| ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 | |
| ; RUN: llc -mtriple=amdgpu6.00 < %s | FileCheck -allow-deprecated-dag-overlap --check-prefix=GCN-NOHSA-SI %s | |
| ; RUN: llc -mtriple=amdgpu7.00--amdhsa < %s | FileCheck -allow-deprecated-dag-overlap --check-prefix=GCN-HSA %s | |
| ; RUN: llc -mtriple=amdgpu8.02 -mattr=-flat-for-global < %s | FileCheck -allow-deprecated-dag-overlap --check-prefix=GCN-NOHSA-VI %s | |
| ; RUN: llc -mtriple=r600 -mcpu=redwood < %s | FileCheck -allow-deprecated-dag-overlap --check-prefix=EG %s | |
| ; RUN: llc -mtriple=r600 -mcpu=cayman < %s | FileCheck -allow-deprecated-dag-overlap --check-prefix=CM %s | |
| ; TODO: NOT AND | |
| define amdgpu_kernel void @global_load_i8(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_load_i8: | |
| ; GCN-NOHSA-SI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_byte v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_load_i8: | |
| ; GCN-HSA: ; %bb.0: ; %entry | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ubyte v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_byte v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_load_i8: | |
| ; GCN-NOHSA-VI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_byte v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_load_i8: | |
| ; EG: ; %bb.0: ; %entry | |
| ; 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) | |
| ; | |
| ; CM-LABEL: global_load_i8: | |
| ; CM: ; %bb.0: ; %entry | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 11, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT MSKOR T0.XW, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: AND_INT * T0.W, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 255(3.573311e-43), 3(4.203895e-45) | |
| ; CM-NEXT: LSHL T0.X, PV.Z, PV.W, | |
| ; CM-NEXT: LSHL * T0.W, literal.x, PV.W, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: MOV T0.Y, 0.0, | |
| ; CM-NEXT: MOV * T0.Z, 0.0, | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| entry: | |
| %ld = load i8, ptr addrspace(1) %in | |
| store i8 %ld, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_load_v2i8(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_load_v2i8: | |
| ; GCN-NOHSA-SI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_short v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_load_v2i8: | |
| ; GCN-HSA: ; %bb.0: ; %entry | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ushort v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_short v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_load_v2i8: | |
| ; GCN-NOHSA-VI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_short v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_load_v2i8: | |
| ; EG: ; %bb.0: ; %entry | |
| ; 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) | |
| ; | |
| ; CM-LABEL: global_load_v2i8: | |
| ; CM: ; %bb.0: ; %entry | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 11, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT MSKOR T0.XW, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: AND_INT * T0.W, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 3(4.203895e-45) | |
| ; CM-NEXT: LSHL T0.X, PV.Z, PV.W, | |
| ; CM-NEXT: LSHL * T0.W, literal.x, PV.W, | |
| ; CM-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; CM-NEXT: MOV T0.Y, 0.0, | |
| ; CM-NEXT: MOV * T0.Z, 0.0, | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| entry: | |
| %ld = load <2 x i8>, ptr addrspace(1) %in | |
| store <2 x i8> %ld, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_load_v3i8(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_load_v3i8: | |
| ; GCN-NOHSA-SI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[2:3] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v0, v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v1, 16, v0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_short v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_byte v1, off, s[4:7], 0 offset:2 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_load_v3i8: | |
| ; GCN-HSA: ; %bb.0: ; %entry | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dword v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 2 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v3, 16, v2 | |
| ; GCN-HSA-NEXT: flat_store_short v[0:1], v2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: flat_store_byte v[0:1], v3 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_load_v3i8: | |
| ; GCN-NOHSA-VI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dword v0, v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v1, 16, v0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_short v0, off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_byte v1, off, s[0:3], 0 offset:2 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_load_v3i8: | |
| ; EG: ; %bb.0: ; %entry | |
| ; EG-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 27, @11, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT MSKOR T6.XW, T8.X | |
| ; EG-NEXT: MEM_RAT MSKOR T7.XW, T0.X | |
| ; EG-NEXT: CF_END | |
| ; 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: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 11: | |
| ; EG-NEXT: MOV * T2.X, T0.X, | |
| ; EG-NEXT: MOV T0.Y, PV.X, | |
| ; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, PV.W, literal.x, | |
| ; EG-NEXT: MOV * T2.W, literal.y, | |
| ; EG-NEXT: 3(4.203895e-45), 8(1.121039e-44) | |
| ; EG-NEXT: BFE_UINT T2.W, T0.Y, literal.x, PS, | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 3(4.203895e-45) | |
| ; EG-NEXT: LSHL T6.X, PV.W, PS, | |
| ; EG-NEXT: LSHL * T6.W, literal.x, PS, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: MOV T6.Y, 0.0, | |
| ; EG-NEXT: AND_INT T1.W, KC0[2].Y, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, T0.X, literal.y, | |
| ; EG-NEXT: 3(4.203895e-45), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; EG-NEXT: LSHL T7.X, T2.W, PV.W, | |
| ; EG-NEXT: LSHL * T7.W, literal.x, PV.W, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: MOV T7.Y, 0.0, | |
| ; EG-NEXT: MOV T6.Z, 0.0, | |
| ; EG-NEXT: MOV * T7.Z, 0.0, | |
| ; EG-NEXT: LSHR T0.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: LSHR * T8.X, T0.W, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_load_v3i8: | |
| ; CM: ; %bb.0: ; %entry | |
| ; CM-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 28, @11, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT MSKOR T6.XW, T8.X | |
| ; CM-NEXT: MEM_RAT MSKOR T7.XW, T0.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 11: | |
| ; CM-NEXT: MOV * T2.X, T0.X, | |
| ; CM-NEXT: MOV T0.Y, PV.X, | |
| ; CM-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.W, literal.x, | |
| ; CM-NEXT: MOV * T1.W, literal.y, | |
| ; CM-NEXT: 3(4.203895e-45), 8(1.121039e-44) | |
| ; CM-NEXT: BFE_UINT T1.Z, T0.Y, literal.x, PV.W, | |
| ; CM-NEXT: LSHL * T1.W, PV.Z, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 3(4.203895e-45) | |
| ; CM-NEXT: LSHL T6.X, PV.Z, PV.W, | |
| ; CM-NEXT: LSHL * T6.W, literal.x, PV.W, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: MOV T6.Y, 0.0, | |
| ; CM-NEXT: AND_INT * T1.W, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 3(4.203895e-45) | |
| ; CM-NEXT: LSHL T7.X, PV.Z, PV.W, | |
| ; CM-NEXT: LSHL * T7.W, literal.x, PV.W, | |
| ; CM-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; CM-NEXT: MOV T7.Y, 0.0, | |
| ; CM-NEXT: MOV * T6.Z, 0.0, | |
| ; CM-NEXT: MOV * T7.Z, 0.0, | |
| ; CM-NEXT: LSHR * T0.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR * T8.X, T0.W, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| entry: | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <3 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %ld = load <3 x i8>, ptr addrspace(1) %in.tid | |
| store <3 x i8> %ld, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_load_v4i8(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_load_v4i8: | |
| ; GCN-NOHSA-SI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[2:3] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v0, v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_load_v4i8: | |
| ; GCN-HSA: ; %bb.0: ; %entry | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dword v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_dword v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_load_v4i8: | |
| ; GCN-NOHSA-VI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dword v0, v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dword v0, off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_load_v4i8: | |
| ; EG: ; %bb.0: ; %entry | |
| ; EG-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 1, @11, 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: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 11: | |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_load_v4i8: | |
| ; CM: ; %bb.0: ; %entry | |
| ; CM-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 1, @11, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0.X, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 11: | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| entry: | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <4 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %ld = load <4 x i8>, ptr addrspace(1) %in.tid | |
| store <4 x i8> %ld, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_load_v8i8(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_load_v8i8: | |
| ; GCN-NOHSA-SI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[2:3] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_load_v8i8: | |
| ; GCN-HSA: ; %bb.0: ; %entry | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx2 v[0:1], v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_dwordx2 v[2:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_load_v8i8: | |
| ; GCN-NOHSA-VI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx2 v[0:1], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_load_v8i8: | |
| ; EG: ; %bb.0: ; %entry | |
| ; EG-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 1, @11, 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: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 11: | |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_load_v8i8: | |
| ; CM: ; %bb.0: ; %entry | |
| ; CM-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 1, @11, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_64 T0.XY, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 11: | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| entry: | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <8 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %ld = load <8 x i8>, ptr addrspace(1) %in.tid | |
| store <8 x i8> %ld, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_load_v16i8(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_load_v16i8: | |
| ; GCN-NOHSA-SI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[2:3] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[0:3], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_load_v16i8: | |
| ; GCN-HSA: ; %bb.0: ; %entry | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_load_v16i8: | |
| ; GCN-NOHSA-VI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_load_v16i8: | |
| ; EG: ; %bb.0: ; %entry | |
| ; EG-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 1, @11, 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_128 T0.XYZW, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 8: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 11: | |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_load_v16i8: | |
| ; CM: ; %bb.0: ; %entry | |
| ; CM-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 1, @11, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 11: | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| entry: | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <16 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %ld = load <16 x i8>, ptr addrspace(1) %in.tid | |
| store <16 x i8> %ld, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_i8_to_i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_i8_to_i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_i8_to_i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ubyte v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_dword v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_i8_to_i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_i8_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) | |
| ; | |
| ; CM-LABEL: global_zextload_i8_to_i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 1, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0.X, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %a = load i8, ptr addrspace(1) %in | |
| %ext = zext i8 %a to i32 | |
| store i32 %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_i8_to_i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_i8_to_i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_i8_to_i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_sbyte v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_dword v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_i8_to_i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_i8_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, literal.x, | |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 2(2.802597e-45) | |
| ; | |
| ; CM-LABEL: global_sextload_i8_to_i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 3, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0.X, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: BFE_INT * T0.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %ld = load i8, ptr addrspace(1) %in | |
| %ext = sext i8 %ld to i32 | |
| store i32 %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v1i8_to_v1i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v1i8_to_v1i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v1i8_to_v1i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ubyte v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_dword v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v1i8_to_v1i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v1i8_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) | |
| ; | |
| ; CM-LABEL: global_zextload_v1i8_to_v1i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 1, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0.X, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %load = load <1 x i8>, ptr addrspace(1) %in | |
| %ext = zext <1 x i8> %load to <1 x i32> | |
| store <1 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v1i8_to_v1i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v1i8_to_v1i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v1i8_to_v1i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_sbyte v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_dword v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v1i8_to_v1i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v1i8_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, literal.x, | |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 2(2.802597e-45) | |
| ; | |
| ; CM-LABEL: global_sextload_v1i8_to_v1i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 3, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0.X, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: BFE_INT * T0.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %load = load <1 x i8>, ptr addrspace(1) %in | |
| %ext = sext <1 x i8> %load to <1 x i32> | |
| store <1 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG/CM should use DST, but for some there are redundant MOVs | |
| define amdgpu_kernel void @global_zextload_v2i8_to_v2i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v2i8_to_v2i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v1, 8, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xff, v0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v2i8_to_v2i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ushort v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v3, 8, v2 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v2, 0xff, v2 | |
| ; GCN-HSA-NEXT: flat_store_dwordx2 v[0:1], v[2:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v2i8_to_v2i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 8 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_sdwa v1, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v0, 0xff, v0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v2i8_to_v2i32: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 1, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 12, @10, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T4.XY, T5.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: PAD | |
| ; EG-NEXT: Fetch clause starting at 6: | |
| ; EG-NEXT: VTX_READ_16 T4.X, T4.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 8: | |
| ; EG-NEXT: MOV * T0.Y, T2.X, | |
| ; EG-NEXT: MOV * T4.X, KC0[2].Z, | |
| ; EG-NEXT: ALU clause starting at 10: | |
| ; EG-NEXT: AND_INT T0.W, T4.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.Y, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T2.X, PV.W, | |
| ; EG-NEXT: MOV T0.Y, PV.X, | |
| ; EG-NEXT: MOV * T1.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT * T4.Y, PV.Y, literal.x, PV.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T4.X, T0.W, literal.x, | |
| ; EG-NEXT: LSHR * T5.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 255(3.573311e-43), 2(2.802597e-45) | |
| ; | |
| ; CM-LABEL: global_zextload_v2i8_to_v2i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 1, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 13, @10, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T4, T5.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_16 T4.X, T4.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.Y, T2.X, | |
| ; CM-NEXT: MOV * T4.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 10: | |
| ; CM-NEXT: AND_INT T0.Z, T4.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; CM-NEXT: OR_INT * T0.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T2.X, PV.W, | |
| ; CM-NEXT: MOV T0.Y, PV.X, | |
| ; CM-NEXT: MOV * T1.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT * T4.Y, PV.Y, literal.x, PV.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT * T4.X, T0.W, literal.x, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR * T5.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %load = load <2 x i8>, ptr addrspace(1) %in | |
| %ext = zext <2 x i8> %load to <2 x i32> | |
| store <2 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG/CM should use DST, but for some there are redundant MOVs | |
| define amdgpu_kernel void @global_sextload_v2i8_to_v2i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v2i8_to_v2i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v1, v0, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v2i8_to_v2i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ushort v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_bfe_i32 v3, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v2, v2, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx2 v[0:1], v[2:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v2i8_to_v2i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v1, v0, 8, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v2i8_to_v2i32: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 1, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 11, @10, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T4.XY, T5.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: PAD | |
| ; EG-NEXT: Fetch clause starting at 6: | |
| ; EG-NEXT: VTX_READ_16 T4.X, T4.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 8: | |
| ; EG-NEXT: MOV * T0.Y, T2.X, | |
| ; EG-NEXT: MOV * T4.X, KC0[2].Z, | |
| ; EG-NEXT: ALU clause starting at 10: | |
| ; EG-NEXT: AND_INT T0.W, T4.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.Y, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T2.X, PV.W, | |
| ; EG-NEXT: MOV * T0.Y, PV.X, | |
| ; EG-NEXT: BFE_INT T4.X, T0.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T0.W, PV.Y, literal.x, | |
| ; EG-NEXT: LSHR * T5.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 2(2.802597e-45) | |
| ; EG-NEXT: BFE_INT * T4.Y, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_sextload_v2i8_to_v2i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 1, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 11, @10, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T4, T5.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_16 T4.X, T4.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.Y, T2.X, | |
| ; CM-NEXT: MOV * T4.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 10: | |
| ; CM-NEXT: AND_INT T0.Z, T4.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; CM-NEXT: OR_INT * T0.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T2.X, PV.W, | |
| ; CM-NEXT: MOV * T0.Y, PV.X, | |
| ; CM-NEXT: BFE_INT T4.X, T0.W, 0.0, literal.x, | |
| ; CM-NEXT: LSHR * T0.W, PV.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR T5.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: BFE_INT * T4.Y, PV.W, 0.0, literal.y, | |
| ; CM-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| %load = load <2 x i8>, ptr addrspace(1) %in | |
| %ext = sext <2 x i8> %load to <2 x i32> | |
| store <2 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG/CM should use DST, but for some there are redundant MOVs | |
