFX12-NEXT: v_mov_b32_e32 v13, v1 |
| ; GFX12-NEXT: v_mov_b32_e32 v15, v1 |
| ; GFX12-NEXT: v_bfe_u32 v14, v12, 1, 1 |
| ; GFX12-NEXT: v_and_b32_e32 v12, 1, v12 |
| ; GFX12-NEXT: s_clause 0x3 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:48 |
| ; GFX12-NEXT: global_store_b128 v1, v[4:7], s[0:1] offset:32 |
| ; GFX12-NEXT: global_store_b128 v1, v[8:11], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v1, v[12:15], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <8 x i1>, ptr addrspace(4) %in |
| %ext = zext <8 x i1> %load to <8 x i64> |
| store <8 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v8i1_to_v8i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v8i1_to_v8i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s2 |
| ; GFX6-NEXT: s_mov_b32 s11, s3 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s6 |
| ; GFX6-NEXT: s_mov_b32 s9, s7 |
| ; GFX6-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s0, s4 |
| ; GFX6-NEXT: s_mov_b32 s1, s5 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v3, 6, v0 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v5, 7, v0 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v7, 4, v0 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v8, 5, v0 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v4, 2, v0 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v6, 3, v0 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v2, 1, v0 |
| ; GFX6-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX6-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v6, v6, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v4, v4, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v10, v8, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v8, v7, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v14, v5, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v12, v3, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v7, 31, v6 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v5, 31, v4 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v11, 31, v10 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v9, 31, v8 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v15, 31, v14 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v13, 31, v12 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v8i1_to_v8i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ubyte v0, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v17, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v16, s0 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_readfirstlane_b32 s3, v0 |
| ; GFX8-NEXT: s_lshr_b32 s2, s3, 6 |
| ; GFX8-NEXT: s_lshr_b32 s4, s3, 7 |
| ; GFX8-NEXT: s_lshr_b32 s6, s3, 4 |
| ; GFX8-NEXT: s_lshr_b32 s8, s3, 5 |
| ; GFX8-NEXT: s_lshr_b32 s10, s3, 2 |
| ; GFX8-NEXT: s_lshr_b32 s12, s3, 3 |
| ; GFX8-NEXT: s_lshr_b32 s14, s3, 1 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s3 |
| ; GFX8-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x10000 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 48 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v19, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v18, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 32 |
| ; GFX8-NEXT: v_mov_b32_e32 v6, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v7, s5 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[18:19], v[4:7] |
| ; GFX8-NEXT: s_add_u32 s0, s0, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v9, s7 |
| ; GFX8-NEXT: v_mov_b32_e32 v10, s8 |
| ; GFX8-NEXT: v_mov_b32_e32 v11, s9 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_addc_u32 s1, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[8:11] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX8-NEXT: v_mov_b32_e32 v12, s10 |
| ; GFX8-NEXT: v_mov_b32_e32 v13, s11 |
| ; GFX8-NEXT: v_mov_b32_e32 v14, s12 |
| ; GFX8-NEXT: v_mov_b32_e32 v15, s13 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s14 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s15 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[12:15] |
| ; GFX8-NEXT: flat_store_dwordx4 v[16:17], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v8i1_to_v8i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @10, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @8 |
| ; EG-NEXT: ALU 37, @11, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T8.XYZW, T12.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.XYZW, T11.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T9.XYZW, T10.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T7.XYZW, T6.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 8: |
| ; EG-NEXT: VTX_READ_8 T5.X, T5.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 10: |
| ; EG-NEXT: MOV * T5.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 11: |
| ; EG-NEXT: LSHR T6.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: LSHR * T0.W, T5.X, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 7(9.809089e-45) |
| ; EG-NEXT: BFE_INT T7.X, T5.X, 0.0, 1, |
| ; EG-NEXT: BFE_INT T8.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T5.X, literal.x, |
| ; EG-NEXT: LSHR * T1.W, T5.X, literal.y, |
| ; EG-NEXT: 3(4.203895e-45), 6(8.407791e-45) |
| ; EG-NEXT: BFE_INT T8.X, PS, 0.0, 1, |
| ; EG-NEXT: BFE_INT T9.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T5.X, 1, |
| ; EG-NEXT: LSHR * T1.W, T5.X, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T9.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T8.Y, PV.X, |
| ; EG-NEXT: BFE_INT T7.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T5.X, literal.x, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 5(7.006492e-45), 16(2.242078e-44) |
| ; EG-NEXT: LSHR T10.X, PS, literal.x, |
| ; EG-NEXT: MOV T9.Y, PV.X, |
| ; EG-NEXT: BFE_INT T5.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T0.W, T5.X, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 4(5.605194e-45) |
| ; EG-NEXT: BFE_INT T5.X, PV.W, 0.0, 1, |
| ; EG-NEXT: MOV T7.Y, T7.X, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 32(4.484155e-44), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T11.X, PV.W, literal.x, |
| ; EG-NEXT: MOV T5.Y, PV.X, |
| ; EG-NEXT: ADD_INT T0.Z, KC0[2].Y, literal.y, |
| ; EG-NEXT: MOV T7.W, T7.Z, |
| ; EG-NEXT: MOV * T9.W, T9.Z, |
| ; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44) |
| ; EG-NEXT: LSHR T12.X, PV.Z, literal.x, |
| ; EG-NEXT: MOV T5.W, T5.Z, |
| ; EG-NEXT: MOV * T8.W, T8.Z, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v8i1_to_v8i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v16, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u8 v0, v16, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_readfirstlane_b32 s3, v0 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) |
| ; GFX12-NEXT: v_mov_b32_e32 v9, s3 |
| ; GFX12-NEXT: s_lshr_b32 s2, s3, 6 |
| ; GFX12-NEXT: s_lshr_b32 s4, s3, 7 |
| ; GFX12-NEXT: s_lshr_b32 s6, s3, 4 |
| ; GFX12-NEXT: s_lshr_b32 s8, s3, 5 |
| ; GFX12-NEXT: s_lshr_b32 s10, s3, 2 |
| ; GFX12-NEXT: s_lshr_b32 s12, s3, 3 |
| ; GFX12-NEXT: s_lshr_b32 s14, s3, 1 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x10000 |
| ; GFX12-NEXT: v_bfe_i32 v12, v9, 0, 1 |
| ; GFX12-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x10000 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3 |
| ; GFX12-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5 |
| ; GFX12-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v4, s6 :: v_dual_mov_b32 v5, s7 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s8 :: v_dual_mov_b32 v7, s9 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s10 :: v_dual_mov_b32 v9, s11 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s12 :: v_dual_mov_b32 v11, s13 |
| ; GFX12-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15 |
| ; GFX12-NEXT: v_ashrrev_i32_e32 v13, 31, v12 |
| ; GFX12-NEXT: s_clause 0x3 |
| ; GFX12-NEXT: global_store_b128 v16, v[0:3], s[0:1] offset:48 |
| ; GFX12-NEXT: global_store_b128 v16, v[4:7], s[0:1] offset:32 |
| ; GFX12-NEXT: global_store_b128 v16, v[8:11], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v16, v[12:15], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <8 x i1>, ptr addrspace(4) %in |
| %ext = sext <8 x i1> %load to <8 x i64> |
| store <8 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v16i1_to_v16i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v16i1_to_v16i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s2 |
| ; GFX6-NEXT: s_mov_b32 s11, s3 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s6 |
| ; GFX6-NEXT: s_mov_b32 s9, s7 |
| ; GFX6-NEXT: buffer_load_ushort v29, off, s[8:11], 0 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v6, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v7, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v10, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v12, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v14, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v16, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v18, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v20, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v22, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v24, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v26, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v28, v1 |
| ; GFX6-NEXT: s_mov_b32 s0, s4 |
| ; GFX6-NEXT: s_mov_b32 s1, s5 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_bfe_u32 v2, v29, 11, 1 |
| ; GFX6-NEXT: v_bfe_u32 v0, v29, 10, 1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80 |
| ; GFX6-NEXT: v_bfe_u32 v5, v29, 9, 1 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_bfe_u32 v3, v29, 8, 1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[3:6], off, s[0:3], 0 offset:64 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v8, 15, v29 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_bfe_u32 v6, v29, 14, 1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[6:9], off, s[0:3], 0 offset:112 |
| ; GFX6-NEXT: v_bfe_u32 v27, v29, 5, 1 |
| ; GFX6-NEXT: v_bfe_u32 v23, v29, 7, 1 |
| ; GFX6-NEXT: v_bfe_u32 v19, v29, 1, 1 |
| ; GFX6-NEXT: v_bfe_u32 v15, v29, 3, 1 |
| ; GFX6-NEXT: v_bfe_u32 v11, v29, 13, 1 |
| ; GFX6-NEXT: v_bfe_u32 v25, v29, 4, 1 |
| ; GFX6-NEXT: v_bfe_u32 v21, v29, 6, 1 |
| ; GFX6-NEXT: v_and_b32_e32 v17, 1, v29 |
| ; GFX6-NEXT: v_bfe_u32 v13, v29, 2, 1 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_bfe_u32 v9, v29, 12, 1 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[9:12], off, s[0:3], 0 offset:96 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[13:16], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[17:20], off, s[0:3], 0 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[21:24], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[25:28], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v16i1_to_v16i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ushort v0, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v7, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v9, v1 |
| ; GFX8-NEXT: v_mov_b32_e32 v11, v1 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_readfirstlane_b32 s2, v0 |
| ; GFX8-NEXT: s_bfe_u32 s3, s2, 0x10009 |
| ; GFX8-NEXT: s_bfe_u32 s4, s2, 0x1000d |
| ; GFX8-NEXT: s_bfe_u32 s5, s2, 0x10007 |
| ; GFX8-NEXT: s_bfe_u32 s6, s2, 0x10003 |
| ; GFX8-NEXT: s_bfe_u32 s7, s2, 0x10001 |
| ; GFX8-NEXT: s_and_b32 s8, s2, 1 |
| ; GFX8-NEXT: s_bfe_u32 s9, s2, 0x10002 |
| ; GFX8-NEXT: s_bfe_u32 s10, s2, 0x10004 |
| ; GFX8-NEXT: s_bfe_u32 s11, s2, 0x10006 |
| ; GFX8-NEXT: s_bfe_u32 s12, s2, 0x1000c |
| ; GFX8-NEXT: s_bfe_u32 s2, s2, 0x1000a |
| ; GFX8-NEXT: v_and_b32_e32 v4, 0xffff, v0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 0x50 |
| ; GFX8-NEXT: v_mov_b32_e32 v6, s3 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v13, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v12, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 64 |
| ; GFX8-NEXT: v_bfe_u32 v2, v4, 11, 1 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[12:13], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v13, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v12, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 0x70 |
| ; GFX8-NEXT: v_lshrrev_b32_e32 v10, 15, v4 |
| ; GFX8-NEXT: v_bfe_u32 v14, v4, 5, 1 |
| ; GFX8-NEXT: v_bfe_u32 v8, v4, 14, 1 |
| ; GFX8-NEXT: v_bfe_u32 v4, v4, 8, 1 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[12:13], v[4:7] |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s12 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 0x60 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[8:11] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 48 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 32 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s11 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s5 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s10 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, v14 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s9 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s8 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s7 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v16i1_to_v16i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @14, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @12 |
| ; EG-NEXT: ALU 62, @15, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T8.XYZW, T22.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T9.XYZW, T21.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T10.XYZW, T20.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T11.XYZW, T19.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T12.XYZW, T18.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T13.XYZW, T17.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T14.XYZW, T16.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T7.XYZW, T15.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 12: |
| ; EG-NEXT: VTX_READ_16 T7.X, T7.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 14: |
| ; EG-NEXT: MOV * T7.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 15: |
| ; EG-NEXT: LSHR * T8.Z, T7.X, literal.x, |
| ; EG-NEXT: 15(2.101948e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T8.X, T7.X, literal.x, 1, |
| ; EG-NEXT: MOV T8.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T9.Z, T7.X, literal.y, 1, |
| ; EG-NEXT: 14(1.961818e-44), 13(1.821688e-44) |
| ; EG-NEXT: BFE_UINT T9.X, T7.X, literal.x, 1, |
| ; EG-NEXT: MOV T9.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T10.Z, T7.X, literal.y, 1, |
| ; EG-NEXT: 12(1.681558e-44), 11(1.541428e-44) |
| ; EG-NEXT: BFE_UINT T10.X, T7.X, literal.x, 1, |
| ; EG-NEXT: MOV T10.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T11.Z, T7.X, literal.y, 1, |
| ; EG-NEXT: 10(1.401298e-44), 9(1.261169e-44) |
| ; EG-NEXT: BFE_UINT T11.X, T7.X, literal.x, 1, |
| ; EG-NEXT: MOV T11.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T12.Z, T7.X, literal.y, 1, |
| ; EG-NEXT: 8(1.121039e-44), 7(9.809089e-45) |
| ; EG-NEXT: BFE_UINT T12.X, T7.X, literal.x, 1, |
| ; EG-NEXT: MOV T12.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T13.Z, T7.X, literal.y, 1, |
| ; EG-NEXT: 6(8.407791e-45), 5(7.006492e-45) |
| ; EG-NEXT: BFE_UINT T13.X, T7.X, literal.x, 1, |
| ; EG-NEXT: MOV T13.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T14.Z, T7.X, literal.y, 1, |
| ; EG-NEXT: 4(5.605194e-45), 3(4.203895e-45) |
| ; EG-NEXT: BFE_UINT T14.X, T7.X, literal.x, 1, |
| ; EG-NEXT: MOV T14.Y, 0.0, |
| ; EG-NEXT: BFE_UINT T7.Z, T7.X, 1, 1, |
| ; EG-NEXT: AND_INT * T7.X, T7.X, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: MOV T7.Y, 0.0, |
| ; EG-NEXT: MOV T8.W, 0.0, |
| ; EG-NEXT: MOV * T9.W, 0.0, |
| ; EG-NEXT: MOV T10.W, 0.0, |
| ; EG-NEXT: MOV * T11.W, 0.0, |
| ; EG-NEXT: MOV T12.W, 0.0, |
| ; EG-NEXT: MOV * T13.W, 0.0, |
| ; EG-NEXT: MOV T14.W, 0.0, |
| ; EG-NEXT: MOV * T7.W, 0.0, |
| ; EG-NEXT: LSHR T15.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44) |
| ; EG-NEXT: LSHR T16.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44) |
| ; EG-NEXT: LSHR T17.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44) |
| ; EG-NEXT: LSHR T18.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 64(8.968310e-44) |
| ; EG-NEXT: LSHR T19.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 80(1.121039e-43) |
| ; EG-NEXT: LSHR T20.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 96(1.345247e-43) |
| ; EG-NEXT: LSHR T21.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 112(1.569454e-43) |
| ; EG-NEXT: LSHR * T22.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v16i1_to_v16i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u16 v0, v1, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_and_b32_e32 v4, 0xffff, v0 |
| ; GFX12-NEXT: v_readfirstlane_b32 s2, v0 |
| ; GFX12-NEXT: v_mov_b32_e32 v7, v1 |
| ; GFX12-NEXT: v_mov_b32_e32 v11, v1 |
| ; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_bfe_u32 v2, v4, 11, 1 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1000a |
| ; GFX12-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v0, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1000d |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x1000c |
| ; GFX12-NEXT: v_mov_b32_e32 v5, v1 |
| ; GFX12-NEXT: v_bfe_u32 v6, v4, 5, 1 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:80 |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10007 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10006 |
| ; GFX12-NEXT: v_mov_b32_e32 v9, v1 |
| ; GFX12-NEXT: s_bfe_u32 s6, s2, 0x10002 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:96 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10004 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10009 |
| ; GFX12-NEXT: s_bfe_u32 s5, s2, 0x10001 |
| ; GFX12-NEXT: v_lshrrev_b32_e32 v10, 15, v4 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:48 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, v6 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10003 |
| ; GFX12-NEXT: s_and_b32 s2, s2, 1 |
| ; GFX12-NEXT: v_bfe_u32 v8, v4, 14, 1 |
| ; GFX12-NEXT: v_bfe_u32 v4, v4, 8, 1 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:32 |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s6 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v6, s3 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:16 |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s5 |
| ; GFX12-NEXT: s_clause 0x2 |
| ; GFX12-NEXT: global_store_b128 v1, v[8:11], s[0:1] offset:112 |
| ; GFX12-NEXT: global_store_b128 v1, v[4:7], s[0:1] offset:64 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <16 x i1>, ptr addrspace(4) %in |
| %ext = zext <16 x i1> %load to <16 x i64> |
| store <16 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v16i1_to_v16i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v16i1_to_v16i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: s_mov_b32 s10, s2 |
| ; GFX6-NEXT: s_mov_b32 s11, s3 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_mov_b32 s8, s6 |
| ; GFX6-NEXT: s_mov_b32 s9, s7 |
| ; GFX6-NEXT: buffer_load_ushort v1, off, s[8:11], 0 |
| ; GFX6-NEXT: s_mov_b32 s0, s4 |
| ; GFX6-NEXT: s_mov_b32 s1, s5 |
| ; GFX6-NEXT: s_waitcnt vmcnt(0) |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v3, 14, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v4, 15, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v7, 12, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v8, 13, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v11, 10, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v12, 11, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v14, 8, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v16, 9, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v15, 6, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v9, 4, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v10, 5, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v0, 2, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v2, 3, v1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v13, 1, v1 |
| ; GFX6-NEXT: v_bfe_i32 v2, v2, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v5, v4, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v3, v3, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v6, 31, v5 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v4, 31, v3 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[3:6], off, s[0:3], 0 offset:112 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_bfe_i32 v6, v10, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v4, v9, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v9, v8, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v7, v7, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v10, 31, v9 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v8, 31, v7 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[7:10], off, s[0:3], 0 offset:96 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_bfe_i32 v9, v12, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v7, v11, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v13, v13, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v11, v1, 0, 1 |
| ; GFX6-NEXT: v_lshrrev_b32_e32 v1, 7, v1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v10, 31, v9 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v8, 31, v7 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[7:10], off, s[0:3], 0 offset:80 |
| ; GFX6-NEXT: v_bfe_i32 v17, v1, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v15, v15, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v21, v16, 0, 1 |
| ; GFX6-NEXT: v_bfe_i32 v19, v14, 0, 1 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v12, 31, v11 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v14, 31, v13 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v3, 31, v2 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v7, 31, v6 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v5, 31, v4 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v18, 31, v17 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v16, 31, v15 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v22, 31, v21 |
| ; GFX6-NEXT: v_ashrrev_i32_e32 v20, 31, v19 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[19:22], off, s[0:3], 0 offset:64 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[15:18], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[11:14], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v16i1_to_v16i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s3 |
| ; GFX8-NEXT: flat_load_ushort v0, v[0:1] |
| ; GFX8-NEXT: v_mov_b32_e32 v19, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v18, s0 |
| ; GFX8-NEXT: s_waitcnt vmcnt(0) |
| ; GFX8-NEXT: v_readfirstlane_b32 s3, v0 |
| ; GFX8-NEXT: s_lshr_b32 s2, s3, 14 |
| ; GFX8-NEXT: s_lshr_b32 s4, s3, 15 |
| ; GFX8-NEXT: s_lshr_b32 s6, s3, 12 |
| ; GFX8-NEXT: s_lshr_b32 s8, s3, 13 |
| ; GFX8-NEXT: s_lshr_b32 s10, s3, 10 |
| ; GFX8-NEXT: s_lshr_b32 s12, s3, 11 |
| ; GFX8-NEXT: s_lshr_b32 s14, s3, 8 |
| ; GFX8-NEXT: s_lshr_b32 s16, s3, 9 |
| ; GFX8-NEXT: s_lshr_b32 s18, s3, 6 |
| ; GFX8-NEXT: s_lshr_b32 s20, s3, 7 |
| ; GFX8-NEXT: s_lshr_b32 s22, s3, 4 |
| ; GFX8-NEXT: s_lshr_b32 s24, s3, 5 |
| ; GFX8-NEXT: s_lshr_b32 s26, s3, 2 |
| ; GFX8-NEXT: s_lshr_b32 s28, s3, 3 |
| ; GFX8-NEXT: s_lshr_b32 s30, s3, 1 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s3 |
| ; GFX8-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x10000 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 0x70 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s3 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v15, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v14, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 0x60 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s5 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[14:15], v[2:5] |
| ; GFX8-NEXT: v_mov_b32_e32 v15, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v14, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 0x50 |
| ; GFX8-NEXT: v_mov_b32_e32 v6, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v7, s7 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, s8 |
| ; GFX8-NEXT: v_mov_b32_e32 v9, s9 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[14:15], v[6:9] |
| ; GFX8-NEXT: v_mov_b32_e32 v15, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v14, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 64 |
| ; GFX8-NEXT: v_mov_b32_e32 v10, s10 |
| ; GFX8-NEXT: v_mov_b32_e32 v11, s11 |
| ; GFX8-NEXT: v_mov_b32_e32 v12, s12 |
| ; GFX8-NEXT: v_mov_b32_e32 v13, s13 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[14:15], v[10:13] |
| ; GFX8-NEXT: v_mov_b32_e32 v15, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v14, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 48 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s14 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s15 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s16 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s17 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[14:15], v[2:5] |
| ; GFX8-NEXT: v_mov_b32_e32 v6, s18 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s0, 32 |
| ; GFX8-NEXT: v_mov_b32_e32 v7, s19 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, s20 |
| ; GFX8-NEXT: v_mov_b32_e32 v9, s21 |
| ; GFX8-NEXT: s_addc_u32 s3, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[6:9] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s3 |
| ; GFX8-NEXT: s_add_u32 s0, s0, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v10, s22 |
| ; GFX8-NEXT: v_mov_b32_e32 v11, s23 |
| ; GFX8-NEXT: v_mov_b32_e32 v12, s24 |
| ; GFX8-NEXT: v_mov_b32_e32 v13, s25 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s2 |
| ; GFX8-NEXT: s_addc_u32 s1, s1, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[10:13] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_bfe_i32 v0, v0, 0, 1 |
| ; GFX8-NEXT: v_mov_b32_e32 v14, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v15, s27 |
| ; GFX8-NEXT: v_mov_b32_e32 v16, s28 |
| ; GFX8-NEXT: v_mov_b32_e32 v17, s29 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: v_ashrrev_i32_e32 v1, 31, v0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s30 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s31 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[14:17] |
| ; GFX8-NEXT: flat_store_dwordx4 v[18:19], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v16i1_to_v16i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @14, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @12 |
| ; EG-NEXT: ALU 78, @15, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T14.XYZW, T22.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T21.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T15.XYZW, T18.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T7.XYZW, T12.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T16.XYZW, T11.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T10.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T17.XYZW, T9.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T13.XYZW, T8.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 12: |
| ; EG-NEXT: VTX_READ_16 T7.X, T7.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 14: |
| ; EG-NEXT: MOV * T7.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 15: |
| ; EG-NEXT: LSHR T8.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44) |
| ; EG-NEXT: LSHR T9.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44) |
| ; EG-NEXT: LSHR T10.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44) |
| ; EG-NEXT: LSHR T11.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 64(8.968310e-44) |
| ; EG-NEXT: LSHR T12.X, PV.W, literal.x, |
| ; EG-NEXT: LSHR * T0.W, T7.X, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 15(2.101948e-44) |
| ; EG-NEXT: BFE_INT T13.X, T7.X, 0.0, 1, |
| ; EG-NEXT: BFE_INT T14.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T7.X, literal.x, |
| ; EG-NEXT: LSHR * T1.W, T7.X, literal.y, |
| ; EG-NEXT: 11(1.541428e-44), 14(1.961818e-44) |
| ; EG-NEXT: BFE_INT T14.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T0.Y, T7.X, literal.x, |
| ; EG-NEXT: BFE_INT T15.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T7.X, literal.y, |
| ; EG-NEXT: LSHR * T1.W, T7.X, literal.z, |
| ; EG-NEXT: 12(1.681558e-44), 7(9.809089e-45) |
| ; EG-NEXT: 10(1.401298e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T15.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T14.Y, PV.X, |
| ; EG-NEXT: BFE_INT T16.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T7.X, literal.x, |
| ; EG-NEXT: LSHR * T1.W, T7.X, literal.y, |
| ; EG-NEXT: 3(4.203895e-45), 6(8.407791e-45) |
| ; EG-NEXT: BFE_INT T16.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T15.Y, PV.X, |
| ; EG-NEXT: BFE_INT T17.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T7.X, 1, |
| ; EG-NEXT: LSHR * T1.W, T7.X, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T17.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T16.Y, PV.X, |
| ; EG-NEXT: BFE_INT T13.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T7.X, literal.x, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 5(7.006492e-45), 80(1.121039e-43) |
| ; EG-NEXT: LSHR T18.X, PS, literal.x, |
| ; EG-NEXT: MOV T17.Y, PV.X, |
| ; EG-NEXT: BFE_INT T19.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T7.X, literal.y, |
| ; EG-NEXT: LSHR * T1.W, T7.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 9(1.261169e-44) |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T19.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T13.Y, T13.X, |
| ; EG-NEXT: BFE_INT T7.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T0.W, T7.X, literal.x, BS:VEC_120/SCL_212 |
| ; EG-NEXT: LSHR * T1.W, T7.X, literal.y, |
| ; EG-NEXT: 13(1.821688e-44), 8(1.121039e-44) |
| ; EG-NEXT: BFE_INT T7.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T19.Y, PV.X, |
| ; EG-NEXT: BFE_INT T20.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: MOV T13.W, T13.Z, |
| ; EG-NEXT: MOV * T17.W, T17.Z, |
| ; EG-NEXT: BFE_INT T20.X, T0.Y, 0.0, 1, |
| ; EG-NEXT: MOV T7.Y, PV.X, |
| ; EG-NEXT: ADD_INT T0.Z, KC0[2].Y, literal.x, |
| ; EG-NEXT: MOV T19.W, T19.Z, |
| ; EG-NEXT: MOV * T16.W, T16.Z, |
| ; EG-NEXT: 96(1.345247e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T21.X, PV.Z, literal.x, |
| ; EG-NEXT: MOV T20.Y, PV.X, |
| ; EG-NEXT: ADD_INT T0.Z, KC0[2].Y, literal.y, |
| ; EG-NEXT: MOV T7.W, T7.Z, |
| ; EG-NEXT: MOV * T15.W, T15.Z, |
| ; EG-NEXT: 2(2.802597e-45), 112(1.569454e-43) |
| ; EG-NEXT: LSHR T22.X, PV.Z, literal.x, |
| ; EG-NEXT: MOV T20.W, T20.Z, |
| ; EG-NEXT: MOV * T14.W, T14.Z, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v16i1_to_v16i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: v_mov_b32_e32 v32, 0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: global_load_u16 v0, v32, s[2:3] |
| ; GFX12-NEXT: s_wait_loadcnt 0x0 |
| ; GFX12-NEXT: v_readfirstlane_b32 s3, v0 |
| ; GFX12-NEXT: s_lshr_b32 s4, s3, 15 |
| ; GFX12-NEXT: s_lshr_b32 s2, s3, 14 |
| ; GFX12-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x10000 |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v28, s3 :: v_dual_mov_b32 v3, s5 |
| ; GFX12-NEXT: s_lshr_b32 s6, s3, 12 |
| ; GFX12-NEXT: s_lshr_b32 s8, s3, 13 |
| ; GFX12-NEXT: s_lshr_b32 s10, s3, 10 |
| ; GFX12-NEXT: s_lshr_b32 s12, s3, 11 |
| ; GFX12-NEXT: s_lshr_b32 s14, s3, 8 |
| ; GFX12-NEXT: s_lshr_b32 s16, s3, 9 |
| ; GFX12-NEXT: s_lshr_b32 s18, s3, 6 |
| ; GFX12-NEXT: s_lshr_b32 s20, s3, 7 |
| ; GFX12-NEXT: s_lshr_b32 s22, s3, 4 |
| ; GFX12-NEXT: s_lshr_b32 s24, s3, 5 |
| ; GFX12-NEXT: s_lshr_b32 s26, s3, 2 |
| ; GFX12-NEXT: s_lshr_b32 s28, s3, 3 |
| ; GFX12-NEXT: s_lshr_b32 s30, s3, 1 |
| ; GFX12-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x10000 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v5, s7 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s3 :: v_dual_mov_b32 v2, s4 |
| ; GFX12-NEXT: v_dual_mov_b32 v7, s9 :: v_dual_mov_b32 v4, s6 |
| ; GFX12-NEXT: v_dual_mov_b32 v9, s11 :: v_dual_mov_b32 v6, s8 |
| ; GFX12-NEXT: v_dual_mov_b32 v11, s13 :: v_dual_mov_b32 v8, s10 |
| ; GFX12-NEXT: v_dual_mov_b32 v13, s15 :: v_dual_mov_b32 v10, s12 |
| ; GFX12-NEXT: v_mov_b32_e32 v15, s17 |
| ; GFX12-NEXT: v_bfe_i32 v28, v28, 0, 1 |
| ; GFX12-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s14 :: v_dual_mov_b32 v17, s19 |
| ; GFX12-NEXT: v_dual_mov_b32 v14, s16 :: v_dual_mov_b32 v19, s21 |
| ; GFX12-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v16, s18 :: v_dual_mov_b32 v21, s23 |
| ; GFX12-NEXT: v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v23, s25 |
| ; GFX12-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v20, s22 :: v_dual_mov_b32 v25, s27 |
| ; GFX12-NEXT: v_dual_mov_b32 v22, s24 :: v_dual_mov_b32 v27, s29 |
| ; GFX12-NEXT: v_dual_mov_b32 v24, s26 :: v_dual_mov_b32 v31, s31 |
| ; GFX12-NEXT: v_mov_b32_e32 v26, s28 |
| ; GFX12-NEXT: v_mov_b32_e32 v30, s30 |
| ; GFX12-NEXT: s_clause 0x1 |
| ; GFX12-NEXT: global_store_b128 v32, v[0:3], s[0:1] offset:112 |
| ; GFX12-NEXT: global_store_b128 v32, v[4:7], s[0:1] offset:96 |
| ; GFX12-NEXT: v_ashrrev_i32_e32 v29, 31, v28 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v32, v[8:11], s[0:1] offset:80 |
| ; GFX12-NEXT: global_store_b128 v32, v[12:15], s[0:1] offset:64 |
| ; GFX12-NEXT: global_store_b128 v32, v[16:19], s[0:1] offset:48 |
| ; GFX12-NEXT: global_store_b128 v32, v[20:23], s[0:1] offset:32 |
| ; GFX12-NEXT: global_store_b128 v32, v[24:27], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v32, v[28:31], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <16 x i1>, ptr addrspace(4) %in |
| %ext = sext <16 x i1> %load to <16 x i64> |
| store <16 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v32i1_to_v32i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v32i1_to_v32i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_load_dword s4, s[2:3], 0x0 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, v1 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_bfe_u32 s5, s4, 0x10001 |
| ; GFX6-NEXT: s_bfe_u32 s6, s4, 0x10003 |
| ; GFX6-NEXT: s_bfe_u32 s7, s4, 0x10005 |
| ; GFX6-NEXT: s_bfe_u32 s8, s4, 0x10007 |
| ; GFX6-NEXT: s_bfe_u32 s9, s4, 0x10009 |
| ; GFX6-NEXT: s_bfe_u32 s10, s4, 0x1000b |
| ; GFX6-NEXT: s_bfe_u32 s11, s4, 0x1000d |
| ; GFX6-NEXT: s_bfe_u32 s12, s4, 0x1000f |
| ; GFX6-NEXT: s_bfe_u32 s13, s4, 0x10011 |
| ; GFX6-NEXT: s_bfe_u32 s14, s4, 0x10013 |
| ; GFX6-NEXT: s_bfe_u32 s15, s4, 0x10015 |
| ; GFX6-NEXT: s_bfe_u32 s16, s4, 0x10017 |
| ; GFX6-NEXT: s_bfe_u32 s17, s4, 0x10019 |
| ; GFX6-NEXT: s_bfe_u32 s18, s4, 0x1001b |
| ; GFX6-NEXT: s_bfe_u32 s19, s4, 0x1001d |
| ; GFX6-NEXT: s_lshr_b32 s20, s4, 31 |
| ; GFX6-NEXT: s_and_b32 s21, s4, 1 |
| ; GFX6-NEXT: s_bfe_u32 s22, s4, 0x10002 |
| ; GFX6-NEXT: s_bfe_u32 s23, s4, 0x10004 |
| ; GFX6-NEXT: s_bfe_u32 s24, s4, 0x10006 |
| ; GFX6-NEXT: s_bfe_u32 s25, s4, 0x10008 |
| ; GFX6-NEXT: s_bfe_u32 s26, s4, 0x1000a |
| ; GFX6-NEXT: s_bfe_u32 s27, s4, 0x1000c |
| ; GFX6-NEXT: s_bfe_u32 s28, s4, 0x1000e |
| ; GFX6-NEXT: s_bfe_u32 s29, s4, 0x10010 |
| ; GFX6-NEXT: s_bfe_u32 s30, s4, 0x10012 |
| ; GFX6-NEXT: s_bfe_u32 s31, s4, 0x10014 |
| ; GFX6-NEXT: s_bfe_u32 s33, s4, 0x10016 |
| ; GFX6-NEXT: s_bfe_u32 s34, s4, 0x10018 |
| ; GFX6-NEXT: s_bfe_u32 s35, s4, 0x1001a |
| ; GFX6-NEXT: s_bfe_u32 s36, s4, 0x1001e |
| ; GFX6-NEXT: s_bfe_u32 s4, s4, 0x1001c |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s20 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:240 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s19 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:224 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s35 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:208 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s34 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s17 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:192 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s33 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s16 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:176 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s31 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s15 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:160 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s30 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s14 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:144 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s29 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s13 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:128 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s28 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s12 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:112 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s27 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s11 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:96 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s26 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s10 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s25 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s9 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s24 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s8 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s23 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s7 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s22 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s6 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s21 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s5 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v32i1_to_v32i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, v1 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_load_dword s6, s[2:3], 0x0 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_lshr_b32 s7, s6, 31 |
| ; GFX8-NEXT: s_bfe_u32 s8, s6, 0x1001d |
| ; GFX8-NEXT: s_bfe_u32 s9, s6, 0x1001b |
| ; GFX8-NEXT: s_bfe_u32 s10, s6, 0x10019 |
| ; GFX8-NEXT: s_bfe_u32 s11, s6, 0x10017 |
| ; GFX8-NEXT: s_bfe_u32 s12, s6, 0x10013 |
| ; GFX8-NEXT: s_bfe_u32 s13, s6, 0x10011 |
| ; GFX8-NEXT: s_bfe_u32 s14, s6, 0x1000f |
| ; GFX8-NEXT: s_bfe_u32 s15, s6, 0x1000d |
| ; GFX8-NEXT: s_bfe_u32 s16, s6, 0x1000b |
| ; GFX8-NEXT: s_bfe_u32 s17, s6, 0x10009 |
| ; GFX8-NEXT: s_bfe_u32 s18, s6, 0x10007 |
| ; GFX8-NEXT: s_bfe_u32 s19, s6, 0x10005 |
| ; GFX8-NEXT: s_bfe_u32 s4, s6, 0x10003 |
| ; GFX8-NEXT: s_bfe_u32 s2, s6, 0x10001 |
| ; GFX8-NEXT: s_and_b32 s3, s6, 1 |
| ; GFX8-NEXT: s_bfe_u32 s5, s6, 0x10002 |
| ; GFX8-NEXT: s_bfe_u32 s20, s6, 0x10004 |
| ; GFX8-NEXT: s_bfe_u32 s21, s6, 0x10006 |
| ; GFX8-NEXT: s_bfe_u32 s22, s6, 0x10008 |
| ; GFX8-NEXT: s_bfe_u32 s23, s6, 0x1000a |
| ; GFX8-NEXT: s_bfe_u32 s24, s6, 0x1000c |
| ; GFX8-NEXT: s_bfe_u32 s25, s6, 0x1000e |
| ; GFX8-NEXT: s_bfe_u32 s26, s6, 0x10010 |
| ; GFX8-NEXT: s_bfe_u32 s27, s6, 0x10012 |
| ; GFX8-NEXT: s_bfe_u32 s28, s6, 0x10014 |
| ; GFX8-NEXT: s_bfe_u32 s29, s6, 0x10015 |
| ; GFX8-NEXT: s_bfe_u32 s30, s6, 0x10016 |
| ; GFX8-NEXT: s_bfe_u32 s31, s6, 0x10018 |
| ; GFX8-NEXT: s_bfe_u32 s33, s6, 0x1001a |
| ; GFX8-NEXT: s_bfe_u32 s34, s6, 0x1001c |
| ; GFX8-NEXT: s_bfe_u32 s6, s6, 0x1001e |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s6 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 0xf0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s7 |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 0xe0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s34 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s8 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 0xd0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s33 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s9 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 0xc0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s31 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s10 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 0xb0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s30 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s11 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 0xa0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s28 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s29 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 0x90 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s27 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s12 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 0x80 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s13 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 0x70 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s25 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s14 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 0x60 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s24 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s15 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 0x50 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s23 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s16 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 64 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s22 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s17 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 48 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s21 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 32 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s20 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s19 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX8-NEXT: s_add_u32 s4, s0, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX8-NEXT: s_addc_u32 s5, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s5 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v32i1_to_v32i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @24, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @22 |
| ; EG-NEXT: ALU 96, @25, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: ALU 30, @122, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T12.XYZW, T42.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T13.XYZW, T41.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T14.XYZW, T40.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T15.XYZW, T39.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T16.XYZW, T38.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T17.XYZW, T37.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T18.XYZW, T36.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T35.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T34.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T21.XYZW, T33.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T22.XYZW, T32.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T23.XYZW, T31.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T24.XYZW, T30.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T25.XYZW, T29.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T26.XYZW, T28.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T11.XYZW, T27.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 22: |
| ; EG-NEXT: VTX_READ_32 T11.X, T11.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 24: |
| ; EG-NEXT: MOV * T11.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 25: |
| ; EG-NEXT: LSHR * T12.Z, T11.X, literal.x, |
| ; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T12.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T12.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T13.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 30(4.203895e-44), 29(4.063766e-44) |
| ; EG-NEXT: BFE_UINT T13.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T13.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T14.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 28(3.923636e-44), 27(3.783506e-44) |
| ; EG-NEXT: BFE_UINT T14.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T14.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T15.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 26(3.643376e-44), 25(3.503246e-44) |
| ; EG-NEXT: BFE_UINT T15.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T15.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T16.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 24(3.363116e-44), 23(3.222986e-44) |
| ; EG-NEXT: BFE_UINT T16.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T16.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T17.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 22(3.082857e-44), 21(2.942727e-44) |
| ; EG-NEXT: BFE_UINT T17.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T17.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T18.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 20(2.802597e-44), 19(2.662467e-44) |
| ; EG-NEXT: BFE_UINT T18.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T18.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T19.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 18(2.522337e-44), 17(2.382207e-44) |
| ; EG-NEXT: BFE_UINT T19.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T19.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T20.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 16(2.242078e-44), 15(2.101948e-44) |
| ; EG-NEXT: BFE_UINT T20.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T20.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T21.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 14(1.961818e-44), 13(1.821688e-44) |