| define amdgpu_kernel void @global_zextload_v3i8_to_v3i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v3i8_to_v3i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v2, v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v1, v2, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xff, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v2, v2, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v2, off, s[0:3], 0 offset:8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v3i8_to_v3i32: | |
| ; GCN-HSA: ; %bb.0: ; %entry | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dword v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_bfe_u32 v1, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v0, 0xff, v2 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v2, v2, 16, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx3 v[3:4], v[0:2] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v3i8_to_v3i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dword v1, v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v0, 0xff, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v2, v1, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v1, v1, 8, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v3i8_to_v3i32: | |
| ; EG: ; %bb.0: ; %entry | |
| ; EG-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 10, @11, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T7.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.XY, T6.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; 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: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 11: | |
| ; EG-NEXT: MOV * T0.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT * T5.Y, T0.X, literal.x, PV.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT * T5.X, T0.X, literal.x, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T0.X, T0.X, literal.x, T0.W, | |
| ; EG-NEXT: LSHR * T6.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 2(2.802597e-45) | |
| ; EG-NEXT: ADD_INT * T7.X, PS, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_zextload_v3i8_to_v3i32: | |
| ; CM: ; %bb.0: ; %entry | |
| ; CM-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 10, @11, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T6.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T5.X, T7.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 11: | |
| ; CM-NEXT: MOV * T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT * T5.X, T0.X, literal.x, PV.W, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR * T6.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T7.X, PV.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T0.Y, T0.X, literal.y, T0.W, | |
| ; CM-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; CM-NEXT: AND_INT * T0.X, T0.X, literal.x, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| entry: | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <3 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %ld = load <3 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <3 x i8> %ld to <3 x i32> | |
| store <3 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG/CM should use DST, but for some there are redundant MOVs | |
| define amdgpu_kernel void @global_sextload_v3i8_to_v3i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v3i8_to_v3i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v2, v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v1, v2, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v2, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v2, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v2, off, s[0:3], 0 offset:8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v3i8_to_v3i32: | |
| ; GCN-HSA: ; %bb.0: ; %entry | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dword v0, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_bfe_i32 v2, v0, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v1, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx3 v[3:4], v[0:2] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v3i8_to_v3i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: ; %entry | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dword v0, v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v2, v0, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v1, v0, 8, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v3i8_to_v3i32: | |
| ; EG: ; %bb.0: ; %entry | |
| ; EG-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 11, @11, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.X, T0.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T6.XY, T7.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; 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: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 11: | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT * T5.X, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T6.X, T0.X, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T7.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 2(2.802597e-45) | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T0.X, T7.X, literal.x, | |
| ; EG-NEXT: BFE_INT * T6.Y, PV.W, 0.0, literal.y, | |
| ; EG-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; | |
| ; CM-LABEL: global_sextload_v3i8_to_v3i32: | |
| ; CM: ; %bb.0: ; %entry | |
| ; CM-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 11, @11, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T6, T7.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T5.X, T0.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 11: | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T5.X, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T6.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR T7.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.y, | |
| ; CM-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T0.X, PV.X, literal.x, | |
| ; CM-NEXT: BFE_INT * T6.Y, PV.W, 0.0, literal.y, | |
| ; CM-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| entry: | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <3 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %ld = load <3 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <3 x i8> %ld to <3 x i32> | |
| store <3 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG/CM should use DST, but for some there are redundant MOVs | |
| define amdgpu_kernel void @global_zextload_v4i8_to_v4i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v4i8_to_v4i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v2, v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v3, 24, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v1, v2, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xff, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v2, v2, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v4i8_to_v4i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dword v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v3, 24, v2 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v1, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v0, 0xff, v2 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v2, v2, 16, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v4i8_to_v4i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dword v1, v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v0, 0xff, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v2, v1, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v1, v1, 8, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v4i8_to_v4i32: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 9, @11, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T5.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: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 11: | |
| ; EG-NEXT: MOV * T0.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT * T0.Z, T0.X, literal.x, PV.W, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T0.Y, T0.X, literal.x, T0.W, | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T0.X, T0.X, literal.x, | |
| ; EG-NEXT: LSHR * T5.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 255(3.573311e-43), 2(2.802597e-45) | |
| ; | |
| ; CM-LABEL: global_zextload_v4i8_to_v4i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 10, @11, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T5.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 11: | |
| ; CM-NEXT: MOV * T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT * T0.Z, T0.X, literal.x, PV.W, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T0.Y, T0.X, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: AND_INT * T0.X, T0.X, literal.x, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR * T5.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <4 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <4 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <4 x i8> %load to <4 x i32> | |
| store <4 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG/CM should use DST, but for some there are redundant MOVs | |
| define amdgpu_kernel void @global_sextload_v4i8_to_v4i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v4i8_to_v4i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v0, v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v3, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v0, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v1, v0, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v4i8_to_v4i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dword v0, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v3, 24, v0 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v2, v0, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v1, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v4i8_to_v4i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dword v0, v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v3, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v2, v0, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v1, v0, 8, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v4i8_to_v4i32: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 11, @11, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.XYZW, T0.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: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 11: | |
| ; EG-NEXT: BFE_INT T5.X, T0.X, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: BFE_INT T5.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT T5.Z, PS, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: LSHR T0.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: BFE_INT * T5.Y, PV.W, 0.0, literal.y, | |
| ; EG-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; | |
| ; CM-LABEL: global_sextload_v4i8_to_v4i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 11, @11, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T5, T0.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 11: | |
| ; CM-NEXT: BFE_INT T5.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: LSHR T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: BFE_INT * T5.W, PV.W, 0.0, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: BFE_INT T5.Z, PV.Z, 0.0, literal.x, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR T0.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: BFE_INT * T5.Y, PV.W, 0.0, literal.y, | |
| ; CM-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <4 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <4 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <4 x i8> %load to <4 x i32> | |
| store <4 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG/CM should use DST, but for some there are redundant MOVs | |
| define amdgpu_kernel void @global_zextload_v8i8_to_v8i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v8i8_to_v8i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx2 v[8:9], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v7, 24, v9 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v5, v9, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v4, 0xff, v9 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v6, v9, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v3, 24, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v1, v8, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xff, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v2, v8, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v8i8_to_v8i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx2 v[8:9], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v11, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v10, s0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v7, 24, v9 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v5, v9, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v4, 0xff, v9 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v6, v9, 16, 8 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v3, 24, v8 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v1, v8, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v0, 0xff, v8 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v2, v8, 16, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[12:13], v[4:7] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[10:11], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v8i8_to_v8i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx2 v[6:7], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 24, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v0, 0xff, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v1, 8, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v2, v7, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v7, 24, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v5, 8, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v4, 0xff, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v6, v6, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v8i8_to_v8i32: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 17, @11, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T7.XYZW, T8.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T6.XYZW, T0.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; 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: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 11: | |
| ; EG-NEXT: MOV * T0.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT * T6.Z, T0.X, literal.x, PV.W, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T6.Y, T0.X, literal.x, T0.W, | |
| ; EG-NEXT: BFE_UINT * T7.Z, T0.Y, literal.y, T0.W, | |
| ; EG-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_UINT T7.Y, T0.Y, literal.x, T0.W, | |
| ; EG-NEXT: LSHR * T6.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T6.X, T0.X, literal.x, | |
| ; EG-NEXT: LSHR T7.W, T0.Y, literal.y, | |
| ; EG-NEXT: AND_INT * T7.X, T0.Y, literal.x, | |
| ; EG-NEXT: 255(3.573311e-43), 24(3.363116e-44) | |
| ; EG-NEXT: LSHR * T0.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T8.X, PV.X, literal.x, | |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_zextload_v8i8_to_v8i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 17, @11, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T7.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T6, T8.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_64 T0.XY, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 11: | |
| ; CM-NEXT: MOV * T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT * T6.Z, T0.Y, literal.x, PV.W, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T6.Y, T0.Y, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T6.W, T0.Y, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: AND_INT T6.X, T0.Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T0.Z, T0.X, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 16(2.242078e-44) | |
| ; CM-NEXT: LSHR T7.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T0.Y, T0.X, literal.y, T0.W, | |
| ; CM-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T8.X, PV.X, literal.x, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.y, | |
| ; CM-NEXT: 4(5.605194e-45), 24(3.363116e-44) | |
| ; CM-NEXT: AND_INT * T0.X, T0.X, literal.x, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <8 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <8 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <8 x i8> %load to <8 x i32> | |
| store <8 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG/CM should use DST, but for some there are redundant MOVs | |
| define amdgpu_kernel void @global_sextload_v8i8_to_v8i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v8i8_to_v8i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx2 v[7:8], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v3, 24, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v7, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v1, v7, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v7, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v7, 24, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v6, v8, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v5, v8, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v4, v8, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v8i8_to_v8i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx2 v[7:8], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v11, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v10, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, s0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v3, 24, v7 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v2, v7, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v1, v7, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v0, v7, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 24, v8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v8, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v8, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v8, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[11:12], v[4:7] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[9:10], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v8i8_to_v8i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx2 v[4:5], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v8, 8, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v1, 8, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v3, 24, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v2, v5, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v0, v5, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v7, 24, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v6, v4, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v4, v4, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v5, v8, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v1, v1, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v8i8_to_v8i32: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 21, @11, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T7.XYZW, T8.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T6.XYZW, T0.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; 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: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 11: | |
| ; EG-NEXT: BFE_INT T6.X, T0.X, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: BFE_INT T7.X, T0.Y, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T0.Z, T0.Y, literal.y, | |
| ; EG-NEXT: BFE_INT T6.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: LSHR T1.Y, T0.Y, literal.x, | |
| ; EG-NEXT: BFE_INT T6.Z, PS, 0.0, literal.y, | |
| ; EG-NEXT: BFE_INT T7.W, PV.Z, 0.0, literal.y, | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; EG-NEXT: LSHR T0.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: BFE_INT T6.Y, PS, 0.0, literal.y, | |
| ; EG-NEXT: BFE_INT T7.Z, PV.Y, 0.0, literal.y, | |
| ; EG-NEXT: LSHR * T0.W, T0.Y, literal.y, | |
| ; EG-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T8.X, PV.X, literal.x, | |
| ; EG-NEXT: BFE_INT * T7.Y, PV.W, 0.0, literal.y, | |
| ; EG-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| ; | |
| ; CM-LABEL: global_sextload_v8i8_to_v8i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 21, @11, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T7, T9.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T6, T0.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_64 T0.XY, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 11: | |
| ; CM-NEXT: BFE_INT T6.X, T0.Y, 0.0, literal.x, | |
| ; CM-NEXT: LSHR * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: BFE_INT T7.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T1.Y, T0.X, literal.y, | |
| ; CM-NEXT: LSHR T0.Z, T0.Y, literal.z, | |
| ; CM-NEXT: BFE_INT * T6.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR T8.X, T0.X, literal.x, | |
| ; CM-NEXT: LSHR T0.Y, T0.Y, literal.y, | |
| ; CM-NEXT: BFE_INT T6.Z, PV.Z, 0.0, literal.y, | |
| ; CM-NEXT: BFE_INT * T7.W, PV.Y, 0.0, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: LSHR T9.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: BFE_INT T6.Y, PV.Y, 0.0, literal.y, | |
| ; CM-NEXT: BFE_INT T7.Z, PV.X, 0.0, literal.y, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.y, | |
| ; CM-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T0.X, PV.X, literal.x, | |
| ; CM-NEXT: BFE_INT * T7.Y, PV.W, 0.0, literal.y, | |
| ; CM-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <8 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <8 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <8 x i8> %load to <8 x i32> | |