| ; EG-NEXT: BFE_UINT T21.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T21.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T22.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 12(1.681558e-44), 11(1.541428e-44) |
| ; EG-NEXT: BFE_UINT T22.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T22.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T23.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 10(1.401298e-44), 9(1.261169e-44) |
| ; EG-NEXT: BFE_UINT T23.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T23.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T24.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 8(1.121039e-44), 7(9.809089e-45) |
| ; EG-NEXT: BFE_UINT T24.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T24.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T25.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 6(8.407791e-45), 5(7.006492e-45) |
| ; EG-NEXT: BFE_UINT T25.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T25.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T26.Z, T11.X, literal.y, 1, |
| ; EG-NEXT: 4(5.605194e-45), 3(4.203895e-45) |
| ; EG-NEXT: BFE_UINT T26.X, T11.X, literal.x, 1, |
| ; EG-NEXT: MOV T26.Y, 0.0, |
| ; EG-NEXT: BFE_UINT T11.Z, T11.X, 1, 1, |
| ; EG-NEXT: AND_INT * T11.X, T11.X, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: MOV T11.Y, 0.0, |
| ; EG-NEXT: MOV T12.W, 0.0, |
| ; EG-NEXT: MOV * T13.W, 0.0, |
| ; EG-NEXT: MOV T14.W, 0.0, |
| ; EG-NEXT: MOV * T15.W, 0.0, |
| ; EG-NEXT: MOV T16.W, 0.0, |
| ; EG-NEXT: MOV * T17.W, 0.0, |
| ; EG-NEXT: MOV T18.W, 0.0, |
| ; EG-NEXT: MOV * T19.W, 0.0, |
| ; EG-NEXT: MOV T20.W, 0.0, |
| ; EG-NEXT: MOV * T21.W, 0.0, |
| ; EG-NEXT: MOV T22.W, 0.0, |
| ; EG-NEXT: MOV * T23.W, 0.0, |
| ; EG-NEXT: MOV T24.W, 0.0, |
| ; EG-NEXT: MOV * T25.W, 0.0, |
| ; EG-NEXT: MOV T26.W, 0.0, |
| ; EG-NEXT: MOV * T11.W, 0.0, |
| ; EG-NEXT: LSHR T27.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44) |
| ; EG-NEXT: LSHR T28.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44) |
| ; EG-NEXT: LSHR T29.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44) |
| ; EG-NEXT: LSHR T30.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 64(8.968310e-44) |
| ; EG-NEXT: LSHR T31.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 80(1.121039e-43) |
| ; EG-NEXT: LSHR * T32.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: ALU clause starting at 122: |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 96(1.345247e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T33.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 112(1.569454e-43) |
| ; EG-NEXT: LSHR T34.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 128(1.793662e-43) |
| ; EG-NEXT: LSHR T35.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 144(2.017870e-43) |
| ; EG-NEXT: LSHR T36.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 160(2.242078e-43) |
| ; EG-NEXT: LSHR T37.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 176(2.466285e-43) |
| ; EG-NEXT: LSHR T38.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 192(2.690493e-43) |
| ; EG-NEXT: LSHR T39.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 208(2.914701e-43) |
| ; EG-NEXT: LSHR T40.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 224(3.138909e-43) |
| ; EG-NEXT: LSHR T41.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 240(3.363116e-43) |
| ; EG-NEXT: LSHR * T42.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v32i1_to_v32i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_b32 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1001e |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s3 |
| ; GFX12-NEXT: s_lshr_b32 s4, s2, 31 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1001d |
| ; GFX12-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, v1 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x1001c |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:240 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1001b |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x1001a |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:224 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10019 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10018 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:208 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10017 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10016 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:192 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10014 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10015 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:176 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s3 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10013 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10012 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:160 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10011 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10010 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:144 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1000f |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x1000e |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:128 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1000d |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x1000c |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:112 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1000b |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x1000a |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:96 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10009 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10008 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:80 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10007 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10006 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:64 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10005 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10004 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:48 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10003 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10002 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:32 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10001 |
| ; GFX12-NEXT: s_and_b32 s2, s2, 1 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:16 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <32 x i1>, ptr addrspace(4) %in |
| %ext = zext <32 x i1> %load to <32 x i64> |
| store <32 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v32i1_to_v32i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v32i1_to_v32i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_load_dword s4, s[2:3], 0x0 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_lshr_b32 s38, s4, 30 |
| ; GFX6-NEXT: s_lshr_b32 s40, s4, 31 |
| ; GFX6-NEXT: s_lshr_b32 s34, s4, 28 |
| ; GFX6-NEXT: s_lshr_b32 s36, s4, 29 |
| ; GFX6-NEXT: s_lshr_b32 s28, s4, 26 |
| ; GFX6-NEXT: s_lshr_b32 s30, s4, 27 |
| ; GFX6-NEXT: s_lshr_b32 s24, s4, 24 |
| ; GFX6-NEXT: s_lshr_b32 s26, s4, 25 |
| ; GFX6-NEXT: s_lshr_b32 s20, s4, 22 |
| ; GFX6-NEXT: s_lshr_b32 s22, s4, 23 |
| ; GFX6-NEXT: s_lshr_b32 s18, s4, 20 |
| ; GFX6-NEXT: s_lshr_b32 s6, s4, 21 |
| ; GFX6-NEXT: s_lshr_b32 s8, s4, 18 |
| ; GFX6-NEXT: s_lshr_b32 s10, s4, 19 |
| ; GFX6-NEXT: s_lshr_b32 s12, s4, 16 |
| ; GFX6-NEXT: s_lshr_b32 s14, s4, 17 |
| ; GFX6-NEXT: s_lshr_b32 s16, s4, 14 |
| ; GFX6-NEXT: s_bfe_i64 s[44:45], s[4:5], 0x10000 |
| ; GFX6-NEXT: s_lshr_b32 s42, s4, 15 |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s44 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s45 |
| ; GFX6-NEXT: s_lshr_b32 s44, s4, 12 |
| ; GFX6-NEXT: s_bfe_i64 s[38:39], s[38:39], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[40:41], s[40:41], 0x10000 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX6-NEXT: s_lshr_b32 s38, s4, 13 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX6-NEXT: s_lshr_b32 s40, s4, 10 |
| ; GFX6-NEXT: s_bfe_i64 s[34:35], s[34:35], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[36:37], s[36:37], 0x10000 |
| ; GFX6-NEXT: v_mov_b32_e32 v6, s34 |
| ; GFX6-NEXT: v_mov_b32_e32 v7, s35 |
| ; GFX6-NEXT: s_lshr_b32 s34, s4, 11 |
| ; GFX6-NEXT: v_mov_b32_e32 v8, s36 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, s37 |
| ; GFX6-NEXT: s_lshr_b32 s36, s4, 8 |
| ; GFX6-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x10000 |
| ; GFX6-NEXT: v_mov_b32_e32 v10, s28 |
| ; GFX6-NEXT: v_mov_b32_e32 v11, s29 |
| ; GFX6-NEXT: s_lshr_b32 s28, s4, 9 |
| ; GFX6-NEXT: v_mov_b32_e32 v12, s30 |
| ; GFX6-NEXT: v_mov_b32_e32 v13, s31 |
| ; GFX6-NEXT: s_lshr_b32 s30, s4, 6 |
| ; GFX6-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x10000 |
| ; GFX6-NEXT: v_mov_b32_e32 v14, s24 |
| ; GFX6-NEXT: v_mov_b32_e32 v15, s25 |
| ; GFX6-NEXT: s_lshr_b32 s24, s4, 7 |
| ; GFX6-NEXT: v_mov_b32_e32 v16, s26 |
| ; GFX6-NEXT: v_mov_b32_e32 v17, s27 |
| ; GFX6-NEXT: s_lshr_b32 s26, s4, 4 |
| ; GFX6-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:240 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s20 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s21 |
| ; GFX6-NEXT: s_lshr_b32 s20, s4, 5 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s22 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s23 |
| ; GFX6-NEXT: s_lshr_b32 s22, s4, 2 |
| ; GFX6-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[6:9], off, s[0:3], 0 offset:224 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v6, s18 |
| ; GFX6-NEXT: v_mov_b32_e32 v7, s19 |
| ; GFX6-NEXT: s_lshr_b32 s18, s4, 3 |
| ; GFX6-NEXT: s_lshr_b32 s4, s4, 1 |
| ; GFX6-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[36:37], s[36:37], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[34:35], s[34:35], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[40:41], s[40:41], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[38:39], s[38:39], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[44:45], s[44:45], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[42:43], s[42:43], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[10:13], off, s[0:3], 0 offset:208 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[14:17], off, s[0:3], 0 offset:192 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:176 |
| ; GFX6-NEXT: v_mov_b32_e32 v8, s6 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, s7 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[6:9], off, s[0:3], 0 offset:160 |
| ; GFX6-NEXT: s_waitcnt expcnt(1) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s8 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s9 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s10 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s11 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:144 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s12 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s13 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s14 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s15 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:128 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s16 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s17 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:112 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s44 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s45 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s38 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s39 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:96 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s40 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s41 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s34 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s35 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:80 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s36 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s37 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s28 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s29 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:64 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s30 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s31 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s24 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s25 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s26 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s27 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s22 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s23 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s18 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s19 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s5 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v32i1_to_v32i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_load_dword s2, s[2:3], 0x0 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_lshr_b32 s44, s2, 30 |
| ; GFX8-NEXT: s_lshr_b32 s46, s2, 31 |
| ; GFX8-NEXT: s_lshr_b32 s48, s2, 28 |
| ; GFX8-NEXT: s_lshr_b32 s50, s2, 29 |
| ; GFX8-NEXT: s_lshr_b32 s52, s2, 26 |
| ; GFX8-NEXT: s_lshr_b32 s54, s2, 27 |
| ; GFX8-NEXT: s_lshr_b32 s56, s2, 24 |
| ; GFX8-NEXT: s_lshr_b32 s58, s2, 25 |
| ; GFX8-NEXT: s_lshr_b32 s60, s2, 22 |
| ; GFX8-NEXT: s_lshr_b32 s62, s2, 23 |
| ; GFX8-NEXT: s_lshr_b32 s64, s2, 20 |
| ; GFX8-NEXT: s_lshr_b32 s66, s2, 21 |
| ; GFX8-NEXT: s_lshr_b32 s42, s2, 18 |
| ; GFX8-NEXT: s_lshr_b32 s40, s2, 19 |
| ; GFX8-NEXT: s_lshr_b32 s38, s2, 16 |
| ; GFX8-NEXT: s_lshr_b32 s36, s2, 17 |
| ; GFX8-NEXT: s_lshr_b32 s34, s2, 14 |
| ; GFX8-NEXT: s_lshr_b32 s30, s2, 15 |
| ; GFX8-NEXT: s_lshr_b32 s28, s2, 12 |
| ; GFX8-NEXT: s_lshr_b32 s26, s2, 13 |
| ; GFX8-NEXT: s_lshr_b32 s24, s2, 10 |
| ; GFX8-NEXT: s_lshr_b32 s22, s2, 11 |
| ; GFX8-NEXT: s_lshr_b32 s20, s2, 8 |
| ; GFX8-NEXT: s_lshr_b32 s18, s2, 9 |
| ; GFX8-NEXT: s_lshr_b32 s16, s2, 6 |
| ; GFX8-NEXT: s_lshr_b32 s14, s2, 7 |
| ; GFX8-NEXT: s_lshr_b32 s12, s2, 4 |
| ; GFX8-NEXT: s_lshr_b32 s10, s2, 5 |
| ; GFX8-NEXT: s_lshr_b32 s8, s2, 2 |
| ; GFX8-NEXT: s_lshr_b32 s6, s2, 3 |
| ; GFX8-NEXT: s_lshr_b32 s68, s2, 1 |
| ; GFX8-NEXT: s_bfe_i64 s[4:5], s[2:3], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[2:3], s[68:69], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[34:35], s[34:35], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[36:37], s[36:37], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[38:39], s[38:39], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[40:41], s[40:41], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[42:43], s[42:43], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[66:67], s[66:67], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[64:65], s[64:65], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[62:63], s[62:63], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[60:61], s[60:61], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[58:59], s[58:59], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[56:57], s[56:57], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[54:55], s[54:55], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[52:53], s[52:53], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[50:51], s[50:51], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[48:49], s[48:49], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[46:47], s[46:47], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[44:45], s[44:45], 0x10000 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s44 |
| ; GFX8-NEXT: s_add_u32 s44, s0, 0xf0 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s45 |
| ; GFX8-NEXT: s_addc_u32 s45, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s44 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s46 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s47 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s45 |
| ; GFX8-NEXT: s_add_u32 s44, s0, 0xe0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s45, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s44 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s48 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s49 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s50 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s51 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s45 |
| ; GFX8-NEXT: s_add_u32 s44, s0, 0xd0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s45, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s44 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s52 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s53 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s54 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s55 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s45 |
| ; GFX8-NEXT: s_add_u32 s44, s0, 0xc0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s45, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s44 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s56 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s57 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s58 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s59 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s45 |
| ; GFX8-NEXT: s_add_u32 s44, s0, 0xb0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s45, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s44 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s60 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s61 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s62 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s63 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s45 |
| ; GFX8-NEXT: s_add_u32 s44, s0, 0xa0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s45, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s44 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s64 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s65 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s66 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s67 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s45 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s40 |
| ; GFX8-NEXT: s_add_u32 s40, s0, 0x90 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s41 |
| ; GFX8-NEXT: s_addc_u32 s41, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s43 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s36 |
| ; GFX8-NEXT: s_add_u32 s36, s0, 0x80 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s37 |
| ; GFX8-NEXT: s_addc_u32 s37, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s36 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s38 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s39 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s37 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s30 |
| ; GFX8-NEXT: s_add_u32 s30, s0, 0x70 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s31 |
| ; GFX8-NEXT: s_addc_u32 s31, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s30 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s34 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s35 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s31 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s26 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0x60 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s27 |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s28 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s29 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s22 |
| ; GFX8-NEXT: s_add_u32 s22, s0, 0x50 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s23 |
| ; GFX8-NEXT: s_addc_u32 s23, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s22 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s24 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s25 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s23 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX8-NEXT: s_add_u32 s18, s0, 64 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX8-NEXT: s_addc_u32 s19, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s18 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s20 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s21 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s19 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s14 |
| ; GFX8-NEXT: s_add_u32 s14, s0, 48 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s15 |
| ; GFX8-NEXT: s_addc_u32 s15, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s14 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s15 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s10 |
| ; GFX8-NEXT: s_add_u32 s10, s0, 32 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s11 |
| ; GFX8-NEXT: s_addc_u32 s11, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s10 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s12 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s13 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s11 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s6 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s7 |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s8 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s9 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s5 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v32i1_to_v32i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @24, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @22 |
| ; EG-NEXT: ALU 92, @25, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: ALU 65, @118, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T26.XYZW, T42.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T40.XYZW, T41.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T27.XYZW, T34.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T39.XYZW, T24.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T28.XYZW, T23.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T38.XYZW, T22.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T29.XYZW, T21.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T37.XYZW, T20.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T30.XYZW, T19.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T36.XYZW, T18.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T31.XYZW, T17.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T11.XYZW, T16.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T32.XYZW, T15.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T35.XYZW, T14.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T33.XYZW, T13.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T25.XYZW, T12.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: PAD |