| store <8 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG/CM should use DST, but for some there are redundant MOVs | |
| define amdgpu_kernel void @global_zextload_v16i8_to_v16i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v16i8_to_v16i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[0:3], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v19, 24, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v17, v3, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v16, 0xff, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v18, v3, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v7, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v5, v0, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v11, 24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v9, v1, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v15, 24, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v13, v2, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v4, 0xff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v6, v0, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v8, 0xff, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v10, v1, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v12, 0xff, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v14, v2, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v16i8_to_v16i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 48 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v11, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 32 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s0 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 16 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s2 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v10, 24, v3 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v8, v3, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v7, 0xff, v3 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v9, v3, 16, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[11:12], v[7:10] | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v14, 24, v2 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v12, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v11, 0xff, v2 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v13, v2, 16, 8 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v6, 24, v0 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v4, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v10, 24, v1 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v8, v1, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v3, 0xff, v0 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v5, v0, 16, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v7, 0xff, v1 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v9, v1, 16, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[17:18], v[11:14] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[7:10] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[15:16], v[3:6] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v16i8_to_v16i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v19, 24, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v16, 0xff, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v17, 8, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v18, v3, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v7, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v4, 0xff, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v5, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v6, v0, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v11, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v8, 0xff, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v9, 8, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v10, v1, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v15, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v12, 0xff, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v13, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v14, v2, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v16i8_to_v16i32: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @10, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @8 | |
| ; EG-NEXT: ALU 33, @13, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T12.XYZW, T14.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T10.XYZW, T13.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T9.XYZW, T11.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T8.XYZW, T0.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: Fetch clause starting at 8: | |
| ; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 10: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 13: | |
| ; EG-NEXT: MOV * T1.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT * T8.Z, T0.X, literal.x, PV.W, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T8.Y, T0.X, literal.x, T1.W, | |
| ; EG-NEXT: BFE_UINT * T9.Z, T0.Y, literal.y, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_UINT T9.Y, T0.Y, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T8.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T8.X, T0.X, literal.x, | |
| ; EG-NEXT: LSHR T9.W, T0.Y, literal.y, | |
| ; EG-NEXT: AND_INT * T9.X, T0.Y, literal.x, | |
| ; EG-NEXT: 255(3.573311e-43), 24(3.363116e-44) | |
| ; EG-NEXT: BFE_UINT * T10.Z, T0.Z, literal.x, T1.W, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: LSHR T0.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: BFE_UINT * T10.Y, T0.Z, literal.y, T1.W, | |
| ; EG-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T11.X, PV.X, literal.x, | |
| ; EG-NEXT: BFE_UINT T12.Z, T0.W, literal.y, T1.W, | |
| ; EG-NEXT: LSHR T10.W, T0.Z, literal.z, | |
| ; EG-NEXT: AND_INT * T10.X, T0.Z, literal.w, | |
| ; EG-NEXT: 4(5.605194e-45), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 255(3.573311e-43) | |
| ; EG-NEXT: BFE_UINT * T12.Y, T0.W, literal.x, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T13.X, T0.X, literal.x, | |
| ; EG-NEXT: LSHR T12.W, T0.W, literal.y, | |
| ; EG-NEXT: AND_INT * T12.X, T0.W, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T14.X, T0.X, literal.x, | |
| ; EG-NEXT: 12(1.681558e-44), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_zextload_v16i8_to_v16i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @10, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @8 | |
| ; CM-NEXT: ALU 33, @13, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T10.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T12, T14.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T9, T13.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T8, T11.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 8: | |
| ; CM-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 10: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 13: | |
| ; CM-NEXT: MOV * T1.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT * T8.Z, T0.W, literal.x, PV.W, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T8.Y, T0.W, literal.x, T1.W, | |
| ; CM-NEXT: LSHR * T8.W, T0.W, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: AND_INT T8.X, T0.W, literal.x, | |
| ; CM-NEXT: BFE_UINT * T9.Z, T0.Z, literal.y, T1.W, | |
| ; CM-NEXT: 255(3.573311e-43), 16(2.242078e-44) | |
| ; CM-NEXT: LSHR T10.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T9.Y, T0.Z, literal.y, T1.W, | |
| ; CM-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T11.X, PV.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T12.Z, T0.Y, literal.y, T1.W, | |
| ; CM-NEXT: LSHR * T9.W, T0.Z, literal.z, | |
| ; CM-NEXT: 12(1.681558e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T9.X, T0.Z, literal.x, | |
| ; CM-NEXT: BFE_UINT * T12.Y, T0.Y, literal.y, T1.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T13.X, T10.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T0.Z, T0.X, literal.y, T1.W, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR * T12.W, T0.Y, literal.z, | |
| ; CM-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T12.X, T0.Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T0.Y, T0.X, literal.y, T1.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T14.X, T10.X, literal.x, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 4(5.605194e-45), 24(3.363116e-44) | |
| ; CM-NEXT: AND_INT * T0.X, T0.X, literal.x, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <16 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <16 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <16 x i8> %load to <16 x i32> | |
| store <16 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG/CM should use DST, but for some there are redundant MOVs | |
| define amdgpu_kernel void @global_sextload_v16i8_to_v16i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v16i8_to_v16i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[0:3], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v19, 24, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v18, v3, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v17, v3, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v16, v3, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v7, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v6, v0, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v5, v0, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v4, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v11, 24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v10, v1, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v9, v1, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v8, v1, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v15, 24, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v14, v2, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v13, v2, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v12, v2, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v16i8_to_v16i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 48 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 32 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s0 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 16 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s2 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v11, 24, v3 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v10, v3, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v9, v3, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v8, v3, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[12:13], v[8:11] | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v15, 24, v2 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v14, v2, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v12, v2, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 24, v0 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v0, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v0, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v11, 24, v1 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v10, v1, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v9, v1, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v8, v1, 0, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[18:19], v[12:15] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[8:11] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[4:7] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v16i8_to_v16i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v19, 24, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v18, v3, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v3, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v3, 8, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v7, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v6, v0, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v4, v0, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v1, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v8, v1, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v15, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v2, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v12, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v2, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v1, 8, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v0, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v17, v3, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v5, v0, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v9, v1, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v13, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v16i8_to_v16i32: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @10, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @8 | |
| ; EG-NEXT: ALU 41, @13, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T8.XYZW, T14.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T11.XYZW, T13.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T10.XYZW, T12.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T9.XYZW, T0.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: Fetch clause starting at 8: | |
| ; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 10: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 13: | |
| ; EG-NEXT: BFE_INT * T8.X, T0.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T9.X, T0.X, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T1.Y, T0.W, literal.y, | |
| ; EG-NEXT: LSHR T1.Z, T0.W, literal.z, | |
| ; EG-NEXT: LSHR T1.W, T0.Z, literal.y, | |
| ; EG-NEXT: LSHR * T2.W, T0.X, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T10.X, T0.Y, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T2.Y, T0.Z, literal.y, | |
| ; EG-NEXT: LSHR T2.Z, T0.Y, literal.y, | |
| ; EG-NEXT: BFE_INT T9.W, PS, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T2.W, T0.X, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T11.X, T0.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T3.Y, T0.Y, literal.y, | |
| ; EG-NEXT: BFE_INT T9.Z, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T10.W, PV.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T2.W, T0.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; EG-NEXT: LSHR T0.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: BFE_INT T9.Y, PS, 0.0, literal.y, | |
| ; EG-NEXT: BFE_INT T10.Z, PV.Y, 0.0, literal.y, | |
| ; EG-NEXT: BFE_INT T11.W, T2.Y, 0.0, literal.y, | |
| ; EG-NEXT: LSHR * T2.W, T0.Y, literal.y, | |
| ; EG-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T12.X, PV.X, literal.x, | |
| ; EG-NEXT: BFE_INT T10.Y, PS, 0.0, literal.y, | |
| ; EG-NEXT: BFE_INT T11.Z, T1.W, 0.0, literal.y, | |
| ; EG-NEXT: BFE_INT T8.W, T1.Z, 0.0, literal.y, | |
| ; EG-NEXT: LSHR * T1.W, T0.Z, literal.y, | |
| ; EG-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T13.X, T0.X, literal.x, | |
| ; EG-NEXT: BFE_INT T11.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T8.Z, T1.Y, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T0.W, T0.W, literal.x, | |
| ; EG-NEXT: ADD_INT * T14.X, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 12(1.681558e-44) | |
| ; EG-NEXT: BFE_INT * T8.Y, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_sextload_v16i8_to_v16i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @10, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @8 | |
| ; CM-NEXT: ALU 43, @13, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T8, T12.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T11, T0.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T10, T14.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T9, T13.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 8: | |
| ; CM-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 10: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 13: | |
| ; CM-NEXT: LSHR T1.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHR * T1.W, T0.Y, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: BFE_INT T8.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T1.Y, T0.X, literal.y, | |
| ; CM-NEXT: LSHR T2.Z, T0.Y, literal.z, | |
| ; CM-NEXT: LSHR * T2.W, T0.Z, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T9.X, T0.W, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T2.Y, T0.Y, literal.y, | |
| ; CM-NEXT: LSHR T3.Z, T0.Z, literal.z, | |
| ; CM-NEXT: LSHR * T3.W, T0.W, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T10.X, T0.Z, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T3.Y, T0.Z, literal.y, | |
| ; CM-NEXT: LSHR T0.Z, T0.W, literal.z, | |
| ; CM-NEXT: BFE_INT * T9.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T11.X, T0.Y, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T0.Y, T0.W, literal.x, | |
| ; CM-NEXT: BFE_INT T9.Z, PV.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T10.W, PV.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR T12.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: BFE_INT T9.Y, PV.Y, 0.0, literal.y, | |
| ; CM-NEXT: BFE_INT T10.Z, T3.Z, 0.0, literal.y, | |
| ; CM-NEXT: BFE_INT * T11.W, T2.Y, 0.0, literal.y, | |
| ; CM-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T13.X, PV.X, literal.x, | |
| ; CM-NEXT: BFE_INT T10.Y, T2.W, 0.0, literal.y, | |
| ; CM-NEXT: BFE_INT T11.Z, T2.Z, 0.0, literal.y, | |
| ; CM-NEXT: BFE_INT * T8.W, T1.Y, 0.0, literal.y, | |
| ; CM-NEXT: 12(1.681558e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T14.X, T12.X, literal.x, | |
| ; CM-NEXT: BFE_INT T11.Y, T1.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T8.Z, T1.Z, 0.0, literal.x, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T0.X, T12.X, literal.x, | |
| ; CM-NEXT: BFE_INT * T8.Y, PV.W, 0.0, literal.y, | |
| ; CM-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <16 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <16 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <16 x i8> %load to <16 x i32> | |
| store <16 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG/CM should use DST, but for some there are redundant MOVs | |
| define amdgpu_kernel void @global_zextload_v32i8_to_v32i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v32i8_to_v32i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[8:11], v[0:1], s[8:11], 0 addr64 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v3, 24, v4 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v35, 24, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v33, v11, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v32, 0xff, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v34, v11, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v1, v4, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v15, 24, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v13, v5, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v19, 24, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v17, v6, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v23, 24, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v21, v7, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xff, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v2, v4, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v12, 0xff, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v14, v5, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v16, 0xff, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v18, v6, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v20, 0xff, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v22, v7, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v7, 24, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v5, v8, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v27, 24, v9 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v25, v9, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v31, 24, v10 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v29, v10, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v4, 0xff, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v6, v8, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v24, 0xff, v9 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v26, v9, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v28, 0xff, v10 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v30, v10, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[32:35], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[28:31], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[24:27], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[20:23], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v32i8_to_v32i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[4:7], v[0:1] | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, 16, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 48 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 32 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x70 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v11, 24, v7 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v9, v7, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v8, 0xff, v7 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v10, v7, 16, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[8:11] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x60 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v10, 24, v6 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v8, v6, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v7, 0xff, v6 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v9, v6, 16, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[7:10] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x50 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v9, 24, v5 