| ; EG-NEXT: Fetch clause starting at 22: |
| ; EG-NEXT: VTX_READ_32 T11.X, T11.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 24: |
| ; EG-NEXT: MOV * T11.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 25: |
| ; EG-NEXT: LSHR T12.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44) |
| ; EG-NEXT: LSHR T13.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44) |
| ; EG-NEXT: LSHR T14.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44) |
| ; EG-NEXT: LSHR T15.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 64(8.968310e-44) |
| ; EG-NEXT: LSHR T16.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 80(1.121039e-43) |
| ; EG-NEXT: LSHR T17.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 96(1.345247e-43) |
| ; EG-NEXT: LSHR T18.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 112(1.569454e-43) |
| ; EG-NEXT: LSHR T19.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 128(1.793662e-43) |
| ; EG-NEXT: LSHR T20.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 144(2.017870e-43) |
| ; EG-NEXT: LSHR T21.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 160(2.242078e-43) |
| ; EG-NEXT: LSHR T22.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 176(2.466285e-43) |
| ; EG-NEXT: LSHR T23.X, PV.W, literal.x, |
| ; EG-NEXT: LSHR T0.Y, T11.X, literal.y, |
| ; EG-NEXT: LSHR T0.Z, T11.X, literal.z, |
| ; EG-NEXT: LSHR * T0.W, T11.X, literal.w, |
| ; EG-NEXT: 2(2.802597e-45), 28(3.923636e-44) |
| ; EG-NEXT: 29(4.063766e-44), 24(3.363116e-44) |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 192(2.690493e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T24.X, PV.W, literal.x, |
| ; EG-NEXT: LSHR T1.Y, T11.X, literal.y, |
| ; EG-NEXT: LSHR T1.Z, T11.X, literal.z, |
| ; EG-NEXT: LSHR * T1.W, T11.X, literal.w, |
| ; EG-NEXT: 2(2.802597e-45), 25(3.503246e-44) |
| ; EG-NEXT: 20(2.802597e-44), 21(2.942727e-44) |
| ; EG-NEXT: LSHR * T2.W, T11.X, literal.x, |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T25.X, T11.X, 0.0, 1, |
| ; EG-NEXT: LSHR T2.Y, T11.X, literal.x, |
| ; EG-NEXT: ASHR T26.Z, T11.X, literal.y, |
| ; EG-NEXT: LSHR T3.W, T11.X, literal.z, |
| ; EG-NEXT: LSHR * T4.W, T11.X, literal.w, |
| ; EG-NEXT: 17(2.382207e-44), 31(4.344025e-44) |
| ; EG-NEXT: 27(3.783506e-44), 30(4.203895e-44) |
| ; EG-NEXT: BFE_INT T26.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T3.Y, T11.X, literal.x, |
| ; EG-NEXT: BFE_INT T27.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T3.W, T11.X, literal.y, |
| ; EG-NEXT: LSHR * T4.W, T11.X, literal.z, |
| ; EG-NEXT: 12(1.681558e-44), 23(3.222986e-44) |
| ; EG-NEXT: 26(3.643376e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T27.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T26.Y, PV.X, |
| ; EG-NEXT: BFE_INT T28.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T3.W, T11.X, literal.x, |
| ; EG-NEXT: LSHR * T4.W, T11.X, literal.y, |
| ; EG-NEXT: 19(2.662467e-44), 22(3.082857e-44) |
| ; EG-NEXT: BFE_INT T28.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T27.Y, PV.X, |
| ; EG-NEXT: BFE_INT T29.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T3.W, T11.X, literal.x, |
| ; EG-NEXT: LSHR * T4.W, T11.X, literal.y, |
| ; EG-NEXT: 15(2.101948e-44), 18(2.522337e-44) |
| ; EG-NEXT: BFE_INT T29.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T28.Y, PV.X, |
| ; EG-NEXT: BFE_INT T30.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T3.W, T11.X, literal.x, |
| ; EG-NEXT: LSHR * T4.W, T11.X, literal.y, |
| ; EG-NEXT: 11(1.541428e-44), 14(1.961818e-44) |
| ; EG-NEXT: BFE_INT T30.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T29.Y, PV.X, |
| ; EG-NEXT: BFE_INT T31.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T3.W, T11.X, literal.x, |
| ; EG-NEXT: LSHR * T4.W, T11.X, literal.y, |
| ; EG-NEXT: 7(9.809089e-45), 10(1.401298e-44) |
| ; EG-NEXT: BFE_INT T31.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T30.Y, PV.X, |
| ; EG-NEXT: BFE_INT T32.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T3.W, T11.X, literal.x, |
| ; EG-NEXT: LSHR * T4.W, T11.X, literal.y, |
| ; EG-NEXT: 3(4.203895e-45), 6(8.407791e-45) |
| ; EG-NEXT: ALU clause starting at 118: |
| ; EG-NEXT: BFE_INT T32.X, T4.W, 0.0, 1, |
| ; EG-NEXT: MOV T31.Y, T31.X, |
| ; EG-NEXT: BFE_INT T33.Z, T3.W, 0.0, 1, BS:VEC_120/SCL_212 |
| ; EG-NEXT: LSHR T3.W, T11.X, 1, BS:VEC_120/SCL_212 |
| ; EG-NEXT: LSHR * T4.W, T11.X, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T33.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T32.Y, PV.X, |
| ; EG-NEXT: BFE_INT T25.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T3.W, T11.X, literal.x, |
| ; EG-NEXT: ADD_INT * T4.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 5(7.006492e-45), 208(2.914701e-43) |
| ; EG-NEXT: LSHR T34.X, PS, literal.x, |
| ; EG-NEXT: MOV T33.Y, PV.X, |
| ; EG-NEXT: BFE_INT T35.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T3.W, T11.X, literal.y, |
| ; EG-NEXT: LSHR * T4.W, T11.X, literal.z, |
| ; EG-NEXT: 2(2.802597e-45), 9(1.261169e-44) |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T35.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T25.Y, T25.X, |
| ; EG-NEXT: BFE_INT T11.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T3.W, T11.X, literal.x, BS:VEC_120/SCL_212 |
| ; EG-NEXT: LSHR * T4.W, T11.X, literal.y, |
| ; EG-NEXT: 13(1.821688e-44), 8(1.121039e-44) |
| ; EG-NEXT: BFE_INT T11.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T35.Y, PV.X, |
| ; EG-NEXT: BFE_INT T36.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: MOV T25.W, T25.Z, |
| ; EG-NEXT: MOV * T33.W, T33.Z, |
| ; EG-NEXT: BFE_INT T36.X, T3.Y, 0.0, 1, |
| ; EG-NEXT: MOV T11.Y, PV.X, |
| ; EG-NEXT: BFE_INT T37.Z, T2.Y, 0.0, 1, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV T35.W, T35.Z, |
| ; EG-NEXT: MOV * T32.W, T32.Z, |
| ; EG-NEXT: BFE_INT T37.X, T2.W, 0.0, 1, |
| ; EG-NEXT: MOV T36.Y, PV.X, |
| ; EG-NEXT: BFE_INT T38.Z, T1.W, 0.0, 1, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV T11.W, T11.Z, |
| ; EG-NEXT: MOV * T31.W, T31.Z, |
| ; EG-NEXT: BFE_INT T38.X, T1.Z, 0.0, 1, |
| ; EG-NEXT: MOV T37.Y, PV.X, |
| ; EG-NEXT: BFE_INT T39.Z, T1.Y, 0.0, 1, |
| ; EG-NEXT: MOV T36.W, T36.Z, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV * T30.W, T30.Z, |
| ; EG-NEXT: BFE_INT T39.X, T0.W, 0.0, 1, |
| ; EG-NEXT: MOV T38.Y, PV.X, |
| ; EG-NEXT: BFE_INT T40.Z, T0.Z, 0.0, 1, |
| ; EG-NEXT: MOV T37.W, T37.Z, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV * T29.W, T29.Z, |
| ; EG-NEXT: BFE_INT T40.X, T0.Y, 0.0, 1, |
| ; EG-NEXT: MOV T39.Y, PV.X, |
| ; EG-NEXT: ADD_INT T0.Z, KC0[2].Y, literal.x, |
| ; EG-NEXT: MOV T38.W, T38.Z, |
| ; EG-NEXT: MOV * T28.W, T28.Z, |
| ; EG-NEXT: 224(3.138909e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T41.X, PV.Z, literal.x, |
| ; EG-NEXT: MOV T40.Y, PV.X, |
| ; EG-NEXT: ADD_INT T0.Z, KC0[2].Y, literal.y, |
| ; EG-NEXT: MOV T39.W, T39.Z, |
| ; EG-NEXT: MOV * T27.W, T27.Z, |
| ; EG-NEXT: 2(2.802597e-45), 240(3.363116e-43) |
| ; EG-NEXT: LSHR T42.X, PV.Z, literal.x, |
| ; EG-NEXT: MOV T40.W, T40.Z, |
| ; EG-NEXT: MOV * T26.W, T26.Z, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v32i1_to_v32i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_b32 s2, s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_lshr_b32 s34, s2, 30 |
| ; GFX12-NEXT: s_lshr_b32 s36, s2, 31 |
| ; GFX12-NEXT: s_lshr_b32 s38, s2, 28 |
| ; GFX12-NEXT: s_lshr_b32 s40, s2, 29 |
| ; GFX12-NEXT: s_lshr_b32 s42, s2, 26 |
| ; GFX12-NEXT: s_lshr_b32 s44, s2, 27 |
| ; GFX12-NEXT: s_bfe_i64 s[34:35], s[34:35], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[36:37], s[36:37], 0x10000 |
| ; GFX12-NEXT: s_lshr_b32 s46, s2, 24 |
| ; GFX12-NEXT: s_lshr_b32 s48, s2, 25 |
| ; GFX12-NEXT: s_bfe_i64 s[40:41], s[40:41], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[38:39], s[38:39], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s35 |
| ; GFX12-NEXT: s_bfe_i64 s[44:45], s[44:45], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[42:43], s[42:43], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, s34 :: v_dual_mov_b32 v3, s37 |
| ; GFX12-NEXT: v_dual_mov_b32 v2, s36 :: v_dual_mov_b32 v5, s39 |
| ; GFX12-NEXT: s_lshr_b32 s26, s2, 22 |
| ; GFX12-NEXT: s_lshr_b32 s50, s2, 23 |
| ; GFX12-NEXT: s_bfe_i64 s[48:49], s[48:49], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[46:47], s[46:47], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v4, s38 :: v_dual_mov_b32 v7, s41 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s40 :: v_dual_mov_b32 v9, s43 |
| ; GFX12-NEXT: s_lshr_b32 s52, s2, 20 |
| ; GFX12-NEXT: s_lshr_b32 s54, s2, 21 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s42 :: v_dual_mov_b32 v11, s45 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s44 :: v_dual_mov_b32 v13, s47 |
| ; GFX12-NEXT: s_lshr_b32 s56, s2, 18 |
| ; GFX12-NEXT: s_lshr_b32 s58, s2, 19 |
| ; GFX12-NEXT: s_bfe_i64 s[50:51], s[50:51], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s46 :: v_dual_mov_b32 v15, s49 |
| ; GFX12-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x10000 |
| ; GFX12-NEXT: v_mov_b32_e32 v14, s48 |
| ; GFX12-NEXT: s_lshr_b32 s60, s2, 16 |
| ; GFX12-NEXT: s_lshr_b32 s62, s2, 17 |
| ; GFX12-NEXT: s_bfe_i64 s[54:55], s[54:55], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[52:53], s[52:53], 0x10000 |
| ; GFX12-NEXT: s_lshr_b32 s64, s2, 14 |
| ; GFX12-NEXT: s_lshr_b32 s66, s2, 15 |
| ; GFX12-NEXT: s_bfe_i64 s[58:59], s[58:59], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[56:57], s[56:57], 0x10000 |
| ; GFX12-NEXT: s_clause 0x3 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:240 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:224 |
| ; GFX12-NEXT: global_store_b128 v24, v[8:11], s[0:1] offset:208 |
| ; GFX12-NEXT: global_store_b128 v24, v[12:15], s[0:1] offset:192 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s27 :: v_dual_mov_b32 v0, s26 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s51 :: v_dual_mov_b32 v2, s50 |
| ; GFX12-NEXT: v_mov_b32_e32 v5, s53 |
| ; GFX12-NEXT: s_lshr_b32 s30, s2, 12 |
| ; GFX12-NEXT: s_lshr_b32 s28, s2, 13 |
| ; GFX12-NEXT: s_lshr_b32 s24, s2, 10 |
| ; GFX12-NEXT: s_lshr_b32 s22, s2, 11 |
| ; GFX12-NEXT: s_bfe_i64 s[62:63], s[62:63], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[60:61], s[60:61], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v4, s52 :: v_dual_mov_b32 v7, s55 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s54 :: v_dual_mov_b32 v9, s57 |
| ; GFX12-NEXT: s_lshr_b32 s20, s2, 8 |
| ; GFX12-NEXT: s_lshr_b32 s18, s2, 9 |
| ; GFX12-NEXT: s_bfe_i64 s[66:67], s[66:67], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[64:65], s[64:65], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s56 :: v_dual_mov_b32 v11, s59 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s58 :: v_dual_mov_b32 v13, s61 |
| ; GFX12-NEXT: s_lshr_b32 s16, s2, 6 |
| ; GFX12-NEXT: s_lshr_b32 s14, s2, 7 |
| ; GFX12-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s60 :: v_dual_mov_b32 v15, s63 |
| ; GFX12-NEXT: v_dual_mov_b32 v14, s62 :: v_dual_mov_b32 v17, s65 |
| ; GFX12-NEXT: s_lshr_b32 s12, s2, 4 |
| ; GFX12-NEXT: s_lshr_b32 s10, s2, 5 |
| ; GFX12-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v16, s64 :: v_dual_mov_b32 v19, s67 |
| ; GFX12-NEXT: v_dual_mov_b32 v18, s66 :: v_dual_mov_b32 v21, s31 |
| ; GFX12-NEXT: s_lshr_b32 s8, s2, 2 |
| ; GFX12-NEXT: s_lshr_b32 s6, s2, 3 |
| ; GFX12-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v20, s30 :: v_dual_mov_b32 v23, s29 |
| ; GFX12-NEXT: v_mov_b32_e32 v22, s28 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:176 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:160 |
| ; GFX12-NEXT: global_store_b128 v24, v[8:11], s[0:1] offset:144 |
| ; GFX12-NEXT: global_store_b128 v24, v[12:15], s[0:1] offset:128 |
| ; GFX12-NEXT: global_store_b128 v24, v[16:19], s[0:1] offset:112 |
| ; GFX12-NEXT: global_store_b128 v24, v[20:23], s[0:1] offset:96 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s25 :: v_dual_mov_b32 v0, s24 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s23 :: v_dual_mov_b32 v2, s22 |
| ; GFX12-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX12-NEXT: s_lshr_b32 s68, s2, 1 |
| ; GFX12-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v4, s20 :: v_dual_mov_b32 v7, s19 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s17 |
| ; GFX12-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s16 :: v_dual_mov_b32 v11, s15 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s14 :: v_dual_mov_b32 v13, s13 |
| ; GFX12-NEXT: s_bfe_i64 s[4:5], s[2:3], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[2:3], s[68:69], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v15, s11 |
| ; GFX12-NEXT: v_dual_mov_b32 v14, s10 :: v_dual_mov_b32 v17, s9 |
| ; GFX12-NEXT: v_dual_mov_b32 v16, s8 :: v_dual_mov_b32 v19, s7 |
| ; GFX12-NEXT: v_dual_mov_b32 v18, s6 :: v_dual_mov_b32 v21, s5 |
| ; GFX12-NEXT: v_dual_mov_b32 v20, s4 :: v_dual_mov_b32 v23, s3 |
| ; GFX12-NEXT: v_mov_b32_e32 v22, s2 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v24, v[0:3], s[0:1] offset:80 |
| ; GFX12-NEXT: global_store_b128 v24, v[4:7], s[0:1] offset:64 |
| ; GFX12-NEXT: global_store_b128 v24, v[8:11], s[0:1] offset:48 |
| ; GFX12-NEXT: global_store_b128 v24, v[12:15], s[0:1] offset:32 |
| ; GFX12-NEXT: global_store_b128 v24, v[16:19], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v24, v[20:23], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <32 x i1>, ptr addrspace(4) %in |
| %ext = sext <32 x i1> %load to <32 x i64> |
| store <32 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_zextload_v64i1_to_v64i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_zextload_v64i1_to_v64i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_bfe_u32 s4, s2, 0x10003 |
| ; GFX6-NEXT: s_bfe_u32 s5, s2, 0x10005 |
| ; GFX6-NEXT: s_bfe_u32 s8, s2, 0x10007 |
| ; GFX6-NEXT: s_bfe_u32 s11, s2, 0x10009 |
| ; GFX6-NEXT: s_bfe_u32 s13, s2, 0x1000b |
| ; GFX6-NEXT: s_bfe_u32 s15, s2, 0x1000d |
| ; GFX6-NEXT: s_bfe_u32 s17, s2, 0x1000f |
| ; GFX6-NEXT: s_bfe_u32 s19, s2, 0x10011 |
| ; GFX6-NEXT: s_bfe_u32 s21, s2, 0x10013 |
| ; GFX6-NEXT: s_bfe_u32 s23, s2, 0x10015 |
| ; GFX6-NEXT: s_bfe_u32 s25, s2, 0x10017 |
| ; GFX6-NEXT: s_bfe_u32 s27, s2, 0x10019 |
| ; GFX6-NEXT: s_bfe_u32 s29, s2, 0x1001b |
| ; GFX6-NEXT: s_bfe_u32 s31, s2, 0x1001d |
| ; GFX6-NEXT: s_lshr_b32 s34, s2, 31 |
| ; GFX6-NEXT: s_bfe_u32 s35, s3, 0x10003 |
| ; GFX6-NEXT: s_bfe_u32 s36, s3, 0x10005 |
| ; GFX6-NEXT: s_bfe_u32 s37, s3, 0x10007 |
| ; GFX6-NEXT: s_bfe_u32 s38, s3, 0x10009 |
| ; GFX6-NEXT: s_bfe_u32 s39, s3, 0x1000b |
| ; GFX6-NEXT: s_bfe_u32 s40, s3, 0x1000d |
| ; GFX6-NEXT: s_bfe_u32 s41, s3, 0x1000f |
| ; GFX6-NEXT: s_bfe_u32 s42, s3, 0x10011 |
| ; GFX6-NEXT: s_bfe_u32 s43, s3, 0x10013 |
| ; GFX6-NEXT: s_bfe_u32 s44, s3, 0x10015 |
| ; GFX6-NEXT: s_bfe_u32 s45, s3, 0x10017 |
| ; GFX6-NEXT: s_bfe_u32 s46, s3, 0x10019 |
| ; GFX6-NEXT: s_bfe_u32 s47, s3, 0x1001b |
| ; GFX6-NEXT: s_bfe_u32 s48, s3, 0x1001d |
| ; GFX6-NEXT: s_lshr_b32 s49, s3, 31 |
| ; GFX6-NEXT: s_bfe_u32 s9, s3, 0x10001 |
| ; GFX6-NEXT: s_bfe_u32 s6, s2, 0x10001 |
| ; GFX6-NEXT: s_and_b32 s7, s2, 1 |
| ; GFX6-NEXT: s_and_b32 s10, s3, 1 |
| ; GFX6-NEXT: s_bfe_u32 s12, s2, 0x10002 |
| ; GFX6-NEXT: s_bfe_u32 s14, s2, 0x10004 |
| ; GFX6-NEXT: s_bfe_u32 s16, s2, 0x10006 |
| ; GFX6-NEXT: s_bfe_u32 s18, s2, 0x10008 |
| ; GFX6-NEXT: s_bfe_u32 s20, s2, 0x1000a |
| ; GFX6-NEXT: s_bfe_u32 s22, s2, 0x1000c |
| ; GFX6-NEXT: s_bfe_u32 s24, s2, 0x1000e |
| ; GFX6-NEXT: s_bfe_u32 s26, s2, 0x10010 |
| ; GFX6-NEXT: s_bfe_u32 s28, s2, 0x10012 |
| ; GFX6-NEXT: s_bfe_u32 s30, s2, 0x10014 |
| ; GFX6-NEXT: s_bfe_u32 s33, s2, 0x10016 |
| ; GFX6-NEXT: s_bfe_u32 s50, s2, 0x10018 |
| ; GFX6-NEXT: s_bfe_u32 s51, s2, 0x1001a |
| ; GFX6-NEXT: s_bfe_u32 s52, s2, 0x1001c |
| ; GFX6-NEXT: s_bfe_u32 s53, s2, 0x1001e |
| ; GFX6-NEXT: s_bfe_u32 s54, s3, 0x10002 |
| ; GFX6-NEXT: s_bfe_u32 s55, s3, 0x10004 |
| ; GFX6-NEXT: s_bfe_u32 s56, s3, 0x10006 |
| ; GFX6-NEXT: s_bfe_u32 s57, s3, 0x10008 |
| ; GFX6-NEXT: s_bfe_u32 s58, s3, 0x1000a |
| ; GFX6-NEXT: s_bfe_u32 s59, s3, 0x1000c |
| ; GFX6-NEXT: s_bfe_u32 s60, s3, 0x1000e |
| ; GFX6-NEXT: s_bfe_u32 s61, s3, 0x10010 |
| ; GFX6-NEXT: s_bfe_u32 s62, s3, 0x10012 |
| ; GFX6-NEXT: s_bfe_u32 s63, s3, 0x10014 |
| ; GFX6-NEXT: s_bfe_u32 s64, s3, 0x10016 |
| ; GFX6-NEXT: s_bfe_u32 s65, s3, 0x10018 |
| ; GFX6-NEXT: s_bfe_u32 s66, s3, 0x1001a |
| ; GFX6-NEXT: s_bfe_u32 s67, s3, 0x1001e |
| ; GFX6-NEXT: s_bfe_u32 s68, s3, 0x1001c |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, v1 |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s67 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s49 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:496 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s68 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s48 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:480 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s66 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s47 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:464 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s65 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s46 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:448 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s64 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s45 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:432 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s63 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s44 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:416 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s62 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s43 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:400 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s61 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s42 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:384 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s60 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s41 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:368 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s59 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s40 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:352 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s58 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s39 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:336 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s57 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:320 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s56 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s37 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:304 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s55 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s36 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:288 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s54 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s35 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:272 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s53 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s34 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:240 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s52 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s31 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:224 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s51 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s29 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:208 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s50 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s27 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:192 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s33 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s25 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:176 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s30 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s23 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:160 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s28 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s21 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:144 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s26 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s19 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:128 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s24 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s17 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:112 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s22 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s15 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:96 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s20 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s13 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s18 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s11 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s8 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s14 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s5 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s12 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s10 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s9 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:256 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s7 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s6 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_zextload_v64i1_to_v64i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, v1 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_load_dwordx2 s[42:43], s[2:3], 0x0 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_lshr_b32 s44, s43, 31 |
| ; GFX8-NEXT: s_bfe_u32 s45, s43, 0x1001d |
| ; GFX8-NEXT: s_bfe_u32 s46, s43, 0x1001b |
| ; GFX8-NEXT: s_bfe_u32 s47, s43, 0x10019 |
| ; GFX8-NEXT: s_bfe_u32 s48, s43, 0x10017 |
| ; GFX8-NEXT: s_bfe_u32 s49, s43, 0x10013 |
| ; GFX8-NEXT: s_bfe_u32 s50, s43, 0x10011 |
| ; GFX8-NEXT: s_bfe_u32 s51, s43, 0x1000f |
| ; GFX8-NEXT: s_bfe_u32 s52, s43, 0x1000d |
| ; GFX8-NEXT: s_bfe_u32 s53, s43, 0x1000b |
| ; GFX8-NEXT: s_bfe_u32 s40, s43, 0x10009 |
| ; GFX8-NEXT: s_bfe_u32 s38, s43, 0x10007 |
| ; GFX8-NEXT: s_bfe_u32 s37, s43, 0x10005 |
| ; GFX8-NEXT: s_bfe_u32 s35, s43, 0x10003 |
| ; GFX8-NEXT: s_bfe_u32 s33, s43, 0x10001 |
| ; GFX8-NEXT: s_lshr_b32 s30, s42, 31 |
| ; GFX8-NEXT: s_bfe_u32 s28, s42, 0x1001d |
| ; GFX8-NEXT: s_bfe_u32 s26, s42, 0x1001b |
| ; GFX8-NEXT: s_bfe_u32 s25, s42, 0x10019 |
| ; GFX8-NEXT: s_bfe_u32 s22, s42, 0x10017 |
| ; GFX8-NEXT: s_bfe_u32 s19, s42, 0x10013 |