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v7, v5, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v6, 0xff, v5 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v8, v5, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[18:19], v[6:9] | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v5, 0xff, v4 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v8, 24, v4 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v6, v4, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v7, v4, 16, 8 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 64 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[12:13], v[5:8] | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(4) | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v4, 0xff, v3 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v7, 24, v3 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v5, v3, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v6, v3, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[4:7] | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v15, 24, v2 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v13, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v12, 0xff, v2 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v14, v2, 16, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v11, 24, v0 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v9, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v6, 24, v1 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v4, v1, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v8, 0xff, v0 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v10, v0, 16, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v3, 0xff, v1 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v5, v1, 16, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[12:15] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[18:19], v[3:6] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[8:11] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v32i8_to_v32i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v4, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[4:5] | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v4, vcc, 16, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[4:7], v[4:5] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v23, 24, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v20, 0xff, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v21, 8, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v22, v3, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v11, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v8, 0xff, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v9, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v10, v0, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v15, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v12, 0xff, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v13, 8, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v14, v1, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v19, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v16, 0xff, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v17, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v18, v2, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 24, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v0, 0xff, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v1, 8, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v2, v4, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v27, 24, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v24, 0xff, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v25, 8, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v26, v5, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v31, 24, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v28, 0xff, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v29, 8, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v30, v6, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v35, 24, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v32, 0xff, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v33, 8, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v34, v7, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[20:23], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[32:35], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[28:31], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[24:27], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v32i8_to_v32i32: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @16, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 1 @12 | |
| ; EG-NEXT: ALU 64, @19, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T24.XYZW, T26.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T22.XYZW, T25.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T21.XYZW, T23.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T12.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 T14.XYZW, T16.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T13.XYZW, T0.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: Fetch clause starting at 12: | |
| ; EG-NEXT: VTX_READ_128 T12.XYZW, T0.X, 16, #1 | |
| ; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 16: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 5(7.006492e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 19: | |
| ; EG-NEXT: MOV * T1.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT * T13.Z, T0.X, literal.x, PV.W, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T13.Y, T0.X, literal.x, T1.W, | |
| ; EG-NEXT: BFE_UINT * T14.Z, T0.Y, literal.y, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_UINT T14.Y, T0.Y, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T13.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T13.X, T0.X, literal.x, | |
| ; EG-NEXT: LSHR T14.W, T0.Y, literal.y, | |
| ; EG-NEXT: AND_INT * T14.X, T0.Y, literal.x, | |
| ; EG-NEXT: 255(3.573311e-43), 24(3.363116e-44) | |
| ; EG-NEXT: BFE_UINT * T15.Z, T0.Z, literal.x, T1.W, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: LSHR T0.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: BFE_UINT * T15.Y, T0.Z, literal.y, T1.W, | |
| ; EG-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T16.X, PV.X, literal.x, | |
| ; EG-NEXT: BFE_UINT T17.Z, T0.W, literal.y, T1.W, | |
| ; EG-NEXT: LSHR T15.W, T0.Z, literal.z, | |
| ; EG-NEXT: AND_INT * T15.X, T0.Z, literal.w, | |
| ; EG-NEXT: 4(5.605194e-45), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 255(3.573311e-43) | |
| ; EG-NEXT: BFE_UINT * T17.Y, T0.W, literal.x, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T18.X, T0.X, literal.x, | |
| ; EG-NEXT: BFE_UINT T19.Z, T12.X, literal.y, T1.W, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR T17.W, T0.W, literal.z, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: AND_INT * T17.X, T0.W, literal.w, | |
| ; EG-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 255(3.573311e-43) | |
| ; EG-NEXT: BFE_UINT * T19.Y, T12.X, literal.x, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T20.X, T0.X, literal.x, | |
| ; EG-NEXT: BFE_UINT T21.Z, T12.Y, literal.y, T1.W, | |
| ; EG-NEXT: LSHR T19.W, T12.X, literal.z, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: AND_INT * T19.X, T12.X, literal.w, | |
| ; EG-NEXT: 12(1.681558e-44), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 255(3.573311e-43) | |
| ; EG-NEXT: BFE_UINT * T21.Y, T12.Y, literal.x, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T12.X, T0.X, literal.x, | |
| ; EG-NEXT: BFE_UINT T22.Z, T12.Z, literal.x, T1.W, | |
| ; EG-NEXT: LSHR T21.W, T12.Y, literal.y, | |
| ; EG-NEXT: AND_INT * T21.X, T12.Y, literal.z, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT * T22.Y, T12.Z, literal.x, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T23.X, T0.X, literal.x, | |
| ; EG-NEXT: BFE_UINT T24.Z, T12.W, literal.y, T1.W, | |
| ; EG-NEXT: LSHR T22.W, T12.Z, literal.z, | |
| ; EG-NEXT: AND_INT * T22.X, T12.Z, literal.w, | |
| ; EG-NEXT: 20(2.802597e-44), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 255(3.573311e-43) | |
| ; EG-NEXT: BFE_UINT * T24.Y, T12.W, literal.x, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T25.X, T0.X, literal.x, | |
| ; EG-NEXT: LSHR T24.W, T12.W, literal.x, | |
| ; EG-NEXT: AND_INT * T24.X, T12.W, literal.y, | |
| ; EG-NEXT: 24(3.363116e-44), 255(3.573311e-43) | |
| ; EG-NEXT: ADD_INT * T26.X, T0.X, literal.x, | |
| ; EG-NEXT: 28(3.923636e-44), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_zextload_v32i8_to_v32i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @16, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 1 @12 | |
| ; CM-NEXT: ALU 63, @19, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T12, T15.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T24, T26.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T22, T25.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T20, T23.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T21.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T17, T19.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T14, T18.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T13, T16.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 12: | |
| ; CM-NEXT: VTX_READ_128 T12.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: VTX_READ_128 T0.XYZW, T0.X, 16, #1 | |
| ; CM-NEXT: ALU clause starting at 16: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 5(7.006492e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 19: | |
| ; CM-NEXT: MOV * T1.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT * T13.Z, T0.W, literal.x, PV.W, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T13.Y, T0.W, literal.x, T1.W, | |
| ; CM-NEXT: LSHR * T13.W, T0.W, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: AND_INT T13.X, T0.W, literal.x, | |
| ; CM-NEXT: BFE_UINT * T14.Z, T0.Z, literal.y, T1.W, | |
| ; CM-NEXT: 255(3.573311e-43), 16(2.242078e-44) | |
| ; CM-NEXT: LSHR T15.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T14.Y, T0.Z, literal.y, T1.W, | |
| ; CM-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T16.X, PV.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T17.Z, T0.Y, literal.y, T1.W, | |
| ; CM-NEXT: LSHR * T14.W, T0.Z, literal.z, | |
| ; CM-NEXT: 28(3.923636e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T14.X, T0.Z, literal.x, | |
| ; CM-NEXT: BFE_UINT * T17.Y, T0.Y, literal.y, T1.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T18.X, T15.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T0.Z, T0.X, literal.y, T1.W, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR * T17.W, T0.Y, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; CM-NEXT: AND_INT T17.X, T0.Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T0.Y, T0.X, literal.y, T1.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T19.X, T15.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T20.Z, T12.W, literal.y, T1.W, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 20(2.802597e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.X, T0.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T20.Y, T12.W, literal.y, T1.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T21.X, T15.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T22.Z, T12.Z, literal.x, T1.W, | |
| ; CM-NEXT: LSHR * T20.W, T12.W, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: AND_INT T20.X, T12.W, literal.x, | |
| ; CM-NEXT: BFE_UINT * T22.Y, T12.Z, literal.y, T1.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T23.X, T15.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T24.Z, T12.Y, literal.y, T1.W, | |
| ; CM-NEXT: LSHR * T22.W, T12.Z, literal.z, | |
| ; CM-NEXT: 12(1.681558e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T22.X, T12.Z, literal.x, | |
| ; CM-NEXT: BFE_UINT * T24.Y, T12.Y, literal.y, T1.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T25.X, T15.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T12.Z, T12.X, literal.y, T1.W, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR * T24.W, T12.Y, literal.z, | |
| ; CM-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T24.X, T12.Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T12.Y, T12.X, literal.y, T1.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T26.X, T15.X, literal.x, | |
| ; CM-NEXT: LSHR * T12.W, T12.X, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 4(5.605194e-45), 24(3.363116e-44) | |
| ; CM-NEXT: AND_INT * T12.X, T12.X, literal.x, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <32 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <32 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <32 x i8> %load to <32 x i32> | |
| store <32 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG/CM should use DST, but for some there are redundant MOVs | |
| define amdgpu_kernel void @global_sextload_v32i8_to_v32i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v32i8_to_v32i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v4, 5, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v5, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[0:3], v[4:5], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[4:7], v[4:5], s[8:11], 0 addr64 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v11, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v35, 24, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v34, v7, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v33, v7, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v32, v7, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v10, v0, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v9, v0, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v8, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v15, 24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v14, v1, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v13, v1, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v12, v1, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v19, 24, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v18, v2, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v17, v2, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v16, v2, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v23, 24, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v22, v3, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v21, v3, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v20, v3, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v3, 24, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v4, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v1, v4, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v4, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v27, 24, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v26, v5, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v25, v5, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v24, v5, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v31, 24, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v30, v6, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v29, v6, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v28, v6, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[32:35], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[28:31], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[24:27], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[20:23], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v32i8_to_v32i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[4:7], v[0:1] | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, 16, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 48 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 32 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x70 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v10, 24, v7 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v9, v7, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v8, v7, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v7, v7, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[7:10] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x60 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v9, 24, v6 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v8, v6, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v7, v6, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v6, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[6:9] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x50 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v8, 24, v5 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v7, v5, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v5, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v5, 0, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[18:19], v[5:8] | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 64 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 24, v4 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v4, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v4, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v4, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[12:13], v[4:7] | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(4) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v6, 24, v3 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v3, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v3, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v3, v3, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[3:6] | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v15, 24, v2 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v14, v2, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v12, v2, 0, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s2 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v11, 24, v0 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v10, v0, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v9, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v8, v0, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v6, 24, v1 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v1, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v1, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v3, v1, 0, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[12:15] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[18:19], v[3:6] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[8:11] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v32i8_to_v32i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v4, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[4:5] | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v4, vcc, 16, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[4:7], v[4:5] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v9, 8, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v17, v9, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v19, 24, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v18, v3, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v3, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v15, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v1, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v12, v1, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v1, 8, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v19, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v18, v2, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v2, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v9, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v17, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v13, v1, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v0, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v8, v0, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(3) | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v19, 24, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v15, 24, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v5, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v12, v5, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v18, v6, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v6, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v23, 24, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v22, v7, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v20, v7, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v7, 8, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v6, 8, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v5, 8, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v1, 8, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v9, v9, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v3, 24, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v2, v4, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v0, v4, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v1, v1, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v13, v5, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v17, v6, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v21, v7, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[20:23], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v32i8_to_v32i32: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @18, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @14 | |
| ; EG-NEXT: ALU 9, @21, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @16 | |