| ; GFX8-NEXT: s_bfe_u32 s17, s42, 0x10011 |
| ; GFX8-NEXT: s_bfe_u32 s15, s42, 0x1000f |
| ; GFX8-NEXT: s_bfe_u32 s13, s42, 0x1000d |
| ; GFX8-NEXT: s_bfe_u32 s12, s42, 0x1000b |
| ; GFX8-NEXT: s_bfe_u32 s10, s42, 0x10009 |
| ; GFX8-NEXT: s_bfe_u32 s8, s42, 0x10007 |
| ; GFX8-NEXT: s_bfe_u32 s6, s42, 0x10005 |
| ; GFX8-NEXT: s_bfe_u32 s4, s42, 0x10003 |
| ; GFX8-NEXT: s_bfe_u32 s2, s42, 0x10001 |
| ; GFX8-NEXT: s_and_b32 s3, s42, 1 |
| ; GFX8-NEXT: s_bfe_u32 s5, s42, 0x10002 |
| ; GFX8-NEXT: s_bfe_u32 s7, s42, 0x10004 |
| ; GFX8-NEXT: s_bfe_u32 s9, s42, 0x10006 |
| ; GFX8-NEXT: s_bfe_u32 s11, s42, 0x10008 |
| ; GFX8-NEXT: s_bfe_u32 s14, s42, 0x1000a |
| ; GFX8-NEXT: s_bfe_u32 s16, s42, 0x1000c |
| ; GFX8-NEXT: s_bfe_u32 s18, s42, 0x1000e |
| ; GFX8-NEXT: s_bfe_u32 s20, s42, 0x10010 |
| ; GFX8-NEXT: s_bfe_u32 s21, s42, 0x10012 |
| ; GFX8-NEXT: s_bfe_u32 s23, s42, 0x10014 |
| ; GFX8-NEXT: s_bfe_u32 s24, s42, 0x10015 |
| ; GFX8-NEXT: s_bfe_u32 s27, s42, 0x10016 |
| ; GFX8-NEXT: s_bfe_u32 s29, s42, 0x10018 |
| ; GFX8-NEXT: s_bfe_u32 s31, s42, 0x1001a |
| ; GFX8-NEXT: s_bfe_u32 s34, s42, 0x1001c |
| ; GFX8-NEXT: s_bfe_u32 s36, s42, 0x1001e |
| ; GFX8-NEXT: s_and_b32 s39, s43, 1 |
| ; GFX8-NEXT: s_bfe_u32 s41, s43, 0x10002 |
| ; GFX8-NEXT: s_bfe_u32 s54, s43, 0x10004 |
| ; GFX8-NEXT: s_bfe_u32 s55, s43, 0x10006 |
| ; GFX8-NEXT: s_bfe_u32 s56, s43, 0x10008 |
| ; GFX8-NEXT: s_bfe_u32 s57, s43, 0x1000a |
| ; GFX8-NEXT: s_bfe_u32 s58, s43, 0x1000c |
| ; GFX8-NEXT: s_bfe_u32 s59, s43, 0x1000e |
| ; GFX8-NEXT: s_bfe_u32 s60, s43, 0x10010 |
| ; GFX8-NEXT: s_bfe_u32 s61, s43, 0x10012 |
| ; GFX8-NEXT: s_bfe_u32 s62, s43, 0x10016 |
| ; GFX8-NEXT: s_bfe_u32 s63, s43, 0x10018 |
| ; GFX8-NEXT: s_bfe_u32 s64, s43, 0x1001a |
| ; GFX8-NEXT: s_bfe_u32 s65, s43, 0x1001c |
| ; GFX8-NEXT: s_bfe_u32 s66, s43, 0x1001e |
| ; GFX8-NEXT: s_bfe_u32 s42, s43, 0x10015 |
| ; GFX8-NEXT: s_bfe_u32 s43, s43, 0x10014 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s42 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x1a0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s43 |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x1f0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s66 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s44 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x1e0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s65 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s45 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x1d0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s64 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s46 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x1c0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s63 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s47 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x1b0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s62 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s48 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x190 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s61 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s49 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x180 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s60 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s50 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x170 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s59 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s51 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x160 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s58 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s52 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x150 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s57 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s53 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x140 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s56 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s40 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x130 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s55 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s42, s0, 0x120 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s43, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s54 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s37 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s43 |
| ; GFX8-NEXT: s_add_u32 s40, s0, 0x110 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s41 |
| ; GFX8-NEXT: s_addc_u32 s41, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s40 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s35 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s41 |
| ; GFX8-NEXT: s_add_u32 s38, s0, 0x100 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s39 |
| ; GFX8-NEXT: s_addc_u32 s39, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s38 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s33 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s39 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX8-NEXT: s_add_u32 s36, s0, 0xf0 |
| ; GFX8-NEXT: s_addc_u32 s37, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s36 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s30 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s37 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s34 |
| ; GFX8-NEXT: s_add_u32 s34, s0, 0xe0 |
| ; GFX8-NEXT: s_addc_u32 s35, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s34 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s28 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s35 |
| ; GFX8-NEXT: s_add_u32 s30, s0, 0xd0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s31 |
| ; GFX8-NEXT: s_addc_u32 s31, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s30 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s31 |
| ; GFX8-NEXT: s_add_u32 s28, s0, 0xc0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s29 |
| ; GFX8-NEXT: s_addc_u32 s29, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s28 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s25 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s29 |
| ; GFX8-NEXT: s_add_u32 s26, s0, 0xb0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s27 |
| ; GFX8-NEXT: s_addc_u32 s27, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s22 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s27 |
| ; GFX8-NEXT: s_add_u32 s22, s0, 0xa0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s23 |
| ; GFX8-NEXT: s_addc_u32 s23, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s22 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s24 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s23 |
| ; GFX8-NEXT: s_add_u32 s22, s0, 0x90 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_addc_u32 s23, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s22 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s21 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s19 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s23 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s20 |
| ; GFX8-NEXT: s_add_u32 s20, s0, 0x80 |
| ; GFX8-NEXT: s_addc_u32 s21, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s20 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s17 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s21 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s18 |
| ; GFX8-NEXT: s_add_u32 s18, s0, 0x70 |
| ; GFX8-NEXT: s_addc_u32 s19, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s18 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s15 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s19 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX8-NEXT: s_add_u32 s16, s0, 0x60 |
| ; GFX8-NEXT: s_addc_u32 s17, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s16 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s13 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s17 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s12 |
| ; GFX8-NEXT: s_add_u32 s12, s0, 0x50 |
| ; GFX8-NEXT: s_addc_u32 s13, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s12 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s14 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s13 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s10 |
| ; GFX8-NEXT: s_add_u32 s10, s0, 64 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s11 |
| ; GFX8-NEXT: s_addc_u32 s11, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s10 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s11 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s8 |
| ; GFX8-NEXT: s_add_u32 s8, s0, 48 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s9 |
| ; GFX8-NEXT: s_addc_u32 s9, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s8 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s9 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s6 |
| ; GFX8-NEXT: s_add_u32 s6, s0, 32 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s7 |
| ; GFX8-NEXT: s_addc_u32 s7, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_nop 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX8-NEXT: s_add_u32 s4, s0, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX8-NEXT: s_addc_u32 s5, s1, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s5 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_zextload_v64i1_to_v64i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 0, @40, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @38 |
| ; EG-NEXT: ALU 95, @41, KC0[], KC1[] |
| ; EG-NEXT: ALU 99, @137, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: ALU 60, @237, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T82.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T81.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T21.XYZW, T80.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T22.XYZW, T79.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T23.XYZW, T78.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T24.XYZW, T77.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T26.XYZW, T76.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T27.XYZW, T75.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T28.XYZW, T74.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T29.XYZW, T73.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T30.XYZW, T72.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T31.XYZW, T71.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T32.XYZW, T70.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T33.XYZW, T69.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T34.XYZW, T68.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T35.XYZW, T67.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T36.XYZW, T66.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T37.XYZW, T65.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T38.XYZW, T64.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T39.XYZW, T63.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T40.XYZW, T62.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T41.XYZW, T61.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T42.XYZW, T60.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T43.XYZW, T59.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T44.XYZW, T58.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T45.XYZW, T57.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T46.XYZW, T56.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T47.XYZW, T55.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T48.XYZW, T54.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T49.XYZW, T53.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T50.XYZW, T52.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T25.XYZW, T51.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 38: |
| ; EG-NEXT: VTX_READ_64 T25.XY, T19.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 40: |
| ; EG-NEXT: MOV * T19.X, KC0[2].Z, |
| ; EG-NEXT: ALU clause starting at 41: |
| ; EG-NEXT: LSHR * T19.Z, T25.Y, literal.x, |
| ; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T19.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T19.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T20.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 30(4.203895e-44), 29(4.063766e-44) |
| ; EG-NEXT: BFE_UINT T20.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T20.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T21.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 28(3.923636e-44), 27(3.783506e-44) |
| ; EG-NEXT: BFE_UINT T21.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T21.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T22.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 26(3.643376e-44), 25(3.503246e-44) |
| ; EG-NEXT: BFE_UINT T22.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T22.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T23.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 24(3.363116e-44), 23(3.222986e-44) |
| ; EG-NEXT: BFE_UINT T23.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T23.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T24.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 22(3.082857e-44), 21(2.942727e-44) |
| ; EG-NEXT: BFE_UINT T24.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T24.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T26.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 20(2.802597e-44), 19(2.662467e-44) |
| ; EG-NEXT: BFE_UINT T26.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T26.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T27.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 18(2.522337e-44), 17(2.382207e-44) |
| ; EG-NEXT: BFE_UINT T27.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T27.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T28.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 16(2.242078e-44), 15(2.101948e-44) |
| ; EG-NEXT: BFE_UINT T28.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T28.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T29.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 14(1.961818e-44), 13(1.821688e-44) |
| ; EG-NEXT: BFE_UINT T29.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T29.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T30.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 12(1.681558e-44), 11(1.541428e-44) |
| ; EG-NEXT: BFE_UINT T30.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T30.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T31.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 10(1.401298e-44), 9(1.261169e-44) |
| ; EG-NEXT: BFE_UINT T31.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T31.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T32.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 8(1.121039e-44), 7(9.809089e-45) |
| ; EG-NEXT: BFE_UINT T32.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T32.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T33.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 6(8.407791e-45), 5(7.006492e-45) |
| ; EG-NEXT: BFE_UINT T33.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T33.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T34.Z, T25.Y, literal.y, 1, |
| ; EG-NEXT: 4(5.605194e-45), 3(4.203895e-45) |
| ; EG-NEXT: BFE_UINT T34.X, T25.Y, literal.x, 1, |
| ; EG-NEXT: MOV T34.Y, 0.0, |
| ; EG-NEXT: BFE_UINT T35.Z, T25.Y, 1, 1, |
| ; EG-NEXT: AND_INT * T35.X, T25.Y, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: MOV T35.Y, 0.0, |
| ; EG-NEXT: LSHR * T36.Z, T25.X, literal.x, |
| ; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T36.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T36.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T37.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 30(4.203895e-44), 29(4.063766e-44) |
| ; EG-NEXT: BFE_UINT T37.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T37.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T38.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 28(3.923636e-44), 27(3.783506e-44) |
| ; EG-NEXT: BFE_UINT T38.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T38.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T39.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 26(3.643376e-44), 25(3.503246e-44) |
| ; EG-NEXT: BFE_UINT T39.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T39.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T40.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 24(3.363116e-44), 23(3.222986e-44) |
| ; EG-NEXT: BFE_UINT T40.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T40.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T41.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 22(3.082857e-44), 21(2.942727e-44) |
| ; EG-NEXT: BFE_UINT T41.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T41.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T42.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 20(2.802597e-44), 19(2.662467e-44) |
| ; EG-NEXT: BFE_UINT T42.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T42.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T43.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 18(2.522337e-44), 17(2.382207e-44) |
| ; EG-NEXT: BFE_UINT * T43.X, T25.X, literal.x, 1, |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) |
| ; EG-NEXT: ALU clause starting at 137: |
| ; EG-NEXT: MOV T43.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T44.Z, T25.X, literal.x, 1, |
| ; EG-NEXT: 15(2.101948e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_UINT T44.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T44.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T45.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 14(1.961818e-44), 13(1.821688e-44) |
| ; EG-NEXT: BFE_UINT T45.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T45.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T46.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 12(1.681558e-44), 11(1.541428e-44) |
| ; EG-NEXT: BFE_UINT T46.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T46.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T47.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 10(1.401298e-44), 9(1.261169e-44) |
| ; EG-NEXT: BFE_UINT T47.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T47.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T48.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 8(1.121039e-44), 7(9.809089e-45) |
| ; EG-NEXT: BFE_UINT T48.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T48.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T49.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 6(8.407791e-45), 5(7.006492e-45) |
| ; EG-NEXT: BFE_UINT T49.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T49.Y, 0.0, |
| ; EG-NEXT: BFE_UINT * T50.Z, T25.X, literal.y, 1, |
| ; EG-NEXT: 4(5.605194e-45), 3(4.203895e-45) |
| ; EG-NEXT: BFE_UINT T50.X, T25.X, literal.x, 1, |
| ; EG-NEXT: MOV T50.Y, 0.0, |
| ; EG-NEXT: BFE_UINT T25.Z, T25.X, 1, 1, |
| ; EG-NEXT: AND_INT * T25.X, T25.X, 1, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: MOV T25.Y, 0.0, |
| ; EG-NEXT: MOV T19.W, 0.0, |
| ; EG-NEXT: MOV * T20.W, 0.0, |
| ; EG-NEXT: MOV T21.W, 0.0, |
| ; EG-NEXT: MOV * T22.W, 0.0, |
| ; EG-NEXT: MOV T23.W, 0.0, |
| ; EG-NEXT: MOV * T24.W, 0.0, |
| ; EG-NEXT: MOV T26.W, 0.0, |
| ; EG-NEXT: MOV * T27.W, 0.0, |
| ; EG-NEXT: MOV T28.W, 0.0, |
| ; EG-NEXT: MOV * T29.W, 0.0, |
| ; EG-NEXT: MOV T30.W, 0.0, |
| ; EG-NEXT: MOV * T31.W, 0.0, |
| ; EG-NEXT: MOV T32.W, 0.0, |
| ; EG-NEXT: MOV * T33.W, 0.0, |
| ; EG-NEXT: MOV T34.W, 0.0, |
| ; EG-NEXT: MOV * T35.W, 0.0, |
| ; EG-NEXT: MOV T36.W, 0.0, |
| ; EG-NEXT: MOV * T37.W, 0.0, |
| ; EG-NEXT: MOV T38.W, 0.0, |
| ; EG-NEXT: MOV * T39.W, 0.0, |
| ; EG-NEXT: MOV T40.W, 0.0, |
| ; EG-NEXT: MOV * T41.W, 0.0, |
| ; EG-NEXT: MOV T42.W, 0.0, |
| ; EG-NEXT: MOV * T43.W, 0.0, |
| ; EG-NEXT: MOV T44.W, 0.0, |
| ; EG-NEXT: MOV * T45.W, 0.0, |
| ; EG-NEXT: MOV T46.W, 0.0, |
| ; EG-NEXT: MOV * T47.W, 0.0, |
| ; EG-NEXT: MOV T48.W, 0.0, |
| ; EG-NEXT: MOV * T49.W, 0.0, |
| ; EG-NEXT: MOV T50.W, 0.0, |
| ; EG-NEXT: MOV * T25.W, 0.0, |
| ; EG-NEXT: LSHR T51.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44) |
| ; EG-NEXT: LSHR T52.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44) |
| ; EG-NEXT: LSHR T53.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44) |
| ; EG-NEXT: LSHR T54.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 64(8.968310e-44) |
| ; EG-NEXT: LSHR T55.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 80(1.121039e-43) |
| ; EG-NEXT: LSHR T56.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 96(1.345247e-43) |
| ; EG-NEXT: LSHR T57.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 112(1.569454e-43) |
| ; EG-NEXT: LSHR T58.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 128(1.793662e-43) |
| ; EG-NEXT: LSHR T59.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 144(2.017870e-43) |
| ; EG-NEXT: LSHR T60.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 160(2.242078e-43) |
| ; EG-NEXT: LSHR T61.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 176(2.466285e-43) |
| ; EG-NEXT: LSHR * T62.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: ALU clause starting at 237: |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 192(2.690493e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T63.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 208(2.914701e-43) |
| ; EG-NEXT: LSHR T64.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 224(3.138909e-43) |
| ; EG-NEXT: LSHR T65.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 240(3.363116e-43) |
| ; EG-NEXT: LSHR T66.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 256(3.587324e-43) |
| ; EG-NEXT: LSHR T67.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 272(3.811532e-43) |
| ; EG-NEXT: LSHR T68.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 288(4.035740e-43) |
| ; EG-NEXT: LSHR T69.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 304(4.259947e-43) |
| ; EG-NEXT: LSHR T70.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 320(4.484155e-43) |
| ; EG-NEXT: LSHR T71.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 336(4.708363e-43) |
| ; EG-NEXT: LSHR T72.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 352(4.932571e-43) |
| ; EG-NEXT: LSHR T73.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 368(5.156778e-43) |
| ; EG-NEXT: LSHR T74.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 384(5.380986e-43) |
| ; EG-NEXT: LSHR T75.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 400(5.605194e-43) |
| ; EG-NEXT: LSHR T76.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 416(5.829402e-43) |
| ; EG-NEXT: LSHR T77.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 432(6.053609e-43) |