| ; EG-NEXT: ALU 73, @31, KC0[], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T24.XYZW, T26.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T23.XYZW, T25.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T17.XYZW, T12.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T22.XYZW, T0.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T21.XYZW, T16.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T15.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T14.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T18.XYZW, T13.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: Fetch clause starting at 14: | |
| ; EG-NEXT: VTX_READ_128 T12.XYZW, T0.X, 16, #1 | |
| ; EG-NEXT: Fetch clause starting at 16: | |
| ; EG-NEXT: VTX_READ_128 T17.XYZW, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 18: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 5(7.006492e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 21: | |
| ; EG-NEXT: LSHR * T13.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T14.X, PV.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T15.X, PV.X, literal.y, | |
| ; EG-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T16.X, T13.X, literal.x, | |
| ; EG-NEXT: LSHR T0.W, T12.W, literal.y, | |
| ; EG-NEXT: LSHR * T1.W, T12.W, literal.z, | |
| ; EG-NEXT: 12(1.681558e-44), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ALU clause starting at 31: | |
| ; EG-NEXT: LSHR T2.W, T12.Z, literal.x, | |
| ; EG-NEXT: LSHR * T3.W, T12.Z, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: ADD_INT T0.X, T13.X, literal.x, | |
| ; EG-NEXT: LSHR T0.Y, T12.Y, literal.x, | |
| ; EG-NEXT: LSHR T0.Z, T12.Y, literal.y, | |
| ; EG-NEXT: LSHR T4.W, T12.X, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR * T5.W, T12.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: BFE_INT T18.X, T17.X, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T1.Y, T17.W, literal.y, | |
| ; EG-NEXT: LSHR T1.Z, T17.W, literal.z, | |
| ; EG-NEXT: LSHR T6.W, T17.Z, literal.y, | |
| ; EG-NEXT: LSHR * T7.W, T17.X, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T19.X, T17.Y, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T2.Y, T17.Z, literal.y, | |
| ; EG-NEXT: LSHR T2.Z, T17.Y, literal.y, | |
| ; EG-NEXT: BFE_INT T18.W, PS, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T7.W, T17.X, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T20.X, T17.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T3.Y, T17.Y, literal.y, | |
| ; EG-NEXT: BFE_INT T18.Z, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T19.W, PV.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T7.W, T17.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT T21.X, T17.W, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T18.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T19.Z, PV.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T20.W, T2.Y, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T7.W, T17.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T22.X, T12.X, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T19.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T20.Z, T6.W, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T21.W, T1.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T6.W, T17.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T17.X, T12.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T20.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T21.Z, T1.Y, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: BFE_INT T22.W, T5.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T5.W, T17.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T23.X, T12.Z, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T21.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T22.Z, T4.W, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T17.W, T0.Z, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR * T4.W, T12.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T24.X, T12.W, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T22.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T17.Z, T0.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T23.W, T3.W, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR * T3.W, T12.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T12.X, T13.X, literal.x, | |
| ; EG-NEXT: BFE_INT T17.Y, PS, 0.0, literal.y, | |
| ; EG-NEXT: BFE_INT T23.Z, T2.W, 0.0, literal.y, | |
| ; EG-NEXT: BFE_INT T24.W, T1.W, 0.0, literal.y, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR * T1.W, T12.Z, literal.y, | |
| ; EG-NEXT: 20(2.802597e-44), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T25.X, T13.X, literal.x, | |
| ; EG-NEXT: BFE_INT T23.Y, PS, 0.0, literal.y, | |
| ; EG-NEXT: BFE_INT T24.Z, T0.W, 0.0, literal.y, | |
| ; EG-NEXT: LSHR T0.W, T12.W, literal.y, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: ADD_INT * T26.X, T13.X, literal.z, | |
| ; EG-NEXT: 24(3.363116e-44), 8(1.121039e-44) | |
| ; EG-NEXT: 28(3.923636e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT * T24.Y, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_sextload_v32i8_to_v32i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @18, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @14 | |
| ; CM-NEXT: ALU 8, @21, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @16 | |
| ; CM-NEXT: ALU 78, @30, KC0[], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T24, T13.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T23, T12.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T22, T26.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T21, T25.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T14, T17.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T20, T16.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T19, T15.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T18, T0.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 14: | |
| ; CM-NEXT: VTX_READ_128 T12.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: Fetch clause starting at 16: | |
| ; CM-NEXT: VTX_READ_128 T14.XYZW, T0.X, 16, #1 | |
| ; CM-NEXT: ALU clause starting at 18: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 5(7.006492e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 21: | |
| ; CM-NEXT: LSHR T0.Z, T12.X, literal.x, | |
| ; CM-NEXT: LSHR * T0.W, T12.Y, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: LSHR T13.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: LSHR T0.Y, T12.X, literal.y, | |
| ; CM-NEXT: LSHR T1.Z, T12.Y, literal.z, | |
| ; CM-NEXT: LSHR * T1.W, T12.Z, literal.w, | |
| ; CM-NEXT: 2(2.802597e-45), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ALU clause starting at 30: | |
| ; CM-NEXT: ADD_INT T0.X, T13.X, literal.x, | |
| ; CM-NEXT: LSHR T1.Y, T12.Y, literal.y, | |
| ; CM-NEXT: LSHR T2.Z, T12.Z, literal.z, | |
| ; CM-NEXT: LSHR * T2.W, T12.W, literal.w, | |
| ; CM-NEXT: 28(3.923636e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T15.X, T13.X, literal.x, | |
| ; CM-NEXT: LSHR T2.Y, T12.Z, literal.x, | |
| ; CM-NEXT: LSHR T3.Z, T12.W, literal.y, | |
| ; CM-NEXT: LSHR * T3.W, T14.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T16.X, T13.X, literal.x, | |
| ; CM-NEXT: LSHR T3.Y, T12.W, literal.y, | |
| ; CM-NEXT: LSHR T4.Z, T14.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR * T4.W, T14.Y, literal.w, | |
| ; CM-NEXT: 20(2.802597e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T17.X, T13.X, literal.x, | |
| ; CM-NEXT: LSHR T4.Y, T14.X, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR T5.Z, T14.Y, literal.x, | |
| ; CM-NEXT: LSHR * T5.W, T14.Z, literal.z, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T18.X, T14.W, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T5.Y, T14.Y, literal.y, | |
| ; CM-NEXT: LSHR T6.Z, T14.Z, literal.z, | |
| ; CM-NEXT: LSHR * T6.W, T14.W, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T19.X, T14.Z, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T6.Y, T14.Z, literal.y, | |
| ; CM-NEXT: LSHR T7.Z, T14.W, literal.z, | |
| ; CM-NEXT: BFE_INT * T18.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T20.X, T14.Y, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T7.Y, T14.W, literal.x, | |
| ; CM-NEXT: BFE_INT T18.Z, PV.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T19.W, PV.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T14.X, T14.X, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T18.Y, PV.Y, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T19.Z, T6.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T20.W, T5.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T21.X, T12.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T19.Y, T5.W, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: BFE_INT T20.Z, T5.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T14.W, T4.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T22.X, T12.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T20.Y, T4.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T14.Z, T4.Z, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: BFE_INT * T21.W, T3.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T23.X, T12.Y, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T14.Y, T3.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T21.Z, T3.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T22.W, T2.Y, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T24.X, T12.X, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T21.Y, T2.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T22.Z, T2.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T23.W, T1.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T25.X, T13.X, literal.x, | |
| ; CM-NEXT: BFE_INT T22.Y, T1.W, 0.0, literal.y, | |
| ; CM-NEXT: BFE_INT T23.Z, T1.Z, 0.0, literal.y, | |
| ; CM-NEXT: BFE_INT * T24.W, T0.Y, 0.0, literal.y, | |
| ; CM-NEXT: 12(1.681558e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T26.X, T13.X, literal.x, | |
| ; CM-NEXT: BFE_INT T23.Y, T0.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T24.Z, T0.Z, 0.0, literal.x, | |
| ; CM-NEXT: LSHR * T0.W, T12.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T12.X, T13.X, literal.x, | |
| ; CM-NEXT: BFE_INT * T24.Y, PV.W, 0.0, literal.y, | |
| ; CM-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <32 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <32 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <32 x i8> %load to <32 x i32> | |
| store <32 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v64i8_to_v64i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v64i8_to_v64i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s14, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s15, 0xe8f000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_add_u32 s12, s12, s11 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 6, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[16:19], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[32:35], v[0:1], s[8:11], 0 addr64 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[36:39], v[0:1], s[8:11], 0 addr64 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[40:43], v[0:1], s[8:11], 0 addr64 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: s_addc_u32 s13, s13, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(3) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v3, 24, v17 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v1, v17, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xff, v17 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v2, v17, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[12:15], 0 ; 4-byte Folded Spill | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v1, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v2, off, s[12:15], 0 offset:8 ; 4-byte Folded Spill | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v3, off, s[12:15], 0 offset:12 ; 4-byte Folded Spill | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(4) expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v3, 24, v42 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v1, v42, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xff, v42 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v2, v42, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v7, 24, v16 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v5, v16, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v11, 24, v19 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v9, v19, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v15, 24, v18 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v13, v18, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v4, 0xff, v16 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v6, v16, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v8, 0xff, v19 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v10, v19, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v12, 0xff, v18 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v14, v18, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v19, 24, v33 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v17, v33, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v23, 24, v32 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v21, v32, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v27, 24, v35 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v25, v35, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v31, 24, v34 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v29, v34, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v16, 0xff, v33 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v18, v33, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v20, 0xff, v32 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v22, v32, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v24, 0xff, v35 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v26, v35, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v28, 0xff, v34 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v30, v34, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v35, 24, v37 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v33, v37, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v47, 24, v36 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v45, v36, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v51, 24, v39 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v49, v39, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v55, 24, v38 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v53, v38, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v32, 0xff, v37 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v34, v37, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v44, 0xff, v36 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v46, v36, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v48, 0xff, v39 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v50, v39, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v52, 0xff, v38 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v54, v38, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v39, 24, v41 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v37, v41, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v59, 24, v40 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v57, v40, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v63, 24, v43 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v61, v43, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v36, 0xff, v41 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v38, v41, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v56, 0xff, v40 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v58, v40, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v60, 0xff, v43 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v62, v43, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:224 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[60:63], off, s[0:3], 0 offset:240 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[56:59], off, s[0:3], 0 offset:192 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[36:39], off, s[0:3], 0 offset:208 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[52:55], off, s[0:3], 0 offset:160 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[48:51], off, s[0:3], 0 offset:176 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[44:47], off, s[0:3], 0 offset:128 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[32:35], off, s[0:3], 0 offset:144 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[28:31], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[24:27], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[20:23], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 ; 4-byte Folded Reload | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v2, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v3, off, s[12:15], 0 offset:12 ; 4-byte Folded Reload | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v64i8_to_v64i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 6, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v4, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[12:15], v[4:5] | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, 48, v4 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v5, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[8:11], v[0:1] | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, 16, v4 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v5, vcc | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v4, vcc, 32, v4 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[4:7], v[4:5] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 32 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v21, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v20, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 48 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(3) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v19, 24, v14 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v17, v14, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v16, 0xff, v14 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v18, v14, 16, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[20:21], v[16:19] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v19, 24, v15 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v17, v15, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v16, 0xff, v15 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v18, v15, 16, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s3 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[16:19] | |
| ; GCN-HSA-NEXT: v_bfe_u32 v15, v12, 8, 8 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v17, 24, v12 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v14, 0xff, v12 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v16, v12, 16, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s0 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[18:19], v[14:17] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v17, 24, v13 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v15, v13, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v14, 0xff, v13 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v16, v13, 16, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, s3 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xe0 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xf0 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[12:13], v[14:17] | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(6) | |
| ; GCN-HSA-NEXT: v_bfe_u32 v13, v10, 8, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xc0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v15, 24, v10 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v12, 0xff, v10 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v14, v10, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[18:19], v[12:15] | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v18, 24, v8 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v15, 24, v11 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v13, v11, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v12, 0xff, v11 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v14, v11, 16, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[12:15] | |
| ; GCN-HSA-NEXT: v_bfe_u32 v16, v8, 8, 8 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v14, 24, v9 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v12, v9, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v11, 0xff, v9 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v13, v9, 16, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v15, 0xff, v8 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v17, v8, 16, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xd0 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v20, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xa0 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[8:9], v[15:18] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[19:20], v[11:14] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xb0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(8) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v14, 24, v6 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v12, v6, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v11, 0xff, v6 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v13, v6, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[15:16], v[11:14] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v6, s2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v14, 24, v7 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v12, v7, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v11, 0xff, v7 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v13, v7, 16, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v7, s3 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x80 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[6:7], v[11:14] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v7, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v6, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x90 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v14, 24, v4 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v12, v4, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v11, 0xff, v4 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v13, v4, 16, 8 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v18, 24, v5 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v16, v5, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v15, 0xff, v5 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v17, v5, 16, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x60 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v10, 24, v1 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v8, v1, 8, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[6:7], v[11:14] | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v7, 0xff, v1 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v14, 24, v0 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v12, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v9, v1, 16, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v11, 0xff, v0 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v13, v0, 16, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x70 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[15:18] | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v6, 24, v2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v18, 24, v3 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v16, v3, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v4, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v15, 0xff, v3 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v17, v3, 16, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v3, 0xff, v2 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v5, v2, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[3:6] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 64 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 0x50 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[15:18] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[11:14] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[7:10] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v64i8_to_v64i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 6, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v4, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v6, vcc, 16, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v5, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[4:5] | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[12:15], v[6:7] | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v6, vcc, 32, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v5, vcc | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v4, vcc, 48, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[36:39], v[6:7] | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[40:43], v[4:5] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(3) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v51, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v48, 0xff, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v49, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v50, v2, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v27, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v24, 0xff, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v25, 8, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v26, v1, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v35, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v32, 0xff, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v33, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v34, v0, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v47, 24, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v44, 0xff, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v45, 8, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v46, v3, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(2) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 24, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v0, 0xff, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v1, 8, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v2, v13, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v7, 24, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v4, 0xff, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v5, 8, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v6, v12, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v11, 24, v15 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v8, 0xff, v15 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v9, 8, v15 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v10, v15, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v15, 24, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v12, 0xff, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v13, 8, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v14, v14, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v19, 24, v37 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v16, 0xff, v37 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v17, 8, v37 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v18, v37, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v23, 24, v36 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v20, 0xff, v36 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v21, 8, v36 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v22, v36, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v31, 24, v39 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v28, 0xff, v39 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v29, 8, v39 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v30, v39, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v39, 24, v38 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v36, 0xff, v38 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v37, 8, v38 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v38, v38, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v55, 24, v41 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v52, 0xff, v41 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v53, 8, v41 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v54, v41, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v59, 24, v40 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v56, 0xff, v40 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v57, 8, v40 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v58, v40, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v63, 24, v43 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v60, 0xff, v43 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v61, 8, v43 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v62, v43, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v43, 24, v42 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v40, 0xff, v42 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v41, 8, v42 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v42, v42, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[48:51], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[44:47], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[32:35], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[24:27], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[40:43], off, s[0:3], 0 offset:224 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[60:63], off, s[0:3], 0 offset:240 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[56:59], off, s[0:3], 0 offset:192 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[52:55], off, s[0:3], 0 offset:208 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[36:39], off, s[0:3], 0 offset:160 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[28:31], off, s[0:3], 0 offset:176 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[20:23], off, s[0:3], 0 offset:128 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:144 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v64i8_to_v64i32: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @30, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 1 @22 | |
| ; EG-NEXT: ALU 37, @33, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 1 @26 | |
| ; EG-NEXT: ALU 92, @71, KC0[], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T48.XYZW, T50.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T47.XYZW, T49.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T45.XYZW, T29.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 T40.XYZW, T42.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T38.XYZW, T30.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T36.XYZW, T39.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T34.XYZW, T37.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T33.XYZW, T35.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T31.XYZW, T22.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T21.XYZW, T32.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T27.XYZW, T28.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T25.XYZW, T23.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T24.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T26.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: Fetch clause starting at 22: | |
| ; EG-NEXT: VTX_READ_128 T22.XYZW, T21.X, 16, #1 | |
| ; EG-NEXT: VTX_READ_128 T23.XYZW, T21.X, 0, #1 | |
| ; EG-NEXT: Fetch clause starting at 26: | |
| ; EG-NEXT: VTX_READ_128 T29.XYZW, T21.X, 48, #1 | |
| ; EG-NEXT: VTX_READ_128 T30.XYZW, T21.X, 32, #1 | |
| ; EG-NEXT: ALU clause starting at 30: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 6(8.407791e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T21.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 33: | |
| ; EG-NEXT: MOV * T1.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT * T0.Z, T23.Y, literal.x, PV.W, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T0.Y, T23.Y, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T0.W, T23.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T0.X, T23.Y, literal.x, | |
| ; EG-NEXT: BFE_UINT T20.Z, T23.X, literal.y, T1.W, | |
| ; EG-NEXT: LSHR * T24.X, KC0[2].Y, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 16(2.242078e-44) | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T20.Y, T23.X, literal.x, T1.W, | |
| ; EG-NEXT: BFE_UINT * T25.Z, T23.W, literal.y, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; EG-NEXT: ADD_INT T26.X, T24.X, literal.x, | |
| ; EG-NEXT: BFE_UINT T25.Y, T23.W, literal.y, T1.W, | |
| ; EG-NEXT: LSHR T20.W, T23.X, literal.z, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: AND_INT * T20.X, T23.X, literal.w, | |
| ; EG-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 255(3.573311e-43) | |
| ; EG-NEXT: BFE_UINT T27.Z, T23.Z, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T25.W, T23.W, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T25.X, T23.W, literal.x, | |
| ; EG-NEXT: BFE_UINT T27.Y, T23.Z, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T23.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; EG-NEXT: 12(1.681558e-44), 0(0.000000e+00) | |
| ; EG-NEXT: LSHR * T27.W, T23.Z, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T27.X, T23.Z, literal.x, | |
| ; EG-NEXT: BFE_UINT T21.Z, T22.Y, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T28.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 16(2.242078e-44) | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT * T21.Y, T22.Y, literal.x, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ALU clause starting at 71: | |
| ; EG-NEXT: BFE_UINT T31.Z, T22.X, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T21.W, T22.Y, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T21.X, T22.Y, literal.x, | |
| ; EG-NEXT: BFE_UINT T31.Y, T22.X, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T32.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; EG-NEXT: 20(2.802597e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T33.Z, T22.W, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T31.W, T22.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T31.X, T22.X, literal.x, | |
| ; EG-NEXT: BFE_UINT T33.Y, T22.W, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T22.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T34.Z, T22.Z, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T33.W, T22.W, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T33.X, T22.W, literal.x, | |
| ; EG-NEXT: BFE_UINT T34.Y, T22.Z, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T35.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; EG-NEXT: 28(3.923636e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T36.Z, T30.Y, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T34.W, T22.Z, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T34.X, T22.Z, literal.x, | |
| ; EG-NEXT: BFE_UINT T36.Y, T30.Y, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T37.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T38.Z, T30.X, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T36.W, T30.Y, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T36.X, T30.Y, literal.x, | |
| ; EG-NEXT: BFE_UINT T38.Y, T30.X, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T39.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; EG-NEXT: 36(5.044674e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T40.Z, T30.W, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T38.W, T30.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T38.X, T30.X, literal.x, | |
| ; EG-NEXT: BFE_UINT T40.Y, T30.W, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T30.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; EG-NEXT: 32(4.484155e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T41.Z, T30.Z, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T40.W, T30.W, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T40.X, T30.W, literal.x, | |
| ; EG-NEXT: BFE_UINT T41.Y, T30.Z, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T42.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; EG-NEXT: 44(6.165713e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T43.Z, T29.Y, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T41.W, T30.Z, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T41.X, T30.Z, literal.x, | |
| ; EG-NEXT: BFE_UINT T43.Y, T29.Y, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T44.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; EG-NEXT: 40(5.605194e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T45.Z, T29.X, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T43.W, T29.Y, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T43.X, T29.Y, literal.x, | |
| ; EG-NEXT: BFE_UINT T45.Y, T29.X, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T46.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; EG-NEXT: 52(7.286752e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T47.Z, T29.W, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T45.W, T29.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T45.X, T29.X, literal.x, | |
| ; EG-NEXT: BFE_UINT T47.Y, T29.W, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T29.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; EG-NEXT: 48(6.726233e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T48.Z, T29.Z, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T47.W, T29.W, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: AND_INT T47.X, T29.W, literal.x, | |
| ; EG-NEXT: BFE_UINT T48.Y, T29.Z, literal.y, T1.W, | |
| ; EG-NEXT: ADD_INT * T49.X, T24.X, literal.z, | |
| ; EG-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; EG-NEXT: 60(8.407791e-44), 0(0.000000e+00) | |
| ; EG-NEXT: LSHR * T48.W, T29.Z, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T48.X, T29.Z, literal.x, | |
| ; EG-NEXT: ADD_INT * T50.X, T24.X, literal.y, | |
| ; EG-NEXT: 255(3.573311e-43), 56(7.847271e-44) | |
| ; | |
| ; CM-LABEL: global_zextload_v64i8_to_v64i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @30, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 1 @22 | |
| ; CM-NEXT: ALU 36, @33, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 1 @26 | |
| ; CM-NEXT: ALU 91, @70, KC0[], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T49, T30.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T48, T24.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T46, T50.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T44, T47.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T43, T45.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T41, T31.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T39, T42.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T37, T40.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T36, T38.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T34, T20.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T32, T35.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T29, T33.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T28, T0.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T26, T21.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T23, T27.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T22, T25.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 22: | |
| ; CM-NEXT: VTX_READ_128 T20.XYZW, T0.X, 32, #1 | |
| ; CM-NEXT: VTX_READ_128 T21.XYZW, T0.X, 48, #1 | |
| ; CM-NEXT: Fetch clause starting at 26: | |
| ; CM-NEXT: VTX_READ_128 T30.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: VTX_READ_128 T31.XYZW, T0.X, 16, #1 | |
| ; CM-NEXT: ALU clause starting at 30: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 6(8.407791e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 33: | |
| ; CM-NEXT: MOV * T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT * T22.Z, T21.Z, literal.x, PV.W, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T22.Y, T21.Z, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T22.W, T21.Z, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: AND_INT T22.X, T21.Z, literal.x, | |
| ; CM-NEXT: BFE_UINT * T23.Z, T21.W, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 16(2.242078e-44) | |
| ; CM-NEXT: LSHR T24.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T23.Y, T21.W, literal.y, T0.W, | |
| ; CM-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T25.X, PV.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T26.Z, T21.X, literal.y, T0.W, | |
| ; CM-NEXT: LSHR * T23.W, T21.W, literal.z, | |
| ; CM-NEXT: 56(7.847271e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T23.X, T21.W, literal.x, | |
| ; CM-NEXT: BFE_UINT * T26.Y, T21.X, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T27.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T28.Z, T21.Y, literal.y, T0.W, | |
| ; CM-NEXT: LSHR * T26.W, T21.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 60(8.407791e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T26.X, T21.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T28.Y, T21.Y, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T21.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T29.Z, T20.Z, literal.y, T0.W, | |
| ; CM-NEXT: LSHR * T28.W, T21.Y, literal.z, | |
| ; CM-NEXT: 48(6.726233e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T28.X, T21.Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T29.Y, T20.Z, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ALU clause starting at 70: | |
| ; CM-NEXT: ADD_INT T0.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T32.Z, T20.W, literal.y, T0.W, | |
| ; CM-NEXT: LSHR * T29.W, T20.Z, literal.z, | |
| ; CM-NEXT: 52(7.286752e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T29.X, T20.Z, literal.x, | |
| ; CM-NEXT: BFE_UINT * T32.Y, T20.W, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T33.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T34.Z, T20.X, literal.y, T0.W, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR * T32.W, T20.W, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 40(5.605194e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T32.X, T20.W, literal.x, | |
| ; CM-NEXT: BFE_UINT * T34.Y, T20.X, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T35.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T36.Z, T20.Y, literal.y, T0.W, | |
| ; CM-NEXT: LSHR * T34.W, T20.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 44(6.165713e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T34.X, T20.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T36.Y, T20.Y, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T20.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T37.Z, T31.Z, literal.y, T0.W, | |
| ; CM-NEXT: LSHR * T36.W, T20.Y, literal.z, | |
| ; CM-NEXT: 32(4.484155e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T36.X, T20.Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T37.Y, T31.Z, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T38.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T39.Z, T31.W, literal.y, T0.W, | |