| ; EG-NEXT: LSHR T78.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 448(6.277817e-43) |
| ; EG-NEXT: LSHR T79.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 464(6.502025e-43) |
| ; EG-NEXT: LSHR T80.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 480(6.726233e-43) |
| ; EG-NEXT: LSHR T81.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 496(6.950440e-43) |
| ; EG-NEXT: LSHR * T82.X, PV.W, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_zextload_v64i1_to_v64i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x10014 |
| ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) |
| ; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s4 |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x10015 |
| ; GFX12-NEXT: s_lshr_b32 s4, s3, 31 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_dual_mov_b32 v2, s5 :: v_dual_mov_b32 v3, v1 |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x1001e |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:416 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x1001d |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x1001c |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:496 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x1001b |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x1001a |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:480 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x10019 |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x10018 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:464 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x10017 |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x10016 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:448 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x10013 |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x10012 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:432 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x10011 |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x10010 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:400 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x1000f |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x1000e |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:384 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x1000d |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x1000c |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:368 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x1000b |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x1000a |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:352 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x10009 |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x10008 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:336 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x10007 |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x10006 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:320 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x10005 |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x10004 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:304 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x10003 |
| ; GFX12-NEXT: s_bfe_u32 s5, s3, 0x10002 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:288 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s4, s3, 0x10001 |
| ; GFX12-NEXT: s_and_b32 s3, s3, 1 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:272 |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s3 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_lshr_b32 s3, s2, 31 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x1001e |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:256 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1001d |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x1001c |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:240 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1001b |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x1001a |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:224 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10019 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10018 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:208 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10017 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10016 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:192 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10014 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10015 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:176 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s3 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s4 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10013 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10012 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:160 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10011 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10010 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:144 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1000f |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x1000e |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:128 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1000d |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x1000c |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:112 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x1000b |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x1000a |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:96 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10009 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10008 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:80 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10007 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10006 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:64 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10005 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10004 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:48 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10003 |
| ; GFX12-NEXT: s_bfe_u32 s4, s2, 0x10002 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:32 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: s_bfe_u32 s3, s2, 0x10001 |
| ; GFX12-NEXT: s_and_b32 s2, s2, 1 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:16 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v0, s2 |
| ; GFX12-NEXT: v_mov_b32_e32 v2, s3 |
| ; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <64 x i1>, ptr addrspace(4) %in |
| %ext = zext <64 x i1> %load to <64 x i64> |
| store <64 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| define amdgpu_kernel void @constant_sextload_v64i1_to_v64i64(ptr addrspace(1) %out, ptr addrspace(4) nocapture %in) #0 { |
| ; GFX6-LABEL: constant_sextload_v64i1_to_v64i64: |
| ; GFX6: ; %bb.0: |
| ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0 |
| ; GFX6-NEXT: s_mov_b32 s3, 0xf000 |
| ; GFX6-NEXT: s_mov_b32 s2, -1 |
| ; GFX6-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX6-NEXT: s_lshr_b32 s42, s5, 30 |
| ; GFX6-NEXT: s_lshr_b32 s36, s5, 28 |
| ; GFX6-NEXT: s_lshr_b32 s38, s5, 29 |
| ; GFX6-NEXT: s_lshr_b32 s30, s5, 26 |
| ; GFX6-NEXT: s_lshr_b32 s34, s5, 27 |
| ; GFX6-NEXT: s_lshr_b32 s26, s5, 24 |
| ; GFX6-NEXT: s_lshr_b32 s28, s5, 25 |
| ; GFX6-NEXT: s_lshr_b32 s22, s5, 22 |
| ; GFX6-NEXT: s_lshr_b32 s24, s5, 23 |
| ; GFX6-NEXT: s_lshr_b32 s18, s5, 20 |
| ; GFX6-NEXT: s_lshr_b32 s20, s5, 21 |
| ; GFX6-NEXT: s_lshr_b32 s14, s5, 18 |
| ; GFX6-NEXT: s_lshr_b32 s16, s5, 19 |
| ; GFX6-NEXT: s_lshr_b32 s10, s5, 16 |
| ; GFX6-NEXT: s_lshr_b32 s12, s5, 17 |
| ; GFX6-NEXT: s_lshr_b32 s6, s5, 14 |
| ; GFX6-NEXT: s_lshr_b32 s8, s5, 15 |
| ; GFX6-NEXT: s_mov_b32 s40, s5 |
| ; GFX6-NEXT: s_ashr_i32 s7, s5, 31 |
| ; GFX6-NEXT: s_bfe_i64 s[44:45], s[40:41], 0x10000 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s7 |
| ; GFX6-NEXT: s_lshr_b32 s40, s5, 12 |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s44 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s45 |
| ; GFX6-NEXT: s_bfe_i64 s[44:45], s[4:5], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[42:43], s[42:43], 0x10000 |
| ; GFX6-NEXT: v_mov_b32_e32 v6, s44 |
| ; GFX6-NEXT: v_mov_b32_e32 v7, s45 |
| ; GFX6-NEXT: s_lshr_b32 s44, s5, 13 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s42 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s43 |
| ; GFX6-NEXT: s_lshr_b32 s42, s5, 10 |
| ; GFX6-NEXT: s_bfe_i64 s[36:37], s[36:37], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[38:39], s[38:39], 0x10000 |
| ; GFX6-NEXT: v_mov_b32_e32 v8, s36 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, s37 |
| ; GFX6-NEXT: s_lshr_b32 s36, s5, 11 |
| ; GFX6-NEXT: v_mov_b32_e32 v10, s38 |
| ; GFX6-NEXT: v_mov_b32_e32 v11, s39 |
| ; GFX6-NEXT: s_lshr_b32 s38, s5, 8 |
| ; GFX6-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[34:35], s[34:35], 0x10000 |
| ; GFX6-NEXT: v_mov_b32_e32 v12, s30 |
| ; GFX6-NEXT: v_mov_b32_e32 v13, s31 |
| ; GFX6-NEXT: s_lshr_b32 s30, s5, 9 |
| ; GFX6-NEXT: v_mov_b32_e32 v14, s34 |
| ; GFX6-NEXT: v_mov_b32_e32 v15, s35 |
| ; GFX6-NEXT: s_lshr_b32 s34, s5, 6 |
| ; GFX6-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x10000 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:496 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s26 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s27 |
| ; GFX6-NEXT: s_lshr_b32 s26, s5, 7 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s28 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s29 |
| ; GFX6-NEXT: s_lshr_b32 s28, s5, 4 |
| ; GFX6-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:480 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v8, s22 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, s23 |
| ; GFX6-NEXT: s_lshr_b32 s22, s5, 5 |
| ; GFX6-NEXT: v_mov_b32_e32 v10, s24 |
| ; GFX6-NEXT: v_mov_b32_e32 v11, s25 |
| ; GFX6-NEXT: s_lshr_b32 s24, s5, 2 |
| ; GFX6-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:464 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v12, s18 |
| ; GFX6-NEXT: v_mov_b32_e32 v13, s19 |
| ; GFX6-NEXT: s_lshr_b32 s18, s5, 3 |
| ; GFX6-NEXT: v_mov_b32_e32 v14, s20 |
| ; GFX6-NEXT: v_mov_b32_e32 v15, s21 |
| ; GFX6-NEXT: s_lshr_b32 s20, s5, 1 |
| ; GFX6-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:448 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s14 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s15 |
| ; GFX6-NEXT: s_lshr_b32 s14, s4, 30 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s16 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s17 |
| ; GFX6-NEXT: s_lshr_b32 s16, s4, 31 |
| ; GFX6-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:432 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v8, s10 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, s11 |
| ; GFX6-NEXT: s_lshr_b32 s10, s4, 28 |
| ; GFX6-NEXT: v_mov_b32_e32 v10, s12 |
| ; GFX6-NEXT: v_mov_b32_e32 v11, s13 |
| ; GFX6-NEXT: s_lshr_b32 s12, s4, 29 |
| ; GFX6-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:416 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v12, s6 |
| ; GFX6-NEXT: v_mov_b32_e32 v13, s7 |
| ; GFX6-NEXT: s_lshr_b32 s46, s4, 26 |
| ; GFX6-NEXT: v_mov_b32_e32 v14, s8 |
| ; GFX6-NEXT: v_mov_b32_e32 v15, s9 |
| ; GFX6-NEXT: s_lshr_b32 s8, s4, 27 |
| ; GFX6-NEXT: s_bfe_i64 s[6:7], s[44:45], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[40:41], s[40:41], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:400 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s40 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s41 |
| ; GFX6-NEXT: s_lshr_b32 s40, s4, 24 |
| ; GFX6-NEXT: v_mov_b32_e32 v4, s6 |
| ; GFX6-NEXT: v_mov_b32_e32 v5, s7 |
| ; GFX6-NEXT: s_lshr_b32 s44, s4, 25 |
| ; GFX6-NEXT: s_bfe_i64 s[6:7], s[36:37], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[36:37], s[42:43], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:384 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v8, s36 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, s37 |
| ; GFX6-NEXT: s_lshr_b32 s36, s4, 22 |
| ; GFX6-NEXT: v_mov_b32_e32 v10, s6 |
| ; GFX6-NEXT: v_mov_b32_e32 v11, s7 |
| ; GFX6-NEXT: s_lshr_b32 s42, s4, 23 |
| ; GFX6-NEXT: s_bfe_i64 s[6:7], s[30:31], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[30:31], s[38:39], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:368 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v12, s30 |
| ; GFX6-NEXT: v_mov_b32_e32 v13, s31 |
| ; GFX6-NEXT: s_lshr_b32 s30, s4, 20 |
| ; GFX6-NEXT: v_mov_b32_e32 v14, s6 |
| ; GFX6-NEXT: v_mov_b32_e32 v15, s7 |
| ; GFX6-NEXT: s_lshr_b32 s6, s4, 21 |
| ; GFX6-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[34:35], s[34:35], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[2:5], off, s[0:3], 0 offset:352 |
| ; GFX6-NEXT: v_mov_b32_e32 v16, s34 |
| ; GFX6-NEXT: v_mov_b32_e32 v17, s35 |
| ; GFX6-NEXT: s_lshr_b32 s34, s4, 18 |
| ; GFX6-NEXT: v_mov_b32_e32 v18, s26 |
| ; GFX6-NEXT: v_mov_b32_e32 v19, s27 |
| ; GFX6-NEXT: s_lshr_b32 s26, s4, 19 |
| ; GFX6-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:336 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v8, s28 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, s29 |
| ; GFX6-NEXT: s_lshr_b32 s28, s4, 16 |
| ; GFX6-NEXT: v_mov_b32_e32 v10, s22 |
| ; GFX6-NEXT: v_mov_b32_e32 v11, s23 |
| ; GFX6-NEXT: s_lshr_b32 s22, s4, 17 |
| ; GFX6-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:320 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v12, s24 |
| ; GFX6-NEXT: v_mov_b32_e32 v13, s25 |
| ; GFX6-NEXT: s_lshr_b32 s24, s4, 14 |
| ; GFX6-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x10000 |
| ; GFX6-NEXT: v_mov_b32_e32 v14, s18 |
| ; GFX6-NEXT: v_mov_b32_e32 v15, s19 |
| ; GFX6-NEXT: s_lshr_b32 s18, s4, 15 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s20 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s21 |
| ; GFX6-NEXT: s_lshr_b32 s20, s4, 12 |
| ; GFX6-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:304 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v16, s14 |
| ; GFX6-NEXT: v_mov_b32_e32 v17, s15 |
| ; GFX6-NEXT: s_lshr_b32 s14, s4, 13 |
| ; GFX6-NEXT: v_mov_b32_e32 v18, s16 |
| ; GFX6-NEXT: v_mov_b32_e32 v19, s17 |
| ; GFX6-NEXT: s_lshr_b32 s16, s4, 10 |
| ; GFX6-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:288 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v8, s10 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, s11 |
| ; GFX6-NEXT: s_lshr_b32 s10, s4, 11 |
| ; GFX6-NEXT: v_mov_b32_e32 v10, s12 |
| ; GFX6-NEXT: v_mov_b32_e32 v11, s13 |
| ; GFX6-NEXT: s_lshr_b32 s12, s4, 8 |
| ; GFX6-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[38:39], s[46:47], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:272 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v12, s38 |
| ; GFX6-NEXT: v_mov_b32_e32 v13, s39 |
| ; GFX6-NEXT: s_lshr_b32 s38, s4, 9 |
| ; GFX6-NEXT: v_mov_b32_e32 v14, s8 |
| ; GFX6-NEXT: v_mov_b32_e32 v15, s9 |
| ; GFX6-NEXT: s_lshr_b32 s8, s4, 6 |
| ; GFX6-NEXT: s_bfe_i64 s[44:45], s[44:45], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[40:41], s[40:41], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:256 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s40 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s41 |
| ; GFX6-NEXT: s_lshr_b32 s40, s4, 7 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s44 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s45 |
| ; GFX6-NEXT: s_lshr_b32 s44, s4, 4 |
| ; GFX6-NEXT: s_bfe_i64 s[42:43], s[42:43], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[36:37], s[36:37], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:240 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v16, s36 |
| ; GFX6-NEXT: v_mov_b32_e32 v17, s37 |
| ; GFX6-NEXT: s_lshr_b32 s36, s4, 5 |
| ; GFX6-NEXT: v_mov_b32_e32 v18, s42 |
| ; GFX6-NEXT: v_mov_b32_e32 v19, s43 |
| ; GFX6-NEXT: s_lshr_b32 s42, s4, 2 |
| ; GFX6-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:224 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v8, s30 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, s31 |
| ; GFX6-NEXT: s_lshr_b32 s30, s4, 3 |
| ; GFX6-NEXT: s_lshr_b32 s4, s4, 1 |
| ; GFX6-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[42:43], s[42:43], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[36:37], s[36:37], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[44:45], s[44:45], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[40:41], s[40:41], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[38:39], s[38:39], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[34:35], s[34:35], 0x10000 |
| ; GFX6-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x10000 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:208 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:192 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:176 |
| ; GFX6-NEXT: v_mov_b32_e32 v10, s6 |
| ; GFX6-NEXT: v_mov_b32_e32 v11, s7 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:160 |
| ; GFX6-NEXT: s_waitcnt expcnt(2) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s34 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s35 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s26 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s27 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:144 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s28 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s29 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s22 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s23 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:128 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s24 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s25 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:112 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s20 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s21 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s14 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s15 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:96 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s10 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s11 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s12 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s13 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s8 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s9 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s40 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s41 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s44 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s45 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s36 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s37 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32 |
| ; GFX6-NEXT: s_waitcnt expcnt(0) |
| ; GFX6-NEXT: v_mov_b32_e32 v0, s42 |
| ; GFX6-NEXT: v_mov_b32_e32 v1, s43 |
| ; GFX6-NEXT: v_mov_b32_e32 v2, s30 |
| ; GFX6-NEXT: v_mov_b32_e32 v3, s31 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 |
| ; GFX6-NEXT: v_mov_b32_e32 v8, s4 |
| ; GFX6-NEXT: v_mov_b32_e32 v9, s5 |
| ; GFX6-NEXT: buffer_store_dwordx4 v[6:9], off, s[0:3], 0 |
| ; GFX6-NEXT: s_endpgm |
| ; |
| ; GFX8-LABEL: constant_sextload_v64i1_to_v64i64: |
| ; GFX8: ; %bb.0: |
| ; GFX8-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x24 |
| ; GFX8-NEXT: ; implicit-def: $vgpr62 : SGPR spill to VGPR lane |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_load_dwordx2 s[2:3], s[10:11], 0x0 |
| ; GFX8-NEXT: s_waitcnt lgkmcnt(0) |
| ; GFX8-NEXT: s_lshr_b32 s0, s3, 8 |
| ; GFX8-NEXT: s_lshr_b32 s48, s3, 15 |
| ; GFX8-NEXT: v_writelane_b32 v62, s0, 0 |
| ; GFX8-NEXT: s_lshr_b32 s74, s3, 30 |
| ; GFX8-NEXT: s_lshr_b32 s30, s3, 31 |
| ; GFX8-NEXT: s_lshr_b32 s72, s3, 28 |
| ; GFX8-NEXT: s_lshr_b32 s34, s3, 29 |
| ; GFX8-NEXT: s_lshr_b32 s70, s3, 26 |
| ; GFX8-NEXT: s_lshr_b32 s36, s3, 27 |
| ; GFX8-NEXT: s_lshr_b32 s68, s3, 24 |
| ; GFX8-NEXT: s_lshr_b32 s38, s3, 25 |
| ; GFX8-NEXT: s_lshr_b32 s64, s3, 22 |
| ; GFX8-NEXT: s_lshr_b32 s40, s3, 23 |
| ; GFX8-NEXT: s_lshr_b32 s60, s3, 20 |
| ; GFX8-NEXT: s_lshr_b32 s42, s3, 21 |
| ; GFX8-NEXT: s_lshr_b32 s66, s3, 18 |
| ; GFX8-NEXT: s_lshr_b32 s44, s3, 19 |
| ; GFX8-NEXT: s_lshr_b32 s56, s3, 16 |
| ; GFX8-NEXT: s_lshr_b32 s46, s3, 17 |
| ; GFX8-NEXT: s_lshr_b32 s58, s3, 14 |
| ; GFX8-NEXT: s_lshr_b32 s62, s3, 12 |
| ; GFX8-NEXT: s_lshr_b32 s54, s3, 10 |
| ; GFX8-NEXT: v_writelane_b32 v62, s1, 1 |
| ; GFX8-NEXT: s_lshr_b32 s0, s3, 9 |
| ; GFX8-NEXT: s_bfe_i64 s[48:49], s[48:49], 0x10000 |
| ; GFX8-NEXT: s_lshr_b32 s52, s3, 11 |
| ; GFX8-NEXT: v_writelane_b32 v62, s0, 2 |
| ; GFX8-NEXT: s_bfe_i64 s[60:61], s[60:61], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[64:65], s[64:65], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[68:69], s[68:69], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[70:71], s[70:71], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[72:73], s[72:73], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[74:75], s[74:75], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[66:67], s[66:67], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[62:63], s[62:63], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[58:59], s[58:59], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[56:57], s[56:57], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[54:55], s[54:55], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[46:47], s[46:47], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[44:45], s[44:45], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[42:43], s[42:43], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[40:41], s[40:41], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[38:39], s[38:39], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[36:37], s[36:37], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[34:35], s[34:35], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x10000 |
| ; GFX8-NEXT: v_mov_b32_e32 v34, s48 |
| ; GFX8-NEXT: s_lshr_b32 s48, s2, 1 |
| ; GFX8-NEXT: s_lshr_b32 s50, s3, 13 |
| ; GFX8-NEXT: v_writelane_b32 v62, s1, 3 |
| ; GFX8-NEXT: s_lshr_b32 s6, s3, 6 |
| ; GFX8-NEXT: s_lshr_b32 s10, s3, 7 |
| ; GFX8-NEXT: s_lshr_b32 s12, s3, 4 |
| ; GFX8-NEXT: s_lshr_b32 s14, s3, 5 |
| ; GFX8-NEXT: s_lshr_b32 s16, s3, 2 |
| ; GFX8-NEXT: s_lshr_b32 s18, s3, 3 |
| ; GFX8-NEXT: s_lshr_b32 s20, s3, 1 |
| ; GFX8-NEXT: s_mov_b32 s22, s3 |
| ; GFX8-NEXT: s_lshr_b32 s24, s2, 30 |
| ; GFX8-NEXT: s_lshr_b32 s26, s2, 31 |