| ; CM-NEXT: LSHR * T37.W, T31.Z, literal.z, | |
| ; CM-NEXT: 36(5.044674e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T37.X, T31.Z, literal.x, | |
| ; CM-NEXT: BFE_UINT * T39.Y, T31.W, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T40.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T41.Z, T31.X, literal.y, T0.W, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR * T39.W, T31.W, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; CM-NEXT: AND_INT T39.X, T31.W, literal.x, | |
| ; CM-NEXT: BFE_UINT * T41.Y, T31.X, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T42.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T43.Z, T31.Y, literal.y, T0.W, | |
| ; CM-NEXT: LSHR * T41.W, T31.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 28(3.923636e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T41.X, T31.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T43.Y, T31.Y, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T31.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T44.Z, T30.Z, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T43.W, T31.Y, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: AND_INT T43.X, T31.Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T44.Y, T30.Z, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T45.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T46.Z, T30.W, literal.y, T0.W, | |
| ; CM-NEXT: LSHR * T44.W, T30.Z, literal.z, | |
| ; CM-NEXT: 20(2.802597e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T44.X, T30.Z, literal.x, | |
| ; CM-NEXT: BFE_UINT * T46.Y, T30.W, literal.y, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T47.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T48.Z, T30.X, literal.y, T0.W, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR * T46.W, T30.W, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T46.X, T30.W, literal.x, | |
| ; CM-NEXT: BFE_UINT T48.Y, T30.X, literal.y, T0.W, | |
| ; CM-NEXT: BFE_UINT * T49.Z, T30.Y, literal.z, T0.W, | |
| ; CM-NEXT: 255(3.573311e-43), 8(1.121039e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T50.X, T24.X, literal.x, | |
| ; CM-NEXT: BFE_UINT T49.Y, T30.Y, literal.y, T0.W, | |
| ; CM-NEXT: LSHR * T48.W, T30.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 12(1.681558e-44), 8(1.121039e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T48.X, T30.X, literal.x, | |
| ; CM-NEXT: LSHR * T49.W, T30.Y, literal.y, | |
| ; CM-NEXT: 255(3.573311e-43), 24(3.363116e-44) | |
| ; CM-NEXT: AND_INT * T49.X, T30.Y, literal.x, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T30.X, T24.X, literal.x, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <64 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <64 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <64 x i8> %load to <64 x i32> | |
| store <64 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v64i8_to_v64i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v64i8_to_v64i32: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s14, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s15, 0xe8f000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_add_u32 s12, s12, s11 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 6, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[10:13], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[26:29], v[0:1], s[8:11], 0 addr64 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[32:35], v[0:1], s[8:11], 0 addr64 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[36:39], v[0:1], s[8:11], 0 addr64 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: s_addc_u32 s13, s13, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(3) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v3, 24, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v11, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v1, v11, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v11, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[12:15], 0 ; 4-byte Folded Spill | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v1, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v2, off, s[12:15], 0 offset:8 ; 4-byte Folded Spill | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v3, off, s[12:15], 0 offset:12 ; 4-byte Folded Spill | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(4) expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v3, 24, v38 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v38, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v1, v38, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v38, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v7, 24, v10 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v6, v10, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v5, v10, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v4, v10, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v11, 24, v13 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v10, v13, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v9, v13, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v8, v13, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v15, 24, v12 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v14, v12, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v13, v12, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v12, v12, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v19, 24, v27 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v18, v27, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v17, v27, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v16, v27, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v23, 24, v26 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v22, v26, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v21, v26, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v20, v26, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v27, 24, v29 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v26, v29, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v25, v29, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v24, v29, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v31, 24, v28 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v30, v28, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v29, v28, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v28, v28, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v43, 24, v33 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v42, v33, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v41, v33, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v40, v33, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v47, 24, v32 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v46, v32, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v45, v32, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v44, v32, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v51, 24, v35 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v50, v35, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v49, v35, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v48, v35, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v55, 24, v34 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v54, v34, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v53, v34, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v52, v34, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v35, 24, v37 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v34, v37, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v33, v37, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v32, v37, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v59, 24, v36 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v58, v36, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v57, v36, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v56, v36, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v63, 24, v39 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v62, v39, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v61, v39, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v60, v39, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:224 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[60:63], off, s[0:3], 0 offset:240 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[56:59], off, s[0:3], 0 offset:192 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[32:35], off, s[0:3], 0 offset:208 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[52:55], off, s[0:3], 0 offset:160 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[48:51], off, s[0:3], 0 offset:176 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[44:47], off, s[0:3], 0 offset:128 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[40:43], off, s[0:3], 0 offset:144 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[28:31], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[24:27], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[20:23], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 ; 4-byte Folded Reload | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v2, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v3, off, s[12:15], 0 offset:12 ; 4-byte Folded Reload | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v64i8_to_v64i32: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 6, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v4, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[12:15], v[4:5] | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, 48, v4 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v5, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[8:11], v[0:1] | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, 16, v4 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v5, vcc | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v4, vcc, 32, v4 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[4:7], v[4:5] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 32 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v21, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v20, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 48 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(3) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v19, 24, v14 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v18, v14, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v17, v14, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v16, v14, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[20:21], v[16:19] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v20, s3 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v18, 24, v15 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v17, v15, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v16, v15, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v15, v15, 0, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[19:20], v[15:18] | |
| ; GCN-HSA-NEXT: v_bfe_i32 v14, v12, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v17, 24, v12 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v16, v12, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v15, v12, 8, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s0 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[18:19], v[14:17] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xe0 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v20, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xf0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v16, 24, v13 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v15, v13, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v14, v13, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v13, 0, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[17:18], v[13:16] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xc0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(6) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v15, 24, v10 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v14, v10, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v10, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v12, v10, 0, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[19:20], v[12:15] | |
| ; GCN-HSA-NEXT: v_bfe_i32 v10, v8, 16, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v14, 24, v11 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v11, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v12, v11, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v11, v11, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[11:14] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xd0 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xa0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v15, 24, v9 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v14, v9, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v9, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v12, v9, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v11, 24, v8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v9, v8, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v8, v8, 0, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[8:11] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xb0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[18:19], v[12:15] | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(8) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v15, 24, v6 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v14, v6, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v6, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v12, v6, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[12:15] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v6, s2 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v15, 24, v7 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v14, v7, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v7, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v12, v7, 0, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v7, s3 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x80 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[6:7], v[12:15] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v7, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v6, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x90 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v15, 24, v4 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v14, v4, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v4, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v12, v4, 0, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v11, 24, v1 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v10, v1, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v9, v1, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v8, v1, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[6:7], v[12:15] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v15, 24, v0 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v14, v0, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v12, v0, 0, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x60 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v19, 24, v5 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v18, v5, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v17, v5, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v16, v5, 0, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[16:19] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x70 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v19, 24, v2 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v18, v2, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v17, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v16, v2, 0, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[16:19] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 64 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v6, 24, v3 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v3, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v3, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v3, v3, 0, 8 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 0x50 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[3:6] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[12:15] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[8:11] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v64i8_to_v64i32: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 6, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[12:15], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v2, vcc, 48, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[8:11], v[2:3] | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v2, vcc, 32, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[4:7], v[2:3] | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, 16, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(3) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v16, 8, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v17, v16, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v19, 24, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v18, v14, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v14, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v14, 8, v15 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(2) | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v21, 24, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v14, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v18, 24, v15 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v17, v15, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v15, v15, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v14, 8, v12 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[15:18], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v20, v6, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v15, v14, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v17, 24, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v12, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v12, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[14:17], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v12, v13, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v15, 24, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v13, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v13, 8, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v16, 8, v10 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v13, v13, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v17, 8, v11 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v18, 8, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v15, 24, v10 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v10, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v12, v10, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v13, v16, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:224 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v9, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v13, 24, v9 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v12, v9, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v15, v17, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v17, 24, v11 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v11, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v11, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v9, 8, v9 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[14:17], off, s[0:3], 0 offset:240 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v11, v9, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v17, 24, v8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v8, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v8, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v8, 8, v8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v19, v18, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v18, v6, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v6, 8, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v9, 8, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v15, v8, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[10:13], off, s[0:3], 0 offset:208 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v8, v4, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 24, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v4, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[18:21], off, s[0:3], 0 offset:160 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v12, 8, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v21, 24, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v20, v7, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v18, v7, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v19, v6, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v9, v9, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[14:17], off, s[0:3], 0 offset:192 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[18:21], off, s[0:3], 0 offset:176 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v17, 24, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v5, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v5, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(10) | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v7, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v6, v1, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v4, v1, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:128 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v21, 24, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v0, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v8, v0, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v20, v3, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v18, v3, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v25, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v24, v2, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v22, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v2, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v3, 8, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v0, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v1, 8, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v15, v12, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v5, v1, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v9, v0, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v19, v3, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v23, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[14:17], off, s[0:3], 0 offset:144 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[22:25], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[18:21], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v64i8_to_v64i32: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @32, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @24 | |