| ; GFX8-NEXT: s_lshr_b32 s28, s2, 28 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s74 |
| ; GFX8-NEXT: v_mov_b32_e32 v12, s72 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s70 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, s68 |
| ; GFX8-NEXT: v_mov_b32_e32 v16, s64 |
| ; GFX8-NEXT: v_mov_b32_e32 v20, s60 |
| ; GFX8-NEXT: v_mov_b32_e32 v24, s66 |
| ; GFX8-NEXT: v_mov_b32_e32 v28, s56 |
| ; GFX8-NEXT: v_mov_b32_e32 v32, s58 |
| ; GFX8-NEXT: v_mov_b32_e32 v36, s62 |
| ; GFX8-NEXT: s_lshr_b32 s86, s2, 29 |
| ; GFX8-NEXT: v_mov_b32_e32 v40, s54 |
| ; GFX8-NEXT: s_lshr_b32 s84, s2, 26 |
| ; GFX8-NEXT: s_lshr_b32 s82, s2, 27 |
| ; GFX8-NEXT: s_bfe_i64 vcc, s[52:53], 0x10000 |
| ; GFX8-NEXT: s_lshr_b32 s80, s2, 24 |
| ; GFX8-NEXT: v_mov_b32_e32 v6, s30 |
| ; GFX8-NEXT: v_mov_b32_e32 v7, s31 |
| ; GFX8-NEXT: s_lshr_b32 s78, s2, 25 |
| ; GFX8-NEXT: s_lshr_b32 s76, s2, 22 |
| ; GFX8-NEXT: v_mov_b32_e32 v14, s34 |
| ; GFX8-NEXT: s_lshr_b32 s74, s2, 23 |
| ; GFX8-NEXT: s_lshr_b32 s72, s2, 20 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s36 |
| ; GFX8-NEXT: s_lshr_b32 s70, s2, 21 |
| ; GFX8-NEXT: s_lshr_b32 s68, s2, 18 |
| ; GFX8-NEXT: v_mov_b32_e32 v10, s38 |
| ; GFX8-NEXT: s_lshr_b32 s66, s2, 19 |
| ; GFX8-NEXT: s_lshr_b32 s64, s2, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v18, s40 |
| ; GFX8-NEXT: s_lshr_b32 s62, s2, 17 |
| ; GFX8-NEXT: s_lshr_b32 s60, s2, 14 |
| ; GFX8-NEXT: v_mov_b32_e32 v22, s42 |
| ; GFX8-NEXT: s_lshr_b32 s58, s2, 15 |
| ; GFX8-NEXT: s_lshr_b32 s56, s2, 12 |
| ; GFX8-NEXT: v_mov_b32_e32 v26, s44 |
| ; GFX8-NEXT: s_lshr_b32 s54, s2, 13 |
| ; GFX8-NEXT: s_lshr_b32 s52, s2, 10 |
| ; GFX8-NEXT: v_mov_b32_e32 v30, s46 |
| ; GFX8-NEXT: s_lshr_b32 s4, s2, 11 |
| ; GFX8-NEXT: s_lshr_b32 s0, s2, 8 |
| ; GFX8-NEXT: s_lshr_b32 s46, s2, 9 |
| ; GFX8-NEXT: s_lshr_b32 s44, s2, 6 |
| ; GFX8-NEXT: s_lshr_b32 s42, s2, 7 |
| ; GFX8-NEXT: s_lshr_b32 s40, s2, 4 |
| ; GFX8-NEXT: s_lshr_b32 s38, s2, 5 |
| ; GFX8-NEXT: s_lshr_b32 s36, s2, 2 |
| ; GFX8-NEXT: s_lshr_b32 s34, s2, 3 |
| ; GFX8-NEXT: s_bfe_i64 s[30:31], s[2:3], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[2:3], s[48:49], 0x10000 |
| ; GFX8-NEXT: v_writelane_b32 v62, s2, 4 |
| ; GFX8-NEXT: v_writelane_b32 v62, s3, 5 |
| ; GFX8-NEXT: v_readlane_b32 s2, v62, 2 |
| ; GFX8-NEXT: s_bfe_i64 s[50:51], s[50:51], 0x10000 |
| ; GFX8-NEXT: v_readlane_b32 s3, v62, 3 |
| ; GFX8-NEXT: v_mov_b32_e32 v38, s50 |
| ; GFX8-NEXT: v_mov_b32_e32 v39, s51 |
| ; GFX8-NEXT: s_bfe_i64 s[50:51], s[4:5], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[4:5], s[6:7], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[6:7], s[2:3], 0x10000 |
| ; GFX8-NEXT: v_readlane_b32 s2, v62, 0 |
| ; GFX8-NEXT: v_readlane_b32 s3, v62, 1 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s75 |
| ; GFX8-NEXT: v_mov_b32_e32 v13, s73 |
| ; GFX8-NEXT: v_mov_b32_e32 v15, s35 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s71 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s37 |
| ; GFX8-NEXT: v_mov_b32_e32 v9, s69 |
| ; GFX8-NEXT: v_mov_b32_e32 v11, s39 |
| ; GFX8-NEXT: v_mov_b32_e32 v17, s65 |
| ; GFX8-NEXT: v_mov_b32_e32 v19, s41 |
| ; GFX8-NEXT: v_mov_b32_e32 v21, s61 |
| ; GFX8-NEXT: v_mov_b32_e32 v23, s43 |
| ; GFX8-NEXT: v_mov_b32_e32 v25, s67 |
| ; GFX8-NEXT: v_mov_b32_e32 v27, s45 |
| ; GFX8-NEXT: v_mov_b32_e32 v29, s57 |
| ; GFX8-NEXT: v_mov_b32_e32 v31, s47 |
| ; GFX8-NEXT: v_mov_b32_e32 v33, s59 |
| ; GFX8-NEXT: v_mov_b32_e32 v35, s49 |
| ; GFX8-NEXT: v_mov_b32_e32 v37, s63 |
| ; GFX8-NEXT: v_mov_b32_e32 v41, s55 |
| ; GFX8-NEXT: s_bfe_i64 s[34:35], s[34:35], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[36:37], s[36:37], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[38:39], s[38:39], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[40:41], s[40:41], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[42:43], s[42:43], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[44:45], s[44:45], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[46:47], s[46:47], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[48:49], s[0:1], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[52:53], s[52:53], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[54:55], s[54:55], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[56:57], s[56:57], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[58:59], s[58:59], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[60:61], s[60:61], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[62:63], s[62:63], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[64:65], s[64:65], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[66:67], s[66:67], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[68:69], s[68:69], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[70:71], s[70:71], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[72:73], s[72:73], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[74:75], s[74:75], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[76:77], s[76:77], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[78:79], s[78:79], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[80:81], s[80:81], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[82:83], s[82:83], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[84:85], s[84:85], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[86:87], s[86:87], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[0:1], s[10:11], 0x10000 |
| ; GFX8-NEXT: s_bfe_i64 s[10:11], s[2:3], 0x10000 |
| ; GFX8-NEXT: s_add_u32 s2, s8, 0x1f0 |
| ; GFX8-NEXT: s_addc_u32 s3, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v43, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v42, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s8, 0x1e0 |
| ; GFX8-NEXT: s_addc_u32 s3, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v45, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v44, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s8, 0x1d0 |
| ; GFX8-NEXT: s_addc_u32 s3, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v47, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v46, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s8, 0x1c0 |
| ; GFX8-NEXT: s_addc_u32 s3, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v49, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v48, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s8, 0x1b0 |
| ; GFX8-NEXT: s_addc_u32 s3, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v51, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v50, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s8, 0x1a0 |
| ; GFX8-NEXT: s_addc_u32 s3, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v53, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v52, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s8, 0x190 |
| ; GFX8-NEXT: s_addc_u32 s3, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v55, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v54, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s8, 0x180 |
| ; GFX8-NEXT: s_addc_u32 s3, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v57, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v56, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s8, 0x170 |
| ; GFX8-NEXT: s_addc_u32 s3, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v59, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v58, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s8, 0x160 |
| ; GFX8-NEXT: s_addc_u32 s3, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v61, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v60, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s8, 0x150 |
| ; GFX8-NEXT: s_addc_u32 s3, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[44:45], v[12:15] |
| ; GFX8-NEXT: flat_store_dwordx4 v[46:47], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v13, s3 |
| ; GFX8-NEXT: v_mov_b32_e32 v12, s2 |
| ; GFX8-NEXT: s_add_u32 s2, s8, 0x140 |
| ; GFX8-NEXT: s_addc_u32 s3, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0x130 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s1 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[42:43], v[4:7] |
| ; GFX8-NEXT: flat_store_dwordx4 v[48:49], v[8:11] |
| ; GFX8-NEXT: flat_store_dwordx4 v[50:51], v[16:19] |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s10 |
| ; GFX8-NEXT: v_mov_b32_e32 v17, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v16, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0x120 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v19, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v18, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0x110 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s11 |
| ; GFX8-NEXT: v_mov_b32_e32 v15, s3 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: v_mov_b32_e32 v42, vcc_lo |
| ; GFX8-NEXT: v_mov_b32_e32 v43, vcc_hi |
| ; GFX8-NEXT: v_mov_b32_e32 v14, s2 |
| ; GFX8-NEXT: v_mov_b32_e32 v6, s6 |
| ; GFX8-NEXT: v_mov_b32_e32 v7, s7 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s4 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s5 |
| ; GFX8-NEXT: v_mov_b32_e32 v8, s12 |
| ; GFX8-NEXT: flat_store_dwordx4 v[52:53], v[20:23] |
| ; GFX8-NEXT: v_mov_b32_e32 v9, s13 |
| ; GFX8-NEXT: flat_store_dwordx4 v[54:55], v[24:27] |
| ; GFX8-NEXT: v_mov_b32_e32 v10, s14 |
| ; GFX8-NEXT: v_mov_b32_e32 v11, s15 |
| ; GFX8-NEXT: flat_store_dwordx4 v[56:57], v[28:31] |
| ; GFX8-NEXT: flat_store_dwordx4 v[58:59], v[32:35] |
| ; GFX8-NEXT: flat_store_dwordx4 v[60:61], v[36:39] |
| ; GFX8-NEXT: flat_store_dwordx4 v[12:13], v[40:43] |
| ; GFX8-NEXT: flat_store_dwordx4 v[14:15], v[4:7] |
| ; GFX8-NEXT: flat_store_dwordx4 v[16:17], v[0:3] |
| ; GFX8-NEXT: flat_store_dwordx4 v[18:19], v[8:11] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0x100 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s16 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s17 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s18 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s19 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0xf0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s22 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s23 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s20 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s21 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0xe0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s24 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s25 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s26 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s27 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0xd0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s28 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s29 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s86 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s87 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0xc0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s84 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s85 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s82 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s83 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0xb0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s80 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s81 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s78 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s79 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0xa0 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s76 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s77 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s74 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s75 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0x90 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s72 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s73 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s70 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s71 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0x80 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s68 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s69 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s66 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s67 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0x70 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s64 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s65 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s62 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s63 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0x60 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s60 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s61 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s58 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s59 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 0x50 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s56 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s57 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s54 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s55 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 64 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s52 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s53 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s50 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s51 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 48 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s48 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s49 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s46 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s47 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 32 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s44 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s45 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s42 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s43 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: s_add_u32 s0, s8, 16 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s40 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s41 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s38 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s39 |
| ; GFX8-NEXT: s_addc_u32 s1, s9, 0 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s36 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s37 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s34 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s35 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s0 |
| ; GFX8-NEXT: v_readlane_b32 s0, v62, 4 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: v_readlane_b32 s1, v62, 5 |
| ; GFX8-NEXT: v_mov_b32_e32 v4, s8 |
| ; GFX8-NEXT: v_mov_b32_e32 v0, s30 |
| ; GFX8-NEXT: v_mov_b32_e32 v1, s31 |
| ; GFX8-NEXT: v_mov_b32_e32 v2, s0 |
| ; GFX8-NEXT: v_mov_b32_e32 v3, s1 |
| ; GFX8-NEXT: v_mov_b32_e32 v5, s9 |
| ; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3] |
| ; GFX8-NEXT: s_endpgm |
| ; |
| ; EG-LABEL: constant_sextload_v64i1_to_v64i64: |
| ; EG: ; %bb.0: |
| ; EG-NEXT: ALU 22, @40, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: TEX 0 @38 |
| ; EG-NEXT: ALU 89, @63, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: ALU 99, @153, KC0[], KC1[] |
| ; EG-NEXT: ALU 107, @253, KC0[CB0:0-32], KC1[] |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T50.XYZW, T82.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T80.XYZW, T81.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T51.XYZW, T73.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T79.XYZW, T48.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T52.XYZW, T47.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T78.XYZW, T46.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T53.XYZW, T45.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T77.XYZW, T44.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T54.XYZW, T43.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T76.XYZW, T42.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T55.XYZW, T41.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T75.XYZW, T39.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T56.XYZW, T38.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T74.XYZW, T37.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T57.XYZW, T36.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T66.XYZW, T35.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T58.XYZW, T34.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T72.XYZW, T33.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T59.XYZW, T32.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T71.XYZW, T31.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T60.XYZW, T30.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T70.XYZW, T29.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T61.XYZW, T28.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T69.XYZW, T27.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T62.XYZW, T26.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T68.XYZW, T25.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T63.XYZW, T24.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T40.XYZW, T23.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T64.XYZW, T22.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T67.XYZW, T21.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T65.XYZW, T20.X, 0 |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T49.XYZW, T19.X, 1 |
| ; EG-NEXT: CF_END |
| ; EG-NEXT: Fetch clause starting at 38: |
| ; EG-NEXT: VTX_READ_64 T40.XY, T26.X, 0, #1 |
| ; EG-NEXT: ALU clause starting at 40: |
| ; EG-NEXT: LSHR T19.X, KC0[2].Y, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44) |
| ; EG-NEXT: LSHR T20.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44) |
| ; EG-NEXT: LSHR T21.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44) |
| ; EG-NEXT: LSHR T22.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 64(8.968310e-44) |
| ; EG-NEXT: LSHR T23.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 80(1.121039e-43) |
| ; EG-NEXT: LSHR T24.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 96(1.345247e-43) |
| ; EG-NEXT: LSHR T25.X, PV.W, literal.x, |
| ; EG-NEXT: MOV * T26.X, KC0[2].Z, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 112(1.569454e-43), 0(0.000000e+00) |
| ; EG-NEXT: ALU clause starting at 63: |
| ; EG-NEXT: LSHR T26.X, T0.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 128(1.793662e-43) |
| ; EG-NEXT: LSHR T27.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 144(2.017870e-43) |
| ; EG-NEXT: LSHR T28.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 160(2.242078e-43) |
| ; EG-NEXT: LSHR T29.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 176(2.466285e-43) |
| ; EG-NEXT: LSHR T30.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 192(2.690493e-43) |
| ; EG-NEXT: LSHR T31.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 208(2.914701e-43) |
| ; EG-NEXT: LSHR T32.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 224(3.138909e-43) |
| ; EG-NEXT: LSHR T33.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 240(3.363116e-43) |
| ; EG-NEXT: LSHR T34.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 256(3.587324e-43) |
| ; EG-NEXT: LSHR T35.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 272(3.811532e-43) |
| ; EG-NEXT: LSHR T36.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 288(4.035740e-43) |
| ; EG-NEXT: LSHR T37.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 304(4.259947e-43) |
| ; EG-NEXT: LSHR T38.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 320(4.484155e-43) |
| ; EG-NEXT: LSHR T39.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 336(4.708363e-43) |
| ; EG-NEXT: LSHR T41.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y, |
| ; EG-NEXT: 2(2.802597e-45), 352(4.932571e-43) |
| ; EG-NEXT: LSHR T42.X, PV.W, literal.x, |
| ; EG-NEXT: LSHR T0.Z, T40.Y, literal.y, |
| ; EG-NEXT: LSHR T0.W, T40.Y, literal.z, |
| ; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.w, |
| ; EG-NEXT: 2(2.802597e-45), 28(3.923636e-44) |
| ; EG-NEXT: 29(4.063766e-44), 368(5.156778e-43) |
| ; EG-NEXT: LSHR T43.X, PS, literal.x, |
| ; EG-NEXT: LSHR T0.Y, T40.Y, literal.y, |
| ; EG-NEXT: LSHR T1.Z, T40.Y, literal.z, |
| ; EG-NEXT: LSHR * T1.W, T40.Y, literal.w, |
| ; EG-NEXT: 2(2.802597e-45), 24(3.363116e-44) |
| ; EG-NEXT: 25(3.503246e-44), 20(2.802597e-44) |
| ; EG-NEXT: ADD_INT * T2.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 384(5.380986e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T44.X, PV.W, literal.x, |
| ; EG-NEXT: LSHR T1.Y, T40.Y, literal.y, |
| ; EG-NEXT: LSHR T2.Z, T40.Y, literal.z, |
| ; EG-NEXT: LSHR * T2.W, T40.Y, literal.w, |
| ; EG-NEXT: 2(2.802597e-45), 21(2.942727e-44) |
| ; EG-NEXT: 16(2.242078e-44), 17(2.382207e-44) |
| ; EG-NEXT: ADD_INT * T3.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 400(5.605194e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T45.X, PV.W, literal.x, |
| ; EG-NEXT: LSHR T2.Y, T40.Y, literal.y, |
| ; EG-NEXT: LSHR T3.Z, T40.Y, literal.z, |
| ; EG-NEXT: LSHR * T3.W, T40.Y, literal.w, |
| ; EG-NEXT: 2(2.802597e-45), 12(1.681558e-44) |
| ; EG-NEXT: 13(1.821688e-44), 8(1.121039e-44) |
| ; EG-NEXT: ADD_INT * T4.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 416(5.829402e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T46.X, PV.W, literal.x, |
| ; EG-NEXT: LSHR T3.Y, T40.Y, literal.y, |
| ; EG-NEXT: LSHR T4.Z, T40.Y, literal.z, |
| ; EG-NEXT: LSHR * T4.W, T40.Y, literal.w, |
| ; EG-NEXT: 2(2.802597e-45), 9(1.261169e-44) |
| ; EG-NEXT: 4(5.605194e-45), 5(7.006492e-45) |
| ; EG-NEXT: ADD_INT * T5.W, KC0[2].Y, literal.x, |
| ; EG-NEXT: 432(6.053609e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T47.X, PV.W, literal.x, |
| ; EG-NEXT: ADD_INT T4.Y, KC0[2].Y, literal.y, |
| ; EG-NEXT: LSHR T5.Z, T40.Y, 1, |
| ; EG-NEXT: LSHR T5.W, T40.X, literal.z, |
| ; EG-NEXT: ADD_INT * T6.W, KC0[2].Y, literal.w, |
| ; EG-NEXT: 2(2.802597e-45), 464(6.502025e-43) |
| ; EG-NEXT: 28(3.923636e-44), 448(6.277817e-43) |
| ; EG-NEXT: ALU clause starting at 153: |
| ; EG-NEXT: LSHR T48.X, T6.W, literal.x, |
| ; EG-NEXT: LSHR T5.Y, T40.X, literal.y, |
| ; EG-NEXT: LSHR T6.Z, T40.X, literal.z, |
| ; EG-NEXT: LSHR * T6.W, T40.X, literal.w, |
| ; EG-NEXT: 2(2.802597e-45), 29(4.063766e-44) |
| ; EG-NEXT: 24(3.363116e-44), 25(3.503246e-44) |
| ; EG-NEXT: LSHR * T7.W, T40.X, literal.x, |
| ; EG-NEXT: 20(2.802597e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T49.X, T40.X, 0.0, 1, |
| ; EG-NEXT: LSHR T6.Y, T40.X, literal.x, |
| ; EG-NEXT: ASHR T50.Z, T40.Y, literal.y, |
| ; EG-NEXT: LSHR T8.W, T40.Y, literal.z, |
| ; EG-NEXT: LSHR * T9.W, T40.Y, literal.w, |
| ; EG-NEXT: 21(2.942727e-44), 31(4.344025e-44) |
| ; EG-NEXT: 27(3.783506e-44), 30(4.203895e-44) |
| ; EG-NEXT: BFE_INT T50.X, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T7.Y, T40.X, literal.x, |
| ; EG-NEXT: BFE_INT T51.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T8.W, T40.Y, literal.y, |
| ; EG-NEXT: LSHR * T9.W, T40.Y, literal.z, |
| ; EG-NEXT: 16(2.242078e-44), 23(3.222986e-44) |
| ; EG-NEXT: 26(3.643376e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T51.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T50.Y, PV.X, |
| ; EG-NEXT: BFE_INT T52.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T8.W, T40.Y, literal.x, |
| ; EG-NEXT: LSHR * T9.W, T40.Y, literal.y, |
| ; EG-NEXT: 19(2.662467e-44), 22(3.082857e-44) |
| ; EG-NEXT: BFE_INT T52.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T51.Y, PV.X, |
| ; EG-NEXT: BFE_INT T53.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T8.W, T40.Y, literal.x, |
| ; EG-NEXT: LSHR * T9.W, T40.Y, literal.y, |
| ; EG-NEXT: 15(2.101948e-44), 18(2.522337e-44) |
| ; EG-NEXT: BFE_INT T53.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T52.Y, PV.X, |
| ; EG-NEXT: BFE_INT T54.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T8.W, T40.Y, literal.x, |
| ; EG-NEXT: LSHR * T9.W, T40.Y, literal.y, |
| ; EG-NEXT: 11(1.541428e-44), 14(1.961818e-44) |
| ; EG-NEXT: BFE_INT T54.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T53.Y, PV.X, |
| ; EG-NEXT: BFE_INT T55.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T8.W, T40.Y, literal.x, |
| ; EG-NEXT: LSHR * T9.W, T40.Y, literal.y, |
| ; EG-NEXT: 7(9.809089e-45), 10(1.401298e-44) |
| ; EG-NEXT: BFE_INT T55.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T54.Y, PV.X, |
| ; EG-NEXT: BFE_INT T56.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T8.W, T40.Y, literal.x, |
| ; EG-NEXT: LSHR * T9.W, T40.Y, literal.y, |
| ; EG-NEXT: 3(4.203895e-45), 6(8.407791e-45) |
| ; EG-NEXT: BFE_INT T56.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T55.Y, PV.X, |
| ; EG-NEXT: BFE_INT T57.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T8.W, T40.X, literal.x, |
| ; EG-NEXT: LSHR * T9.W, T40.Y, literal.y, |
| ; EG-NEXT: 17(2.382207e-44), 2(2.802597e-45) |
| ; EG-NEXT: BFE_INT T57.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T56.Y, PV.X, |
| ; EG-NEXT: ASHR T58.Z, T40.X, literal.x, |
| ; EG-NEXT: LSHR T9.W, T40.X, literal.y, |
| ; EG-NEXT: LSHR * T10.W, T40.X, literal.z, |
| ; EG-NEXT: 31(4.344025e-44), 27(3.783506e-44) |
| ; EG-NEXT: 30(4.203895e-44), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T58.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T57.Y, PV.X, |
| ; EG-NEXT: BFE_INT T59.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T9.W, T40.X, literal.x, |
| ; EG-NEXT: LSHR * T10.W, T40.X, literal.y, |
| ; EG-NEXT: 23(3.222986e-44), 26(3.643376e-44) |
| ; EG-NEXT: BFE_INT T59.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T58.Y, PV.X, |
| ; EG-NEXT: BFE_INT T60.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T9.W, T40.X, literal.x, |
| ; EG-NEXT: LSHR * T10.W, T40.X, literal.y, |
| ; EG-NEXT: 19(2.662467e-44), 22(3.082857e-44) |
| ; EG-NEXT: BFE_INT T60.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T59.Y, PV.X, |
| ; EG-NEXT: BFE_INT T61.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T9.W, T40.X, literal.x, |
| ; EG-NEXT: LSHR * T10.W, T40.X, literal.y, |
| ; EG-NEXT: 15(2.101948e-44), 18(2.522337e-44) |
| ; EG-NEXT: BFE_INT T61.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T60.Y, PV.X, |
| ; EG-NEXT: BFE_INT T62.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T9.W, T40.X, literal.x, |
| ; EG-NEXT: LSHR * T10.W, T40.X, literal.y, |
| ; EG-NEXT: 11(1.541428e-44), 14(1.961818e-44) |
| ; EG-NEXT: BFE_INT T62.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T61.Y, PV.X, |
| ; EG-NEXT: BFE_INT T63.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T9.W, T40.X, literal.x, |
| ; EG-NEXT: LSHR * T10.W, T40.X, literal.y, |
| ; EG-NEXT: 7(9.809089e-45), 10(1.401298e-44) |
| ; EG-NEXT: BFE_INT T63.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T62.Y, PV.X, |
| ; EG-NEXT: BFE_INT T64.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR * T9.W, T40.X, literal.x, |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) |
| ; EG-NEXT: ALU clause starting at 253: |
| ; EG-NEXT: LSHR * T10.W, T40.X, literal.x, |
| ; EG-NEXT: 6(8.407791e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T64.X, PV.W, 0.0, 1, |
| ; EG-NEXT: MOV T63.Y, T63.X, |
| ; EG-NEXT: BFE_INT T65.Z, T9.W, 0.0, 1, |
| ; EG-NEXT: LSHR T9.W, T40.X, 1, BS:VEC_120/SCL_212 |
| ; EG-NEXT: LSHR * T10.W, T40.X, literal.x, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T65.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T64.Y, PV.X, |
| ; EG-NEXT: BFE_INT T49.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T9.W, T40.X, literal.x, |
| ; EG-NEXT: LSHR * T10.W, T40.X, literal.y, |
| ; EG-NEXT: 12(1.681558e-44), 5(7.006492e-45) |
| ; EG-NEXT: BFE_INT T66.X, T40.Y, 0.0, 1, |
| ; EG-NEXT: MOV T65.Y, PV.X, |
| ; EG-NEXT: BFE_INT T67.Z, PS, 0.0, 1, |
| ; EG-NEXT: LSHR T10.W, T40.X, literal.x, |
| ; EG-NEXT: LSHR * T11.W, T40.X, literal.y, |
| ; EG-NEXT: 9(1.261169e-44), 4(5.605194e-45) |
| ; EG-NEXT: BFE_INT T67.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T49.Y, T49.X, |
| ; EG-NEXT: BFE_INT T40.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: LSHR T10.W, T40.X, literal.x, BS:VEC_120/SCL_212 |
| ; EG-NEXT: LSHR * T11.W, T40.X, literal.y, |
| ; EG-NEXT: 13(1.821688e-44), 8(1.121039e-44) |
| ; EG-NEXT: BFE_INT T40.X, PS, 0.0, 1, |
| ; EG-NEXT: MOV T67.Y, PV.X, |
| ; EG-NEXT: BFE_INT T68.Z, PV.W, 0.0, 1, |
| ; EG-NEXT: MOV T49.W, T49.Z, |
| ; EG-NEXT: MOV * T65.W, T65.Z, |
| ; EG-NEXT: BFE_INT T68.X, T9.W, 0.0, 1, |
| ; EG-NEXT: MOV T40.Y, PV.X, |
| ; EG-NEXT: BFE_INT T69.Z, T8.W, 0.0, 1, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV T67.W, T67.Z, |
| ; EG-NEXT: MOV * T64.W, T64.Z, |
| ; EG-NEXT: BFE_INT T69.X, T7.Y, 0.0, 1, |
| ; EG-NEXT: MOV T68.Y, PV.X, |
| ; EG-NEXT: BFE_INT T70.Z, T6.Y, 0.0, 1, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV T40.W, T40.Z, |
| ; EG-NEXT: MOV * T63.W, T63.Z, |
| ; EG-NEXT: BFE_INT T70.X, T7.W, 0.0, 1, |
| ; EG-NEXT: MOV T69.Y, PV.X, |
| ; EG-NEXT: BFE_INT T71.Z, T6.W, 0.0, 1, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV T68.W, T68.Z, |
| ; EG-NEXT: MOV * T62.W, T62.Z, |
| ; EG-NEXT: BFE_INT T71.X, T6.Z, 0.0, 1, |
| ; EG-NEXT: MOV T70.Y, PV.X, |
| ; EG-NEXT: BFE_INT T72.Z, T5.Y, 0.0, 1, |
| ; EG-NEXT: MOV T69.W, T69.Z, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV * T61.W, T61.Z, |
| ; EG-NEXT: BFE_INT T72.X, T5.W, 0.0, 1, |
| ; EG-NEXT: MOV T71.Y, PV.X, |
| ; EG-NEXT: BFE_INT T66.Z, T5.Z, 0.0, 1, |
| ; EG-NEXT: MOV T70.W, T70.Z, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV * T60.W, T60.Z, |
| ; EG-NEXT: LSHR T73.X, T4.Y, literal.x, |
| ; EG-NEXT: MOV T72.Y, PV.X, |
| ; EG-NEXT: BFE_INT T74.Z, T4.W, 0.0, 1, |
| ; EG-NEXT: MOV T71.W, T71.Z, |
| ; EG-NEXT: MOV * T59.W, T59.Z, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; EG-NEXT: BFE_INT T74.X, T4.Z, 0.0, 1, |
| ; EG-NEXT: MOV T66.Y, T66.X, |
| ; EG-NEXT: BFE_INT T75.Z, T3.Y, 0.0, 1, |
| ; EG-NEXT: MOV T72.W, T72.Z, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV * T58.W, T58.Z, |
| ; EG-NEXT: BFE_INT T75.X, T3.W, 0.0, 1, |
| ; EG-NEXT: MOV T74.Y, PV.X, |
| ; EG-NEXT: BFE_INT T76.Z, T3.Z, 0.0, 1, |
| ; EG-NEXT: MOV T66.W, T66.Z, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV * T57.W, T57.Z, |
| ; EG-NEXT: BFE_INT T76.X, T2.Y, 0.0, 1, |
| ; EG-NEXT: MOV T75.Y, PV.X, |
| ; EG-NEXT: BFE_INT T77.Z, T2.W, 0.0, 1, |
| ; EG-NEXT: MOV T74.W, T74.Z, |
| ; EG-NEXT: MOV * T56.W, T56.Z, |
| ; EG-NEXT: BFE_INT T77.X, T2.Z, 0.0, 1, |
| ; EG-NEXT: MOV T76.Y, PV.X, |
| ; EG-NEXT: BFE_INT T78.Z, T1.Y, 0.0, 1, |
| ; EG-NEXT: MOV T75.W, T75.Z, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV * T55.W, T55.Z, |
| ; EG-NEXT: BFE_INT T78.X, T1.W, 0.0, 1, |
| ; EG-NEXT: MOV T77.Y, PV.X, |
| ; EG-NEXT: BFE_INT T79.Z, T1.Z, 0.0, 1, |
| ; EG-NEXT: MOV T76.W, T76.Z, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV * T54.W, T54.Z, |
| ; EG-NEXT: BFE_INT T79.X, T0.Y, 0.0, 1, |
| ; EG-NEXT: MOV T78.Y, PV.X, |
| ; EG-NEXT: BFE_INT T80.Z, T0.W, 0.0, 1, |
| ; EG-NEXT: MOV T77.W, T77.Z, |
| ; EG-NEXT: MOV * T53.W, T53.Z, |
| ; EG-NEXT: BFE_INT T80.X, T0.Z, 0.0, 1, |
| ; EG-NEXT: MOV T79.Y, PV.X, |
| ; EG-NEXT: ADD_INT T0.Z, KC0[2].Y, literal.x, |
| ; EG-NEXT: MOV T78.W, T78.Z, BS:VEC_120/SCL_212 |
| ; EG-NEXT: MOV * T52.W, T52.Z, |
| ; EG-NEXT: 480(6.726233e-43), 0(0.000000e+00) |
| ; EG-NEXT: LSHR T81.X, PV.Z, literal.x, |
| ; EG-NEXT: MOV T80.Y, PV.X, |
| ; EG-NEXT: ADD_INT T0.Z, KC0[2].Y, literal.y, |
| ; EG-NEXT: MOV T79.W, T79.Z, |
| ; EG-NEXT: MOV * T51.W, T51.Z, |
| ; EG-NEXT: 2(2.802597e-45), 496(6.950440e-43) |
| ; EG-NEXT: LSHR T82.X, PV.Z, literal.x, |
| ; EG-NEXT: MOV T80.W, T80.Z, |
| ; EG-NEXT: MOV * T50.W, T50.Z, |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) |
| ; |
| ; GFX12-LABEL: constant_sextload_v64i1_to_v64i64: |
| ; GFX12: ; %bb.0: |
| ; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_load_b64 s[10:11], s[2:3], 0x0 |
| ; GFX12-NEXT: s_wait_kmcnt 0x0 |
| ; GFX12-NEXT: s_lshr_b32 s96, s11, 30 |
| ; GFX12-NEXT: s_lshr_b32 s98, s11, 31 |
| ; GFX12-NEXT: s_lshr_b32 s92, s11, 28 |
| ; GFX12-NEXT: s_lshr_b32 s94, s11, 29 |
| ; GFX12-NEXT: s_lshr_b32 s78, s11, 26 |
| ; GFX12-NEXT: s_lshr_b32 s88, s11, 27 |
| ; GFX12-NEXT: s_bfe_i64 s[96:97], s[96:97], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[100:101], s[98:99], 0x10000 |
| ; GFX12-NEXT: s_lshr_b32 s66, s11, 24 |
| ; GFX12-NEXT: s_lshr_b32 s74, s11, 25 |
| ; GFX12-NEXT: s_bfe_i64 s[92:93], s[92:93], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[94:95], s[94:95], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s96 |
| ; GFX12-NEXT: s_lshr_b32 s56, s11, 22 |
| ; GFX12-NEXT: s_lshr_b32 s62, s11, 23 |
| ; GFX12-NEXT: v_dual_mov_b32 v2, s97 :: v_dual_mov_b32 v3, s100 |
| ; GFX12-NEXT: v_dual_mov_b32 v4, s101 :: v_dual_mov_b32 v5, s92 |
| ; GFX12-NEXT: s_bfe_i64 s[78:79], s[78:79], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[88:89], s[88:89], 0x10000 |
| ; GFX12-NEXT: s_lshr_b32 s44, s11, 20 |
| ; GFX12-NEXT: s_lshr_b32 s52, s11, 21 |
| ; GFX12-NEXT: s_lshr_b32 s30, s11, 18 |
| ; GFX12-NEXT: s_lshr_b32 s40, s11, 19 |
| ; GFX12-NEXT: s_lshr_b32 s18, s11, 16 |
| ; GFX12-NEXT: s_lshr_b32 s26, s11, 17 |
| ; GFX12-NEXT: s_lshr_b32 s2, s11, 14 |
| ; GFX12-NEXT: s_lshr_b32 s4, s11, 15 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s93 :: v_dual_mov_b32 v7, s94 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s95 :: v_dual_mov_b32 v9, s78 |
| ; GFX12-NEXT: s_bfe_i64 s[66:67], s[66:67], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[74:75], s[74:75], 0x10000 |
| ; GFX12-NEXT: s_lshr_b32 s6, s11, 12 |
| ; GFX12-NEXT: s_lshr_b32 s8, s11, 13 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s79 :: v_dual_mov_b32 v11, s88 |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s89 :: v_dual_mov_b32 v13, s66 |
| ; GFX12-NEXT: s_bfe_i64 s[56:57], s[56:57], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[62:63], s[62:63], 0x10000 |
| ; GFX12-NEXT: s_lshr_b32 s12, s11, 10 |
| ; GFX12-NEXT: s_lshr_b32 s14, s11, 11 |
| ; GFX12-NEXT: v_dual_mov_b32 v14, s67 :: v_dual_mov_b32 v15, s74 |
| ; GFX12-NEXT: v_dual_mov_b32 v16, s75 :: v_dual_mov_b32 v17, s56 |
| ; GFX12-NEXT: s_bfe_i64 s[44:45], s[44:45], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[52:53], s[52:53], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[40:41], s[40:41], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x10000 |
| ; GFX12-NEXT: s_lshr_b32 s16, s11, 8 |
| ; GFX12-NEXT: s_lshr_b32 s20, s11, 9 |
| ; GFX12-NEXT: v_dual_mov_b32 v18, s57 :: v_dual_mov_b32 v19, s62 |
| ; GFX12-NEXT: v_dual_mov_b32 v20, s63 :: v_dual_mov_b32 v21, s44 |
| ; GFX12-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x10000 |
| ; GFX12-NEXT: s_lshr_b32 s22, s11, 6 |
| ; GFX12-NEXT: s_lshr_b32 s24, s11, 7 |
| ; GFX12-NEXT: v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v23, s52 |
| ; GFX12-NEXT: v_dual_mov_b32 v24, s53 :: v_dual_mov_b32 v25, s30 |
| ; GFX12-NEXT: v_dual_mov_b32 v26, s31 :: v_dual_mov_b32 v27, s40 |
| ; GFX12-NEXT: v_dual_mov_b32 v28, s41 :: v_dual_mov_b32 v29, s18 |
| ; GFX12-NEXT: v_dual_mov_b32 v30, s19 :: v_dual_mov_b32 v31, s26 |
| ; GFX12-NEXT: v_mov_b32_e32 v32, s27 |
| ; GFX12-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x10000 |
| ; GFX12-NEXT: s_clause 0x7 |
| ; GFX12-NEXT: global_store_b128 v0, v[1:4], s[0:1] offset:496 |
| ; GFX12-NEXT: global_store_b128 v0, v[5:8], s[0:1] offset:480 |
| ; GFX12-NEXT: global_store_b128 v0, v[9:12], s[0:1] offset:464 |
| ; GFX12-NEXT: global_store_b128 v0, v[13:16], s[0:1] offset:448 |
| ; GFX12-NEXT: global_store_b128 v0, v[17:20], s[0:1] offset:432 |
| ; GFX12-NEXT: global_store_b128 v0, v[21:24], s[0:1] offset:416 |
| ; GFX12-NEXT: global_store_b128 v0, v[25:28], s[0:1] offset:400 |
| ; GFX12-NEXT: global_store_b128 v0, v[29:32], s[0:1] offset:384 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s2 :: v_dual_mov_b32 v2, s3 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s4 :: v_dual_mov_b32 v4, s5 |
| ; GFX12-NEXT: v_mov_b32_e32 v5, s6 |
| ; GFX12-NEXT: s_lshr_b32 s28, s11, 4 |
| ; GFX12-NEXT: s_lshr_b32 s34, s11, 5 |
| ; GFX12-NEXT: s_lshr_b32 s36, s11, 2 |
| ; GFX12-NEXT: s_lshr_b32 s38, s11, 3 |
| ; GFX12-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s12 |
| ; GFX12-NEXT: s_lshr_b32 s42, s11, 1 |
| ; GFX12-NEXT: s_mov_b32 s46, s11 |
| ; GFX12-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s13 :: v_dual_mov_b32 v11, s14 |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s15 :: v_dual_mov_b32 v13, s16 |
| ; GFX12-NEXT: s_lshr_b32 s48, s10, 30 |
| ; GFX12-NEXT: s_lshr_b32 s50, s10, 31 |
| ; GFX12-NEXT: s_bfe_i64 s[38:39], s[38:39], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[36:37], s[36:37], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[34:35], s[34:35], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v14, s17 :: v_dual_mov_b32 v15, s20 |
| ; GFX12-NEXT: v_dual_mov_b32 v16, s21 :: v_dual_mov_b32 v17, s22 |
| ; GFX12-NEXT: s_lshr_b32 s54, s10, 28 |
| ; GFX12-NEXT: s_lshr_b32 s58, s10, 29 |
| ; GFX12-NEXT: s_bfe_i64 s[46:47], s[46:47], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[42:43], s[42:43], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v18, s23 :: v_dual_mov_b32 v19, s24 |
| ; GFX12-NEXT: v_dual_mov_b32 v20, s25 :: v_dual_mov_b32 v21, s28 |
| ; GFX12-NEXT: s_lshr_b32 s60, s10, 26 |
| ; GFX12-NEXT: s_lshr_b32 s64, s10, 27 |
| ; GFX12-NEXT: s_bfe_i64 s[50:51], s[50:51], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[48:49], s[48:49], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v22, s29 :: v_dual_mov_b32 v23, s34 |
| ; GFX12-NEXT: v_mov_b32_e32 v24, s35 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v0, v[1:4], s[0:1] offset:368 |
| ; GFX12-NEXT: global_store_b128 v0, v[5:8], s[0:1] offset:352 |
| ; GFX12-NEXT: global_store_b128 v0, v[9:12], s[0:1] offset:336 |
| ; GFX12-NEXT: global_store_b128 v0, v[13:16], s[0:1] offset:320 |
| ; GFX12-NEXT: global_store_b128 v0, v[17:20], s[0:1] offset:304 |
| ; GFX12-NEXT: global_store_b128 v0, v[21:24], s[0:1] offset:288 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s36 :: v_dual_mov_b32 v2, s37 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s38 :: v_dual_mov_b32 v4, s39 |
| ; GFX12-NEXT: v_mov_b32_e32 v5, s46 |
| ; GFX12-NEXT: s_lshr_b32 s68, s10, 24 |
| ; GFX12-NEXT: s_lshr_b32 s70, s10, 25 |
| ; GFX12-NEXT: s_lshr_b32 s72, s10, 22 |
| ; GFX12-NEXT: s_lshr_b32 s76, s10, 23 |
| ; GFX12-NEXT: s_bfe_i64 s[58:59], s[58:59], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[54:55], s[54:55], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s47 :: v_dual_mov_b32 v7, s42 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s43 :: v_dual_mov_b32 v9, s48 |
| ; GFX12-NEXT: s_lshr_b32 s80, s10, 20 |
| ; GFX12-NEXT: s_lshr_b32 s82, s10, 21 |
| ; GFX12-NEXT: s_bfe_i64 s[64:65], s[64:65], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[60:61], s[60:61], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s49 :: v_dual_mov_b32 v11, s50 |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s51 :: v_dual_mov_b32 v13, s54 |
| ; GFX12-NEXT: s_lshr_b32 s84, s10, 18 |
| ; GFX12-NEXT: s_lshr_b32 s86, s10, 19 |
| ; GFX12-NEXT: s_bfe_i64 s[76:77], s[76:77], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[72:73], s[72:73], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[70:71], s[70:71], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[68:69], s[68:69], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v14, s55 :: v_dual_mov_b32 v15, s58 |
| ; GFX12-NEXT: v_dual_mov_b32 v16, s59 :: v_dual_mov_b32 v17, s60 |
| ; GFX12-NEXT: s_lshr_b32 s90, s10, 16 |
| ; GFX12-NEXT: s_lshr_b32 s98, s10, 17 |
| ; GFX12-NEXT: s_bfe_i64 s[82:83], s[82:83], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[80:81], s[80:81], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v18, s61 :: v_dual_mov_b32 v19, s64 |
| ; GFX12-NEXT: v_dual_mov_b32 v20, s65 :: v_dual_mov_b32 v21, s68 |
| ; GFX12-NEXT: s_lshr_b32 s96, s10, 14 |
| ; GFX12-NEXT: s_lshr_b32 s100, s10, 15 |
| ; GFX12-NEXT: s_lshr_b32 s94, s10, 13 |
| ; GFX12-NEXT: s_lshr_b32 s88, s10, 11 |
| ; GFX12-NEXT: s_lshr_b32 s74, s10, 9 |
| ; GFX12-NEXT: s_lshr_b32 s62, s10, 7 |
| ; GFX12-NEXT: s_lshr_b32 s52, s10, 5 |
| ; GFX12-NEXT: s_lshr_b32 s40, s10, 3 |
| ; GFX12-NEXT: s_lshr_b32 s26, s10, 1 |
| ; GFX12-NEXT: s_bfe_i64 s[86:87], s[86:87], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[84:85], s[84:85], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v22, s69 :: v_dual_mov_b32 v23, s70 |
| ; GFX12-NEXT: v_mov_b32_e32 v24, s71 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v0, v[1:4], s[0:1] offset:272 |
| ; GFX12-NEXT: global_store_b128 v0, v[5:8], s[0:1] offset:256 |
| ; GFX12-NEXT: global_store_b128 v0, v[9:12], s[0:1] offset:240 |
| ; GFX12-NEXT: global_store_b128 v0, v[13:16], s[0:1] offset:224 |
| ; GFX12-NEXT: global_store_b128 v0, v[17:20], s[0:1] offset:208 |
| ; GFX12-NEXT: global_store_b128 v0, v[21:24], s[0:1] offset:192 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s72 :: v_dual_mov_b32 v2, s73 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s76 :: v_dual_mov_b32 v4, s77 |
| ; GFX12-NEXT: v_mov_b32_e32 v5, s80 |
| ; GFX12-NEXT: s_lshr_b32 s92, s10, 12 |
| ; GFX12-NEXT: s_lshr_b32 s78, s10, 10 |
| ; GFX12-NEXT: s_bfe_i64 s[98:99], s[98:99], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[90:91], s[90:91], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s81 :: v_dual_mov_b32 v7, s82 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s83 :: v_dual_mov_b32 v9, s84 |
| ; GFX12-NEXT: s_lshr_b32 s66, s10, 8 |
| ; GFX12-NEXT: s_lshr_b32 s56, s10, 6 |
| ; GFX12-NEXT: s_lshr_b32 s44, s10, 4 |
| ; GFX12-NEXT: s_lshr_b32 s30, s10, 2 |
| ; GFX12-NEXT: s_bfe_i64 s[18:19], s[10:11], 0x10000 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: s_bfe_i64 s[10:11], s[26:27], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[26:27], s[40:41], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[40:41], s[52:53], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[52:53], s[62:63], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[62:63], s[74:75], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[74:75], s[88:89], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[88:89], s[94:95], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[94:95], s[100:101], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[96:97], s[96:97], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s85 :: v_dual_mov_b32 v11, s86 |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s87 :: v_dual_mov_b32 v13, s90 |
| ; GFX12-NEXT: s_bfe_i64 s[78:79], s[78:79], 0x10000 |
| ; GFX12-NEXT: s_bfe_i64 s[92:93], s[92:93], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v14, s91 :: v_dual_mov_b32 v15, s98 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_dual_mov_b32 v16, s99 :: v_dual_mov_b32 v17, s96 |
| ; GFX12-NEXT: s_bfe_i64 s[66:67], s[66:67], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v18, s97 :: v_dual_mov_b32 v19, s94 |
| ; GFX12-NEXT: v_dual_mov_b32 v20, s95 :: v_dual_mov_b32 v21, s92 |
| ; GFX12-NEXT: s_bfe_i64 s[56:57], s[56:57], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v22, s93 :: v_dual_mov_b32 v23, s88 |
| ; GFX12-NEXT: v_mov_b32_e32 v24, s89 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v0, v[1:4], s[0:1] offset:176 |
| ; GFX12-NEXT: global_store_b128 v0, v[5:8], s[0:1] offset:160 |
| ; GFX12-NEXT: global_store_b128 v0, v[9:12], s[0:1] offset:144 |
| ; GFX12-NEXT: global_store_b128 v0, v[13:16], s[0:1] offset:128 |
| ; GFX12-NEXT: global_store_b128 v0, v[17:20], s[0:1] offset:112 |
| ; GFX12-NEXT: global_store_b128 v0, v[21:24], s[0:1] offset:96 |
| ; GFX12-NEXT: v_dual_mov_b32 v1, s78 :: v_dual_mov_b32 v2, s79 |
| ; GFX12-NEXT: v_dual_mov_b32 v3, s74 :: v_dual_mov_b32 v4, s75 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_mov_b32_e32 v5, s66 |
| ; GFX12-NEXT: s_bfe_i64 s[44:45], s[44:45], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v6, s67 :: v_dual_mov_b32 v7, s62 |
| ; GFX12-NEXT: v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v9, s56 |
| ; GFX12-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x10000 |
| ; GFX12-NEXT: v_dual_mov_b32 v10, s57 :: v_dual_mov_b32 v11, s52 |
| ; GFX12-NEXT: s_wait_alu 0xfffe |
| ; GFX12-NEXT: v_dual_mov_b32 v12, s53 :: v_dual_mov_b32 v13, s44 |
| ; GFX12-NEXT: v_dual_mov_b32 v14, s45 :: v_dual_mov_b32 v15, s40 |
| ; GFX12-NEXT: v_dual_mov_b32 v16, s41 :: v_dual_mov_b32 v17, s30 |
| ; GFX12-NEXT: v_dual_mov_b32 v18, s31 :: v_dual_mov_b32 v19, s26 |
| ; GFX12-NEXT: v_dual_mov_b32 v20, s27 :: v_dual_mov_b32 v21, s18 |
| ; GFX12-NEXT: v_dual_mov_b32 v22, s19 :: v_dual_mov_b32 v23, s10 |
| ; GFX12-NEXT: v_mov_b32_e32 v24, s11 |
| ; GFX12-NEXT: s_clause 0x5 |
| ; GFX12-NEXT: global_store_b128 v0, v[1:4], s[0:1] offset:80 |
| ; GFX12-NEXT: global_store_b128 v0, v[5:8], s[0:1] offset:64 |
| ; GFX12-NEXT: global_store_b128 v0, v[9:12], s[0:1] offset:48 |
| ; GFX12-NEXT: global_store_b128 v0, v[13:16], s[0:1] offset:32 |
| ; GFX12-NEXT: global_store_b128 v0, v[17:20], s[0:1] offset:16 |
| ; GFX12-NEXT: global_store_b128 v0, v[21:24], s[0:1] |
| ; GFX12-NEXT: s_endpgm |
| %load = load <64 x i1>, ptr addrspace(4) %in |
| %ext = sext <64 x i1> %load to <64 x i64> |
| store <64 x i64> %ext, ptr addrspace(1) %out |
| ret void |
| } |
| |
| attributes #0 = { nounwind } |