| ; EG-NEXT: ALU 16, @35, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 2 @26 | |
| ; EG-NEXT: ALU 76, @52, KC0[], KC1[] | |
| ; EG-NEXT: ALU 71, @129, KC0[], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T48.XYZW, T50.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T47.XYZW, T49.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T46.XYZW, T0.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T35.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T45.XYZW, T34.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T44.XYZW, T33.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T43.XYZW, T32.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T30.XYZW, T31.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T42.XYZW, T28.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T41.XYZW, T27.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T40.XYZW, T26.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T29.XYZW, T25.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T39.XYZW, T24.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T38.XYZW, T23.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T37.XYZW, T21.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T36.XYZW, T22.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: PAD | |
| ; EG-NEXT: Fetch clause starting at 24: | |
| ; EG-NEXT: VTX_READ_128 T0.XYZW, T20.X, 48, #1 | |
| ; EG-NEXT: Fetch clause starting at 26: | |
| ; EG-NEXT: VTX_READ_128 T29.XYZW, T20.X, 0, #1 | |
| ; EG-NEXT: VTX_READ_128 T30.XYZW, T20.X, 16, #1 | |
| ; EG-NEXT: VTX_READ_128 T20.XYZW, T20.X, 32, #1 | |
| ; EG-NEXT: ALU clause starting at 32: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 6(8.407791e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T20.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 35: | |
| ; EG-NEXT: LSHR * T21.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T22.X, PV.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T23.X, PV.X, literal.y, | |
| ; EG-NEXT: 4(5.605194e-45), 12(1.681558e-44) | |
| ; EG-NEXT: ADD_INT T24.X, T21.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T25.X, T21.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 20(2.802597e-44) | |
| ; EG-NEXT: ADD_INT T26.X, T21.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T27.X, T21.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 28(3.923636e-44) | |
| ; EG-NEXT: ADD_INT T28.X, T21.X, literal.x, | |
| ; EG-NEXT: LSHR T1.Y, T0.Z, literal.y, | |
| ; EG-NEXT: LSHR T1.Z, T0.Z, literal.x, | |
| ; EG-NEXT: LSHR T1.W, T0.W, literal.y, | |
| ; EG-NEXT: LSHR * T2.W, T0.W, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; EG-NEXT: ALU clause starting at 52: | |
| ; EG-NEXT: ADD_INT T31.X, T21.X, literal.x, | |
| ; EG-NEXT: LSHR T2.Y, T0.X, literal.y, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR T2.Z, T0.X, literal.z, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR T3.W, T0.Y, literal.y, | |
| ; EG-NEXT: LSHR * T4.W, T0.Y, literal.z, | |
| ; EG-NEXT: 36(5.044674e-44), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T32.X, T21.X, literal.x, | |
| ; EG-NEXT: LSHR T3.Y, T20.Z, literal.y, | |
| ; EG-NEXT: LSHR T3.Z, T20.Z, literal.z, | |
| ; EG-NEXT: LSHR T5.W, T20.W, literal.y, | |
| ; EG-NEXT: LSHR * T6.W, T20.W, literal.z, | |
| ; EG-NEXT: 32(4.484155e-44), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T33.X, T21.X, literal.x, | |
| ; EG-NEXT: LSHR T4.Y, T20.X, literal.y, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR T4.Z, T20.X, literal.z, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR T7.W, T20.Y, literal.y, | |
| ; EG-NEXT: LSHR * T8.W, T20.Y, literal.z, | |
| ; EG-NEXT: 44(6.165713e-44), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T34.X, T21.X, literal.x, | |
| ; EG-NEXT: LSHR T5.Y, T30.Z, literal.y, | |
| ; EG-NEXT: LSHR T5.Z, T30.Z, literal.z, | |
| ; EG-NEXT: LSHR T9.W, T30.W, literal.y, | |
| ; EG-NEXT: LSHR * T10.W, T30.W, literal.z, | |
| ; EG-NEXT: 40(5.605194e-44), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T35.X, T21.X, literal.x, | |
| ; EG-NEXT: LSHR T6.Y, T30.X, literal.y, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR T6.Z, T30.X, literal.z, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR T11.W, T30.Y, literal.y, | |
| ; EG-NEXT: LSHR * T12.W, T30.Y, literal.z, | |
| ; EG-NEXT: 52(7.286752e-44), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T36.X, T29.Y, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T7.Y, T29.Z, literal.y, | |
| ; EG-NEXT: LSHR T7.Z, T29.Z, literal.z, | |
| ; EG-NEXT: LSHR T13.W, T29.W, literal.y, | |
| ; EG-NEXT: LSHR * T14.W, T29.Y, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T37.X, T29.X, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T8.Y, T29.W, literal.y, | |
| ; EG-NEXT: LSHR T8.Z, T29.X, literal.y, | |
| ; EG-NEXT: BFE_INT T36.W, PS, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T14.W, T29.Y, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T38.X, T29.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T9.Y, T29.X, literal.y, | |
| ; EG-NEXT: BFE_INT T36.Z, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T37.W, PV.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T14.W, T29.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT T39.X, T29.Z, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T36.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T37.Z, PV.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T38.W, T8.Y, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T14.W, T29.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T29.X, T30.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T37.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T38.Z, T13.W, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T39.W, T7.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T13.W, T29.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T40.X, T30.X, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T38.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T39.Z, T7.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T29.W, T12.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T12.W, T29.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T41.X, T30.W, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T39.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT * T29.Z, T11.W, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ALU clause starting at 129: | |
| ; EG-NEXT: BFE_INT T40.W, T6.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T11.W, T30.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T42.X, T30.Z, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T29.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T40.Z, T6.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T41.W, T10.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T10.W, T30.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T30.X, T20.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T40.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T41.Z, T9.W, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T42.W, T5.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T9.W, T30.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T43.X, T20.X, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T41.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T42.Z, T5.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T30.W, T8.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T8.W, T30.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T44.X, T20.W, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T42.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T30.Z, T7.W, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: BFE_INT T43.W, T4.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T7.W, T20.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T45.X, T20.Z, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T30.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T43.Z, T4.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T44.W, T6.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T6.W, T20.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T20.X, T0.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T43.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T44.Z, T5.W, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T45.W, T3.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T5.W, T20.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T46.X, T0.X, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T44.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T45.Z, T3.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T20.W, T4.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T4.W, T20.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T47.X, T0.W, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T45.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T20.Z, T3.W, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: BFE_INT T46.W, T2.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T3.W, T0.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T48.X, T0.Z, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T20.Y, PS, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T46.Z, T2.Y, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT T47.W, T2.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T2.W, T0.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T0.X, T21.X, literal.x, | |
| ; EG-NEXT: BFE_INT T46.Y, PS, 0.0, literal.y, | |
| ; EG-NEXT: BFE_INT T47.Z, T1.W, 0.0, literal.y, | |
| ; EG-NEXT: BFE_INT T48.W, T1.Z, 0.0, literal.y, | |
| ; EG-NEXT: LSHR * T0.W, T0.W, literal.y, | |
| ; EG-NEXT: 48(6.726233e-44), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T49.X, T21.X, literal.x, | |
| ; EG-NEXT: BFE_INT T47.Y, PS, 0.0, literal.y, | |
| ; EG-NEXT: BFE_INT T48.Z, T1.Y, 0.0, literal.y, | |
| ; EG-NEXT: LSHR T0.W, T0.Z, literal.y, | |
| ; EG-NEXT: ADD_INT * T50.X, T21.X, literal.z, | |
| ; EG-NEXT: 60(8.407791e-44), 8(1.121039e-44) | |
| ; EG-NEXT: 56(7.847271e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT * T48.Y, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_sextload_v64i8_to_v64i32: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @32, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @24 | |
| ; CM-NEXT: ALU 17, @35, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 2 @26 | |
| ; CM-NEXT: ALU 84, @53, KC0[], KC1[] | |
| ; CM-NEXT: ALU 72, @138, KC0[], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T50.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T47, T21.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T46, T49.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T45, T48.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T44, T35.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T20, T34.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T43, T33.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T42, T32.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T41, T31.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T25, T30.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T40, T29.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T39, T28.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T38, T27.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T24, T26.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T37, T23.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T36, T22.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 24: | |
| ; CM-NEXT: VTX_READ_128 T0.XYZW, T20.X, 0, #1 | |
| ; CM-NEXT: Fetch clause starting at 26: | |
| ; CM-NEXT: VTX_READ_128 T24.XYZW, T20.X, 48, #1 | |
| ; CM-NEXT: VTX_READ_128 T25.XYZW, T20.X, 32, #1 | |
| ; CM-NEXT: VTX_READ_128 T20.XYZW, T20.X, 16, #1 | |
| ; CM-NEXT: ALU clause starting at 32: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 6(8.407791e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T20.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 35: | |
| ; CM-NEXT: LSHR T21.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: LSHR T1.Y, T0.Y, literal.y, | |
| ; CM-NEXT: LSHR T1.Z, T0.X, literal.z, | |
| ; CM-NEXT: LSHR * T1.W, T0.Y, literal.w, | |
| ; CM-NEXT: 2(2.802597e-45), 16(2.242078e-44) | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: ADD_INT T22.X, PV.X, literal.x, | |
| ; CM-NEXT: LSHR T2.Y, T0.X, literal.y, | |
| ; CM-NEXT: LSHR T2.Z, T0.W, literal.z, | |
| ; CM-NEXT: LSHR * T2.W, T0.X, literal.w, | |
| ; CM-NEXT: 56(7.847271e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: ADD_INT T23.X, T21.X, literal.x, | |
| ; CM-NEXT: LSHR T3.Y, T0.W, literal.y, | |
| ; CM-NEXT: LSHR T3.Z, T0.Z, literal.z, | |
| ; CM-NEXT: LSHR * T3.W, T0.W, literal.w, | |
| ; CM-NEXT: 60(8.407791e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: ALU clause starting at 53: | |
| ; CM-NEXT: LSHR T4.Z, T0.Z, literal.x, | |
| ; CM-NEXT: LSHR * T4.W, T20.Y, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T26.X, T21.X, literal.x, | |
| ; CM-NEXT: LSHR T4.Y, T0.Z, literal.y, | |
| ; CM-NEXT: LSHR T5.Z, T20.Y, literal.z, | |
| ; CM-NEXT: LSHR * T5.W, T20.X, literal.w, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 48(6.726233e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T27.X, T21.X, literal.x, | |
| ; CM-NEXT: LSHR T5.Y, T20.Y, literal.y, | |
| ; CM-NEXT: LSHR T6.Z, T20.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR * T6.W, T20.W, literal.w, | |
| ; CM-NEXT: 52(7.286752e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T28.X, T21.X, literal.x, | |
| ; CM-NEXT: LSHR T6.Y, T20.X, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR T7.Z, T20.W, literal.z, | |
| ; CM-NEXT: LSHR * T7.W, T20.Z, literal.w, | |
| ; CM-NEXT: 40(5.605194e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T29.X, T21.X, literal.x, | |
| ; CM-NEXT: LSHR T7.Y, T20.W, literal.y, | |
| ; CM-NEXT: LSHR T8.Z, T20.Z, literal.z, | |
| ; CM-NEXT: LSHR * T8.W, T25.Y, literal.w, | |
| ; CM-NEXT: 44(6.165713e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T30.X, T21.X, literal.x, | |
| ; CM-NEXT: LSHR T8.Y, T20.Z, literal.y, | |
| ; CM-NEXT: LSHR T9.Z, T25.Y, literal.z, | |
| ; CM-NEXT: LSHR * T9.W, T25.X, literal.w, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 32(4.484155e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T31.X, T21.X, literal.x, | |
| ; CM-NEXT: LSHR T9.Y, T25.Y, literal.y, | |
| ; CM-NEXT: LSHR T10.Z, T25.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR * T10.W, T25.W, literal.w, | |
| ; CM-NEXT: 36(5.044674e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T32.X, T21.X, literal.x, | |
| ; CM-NEXT: LSHR T10.Y, T25.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR T11.Z, T25.W, literal.y, | |
| ; CM-NEXT: LSHR * T11.W, T25.Z, literal.z, | |
| ; CM-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T33.X, T21.X, literal.x, | |
| ; CM-NEXT: LSHR T11.Y, T25.W, literal.y, | |
| ; CM-NEXT: LSHR T12.Z, T25.Z, literal.z, | |
| ; CM-NEXT: LSHR * T12.W, T24.Y, literal.w, | |
| ; CM-NEXT: 28(3.923636e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T34.X, T21.X, literal.x, | |
| ; CM-NEXT: LSHR T12.Y, T25.Z, literal.y, | |
| ; CM-NEXT: LSHR T13.Z, T24.Y, literal.x, | |
| ; CM-NEXT: LSHR * T13.W, T24.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T35.X, T21.X, literal.x, | |
| ; CM-NEXT: LSHR T13.Y, T24.Y, literal.y, | |
| ; CM-NEXT: LSHR T14.Z, T24.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR * T14.W, T24.W, literal.w, | |
| ; CM-NEXT: 20(2.802597e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: BFE_INT T36.X, T24.Z, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T14.Y, T24.X, literal.y, | |
| ; CM-NEXT: LSHR T15.Z, T24.W, literal.z, | |
| ; CM-NEXT: LSHR * T15.W, T24.Z, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T37.X, T24.W, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T15.Y, T24.W, literal.y, | |
| ; CM-NEXT: LSHR T16.Z, T24.Z, literal.z, | |
| ; CM-NEXT: BFE_INT * T36.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T24.X, T24.X, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T16.Y, T24.Z, literal.x, | |
| ; CM-NEXT: BFE_INT T36.Z, PV.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T37.W, PV.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T38.X, T24.Y, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T36.Y, PV.Y, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T37.Z, T15.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T24.W, T14.Y, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ALU clause starting at 138: | |
| ; CM-NEXT: BFE_INT T39.X, T25.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T37.Y, T14.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T24.Z, T14.Z, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: BFE_INT * T38.W, T13.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T40.X, T25.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T24.Y, T13.W, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: BFE_INT T38.Z, T13.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T39.W, T12.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T25.X, T25.X, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T38.Y, T12.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T39.Z, T12.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T40.W, T11.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T41.X, T25.Y, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T39.Y, T11.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T40.Z, T11.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T25.W, T10.Y, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T42.X, T20.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T40.Y, T10.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T25.Z, T10.Z, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: BFE_INT * T41.W, T9.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T43.X, T20.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T25.Y, T9.W, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: BFE_INT T41.Z, T9.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T42.W, T8.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T20.X, T20.X, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T41.Y, T8.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T42.Z, T8.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T43.W, T7.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T44.X, T20.Y, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T42.Y, T7.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T43.Z, T7.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T20.W, T6.Y, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T45.X, T0.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T43.Y, T6.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T20.Z, T6.Z, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: BFE_INT * T44.W, T5.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T46.X, T0.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T20.Y, T5.W, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: BFE_INT T44.Z, T5.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T45.W, T4.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T47.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T44.Y, T4.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T45.Z, T4.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T46.W, T3.W, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T0.X, T0.Y, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T45.Y, T3.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T46.Z, T3.Y, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: BFE_INT * T47.W, T2.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T48.X, T21.X, literal.x, | |
| ; CM-NEXT: BFE_INT T46.Y, T2.Z, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT T47.Z, T2.Y, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T0.W, T1.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T49.X, T21.X, literal.x, | |
| ; CM-NEXT: BFE_INT T47.Y, T1.Z, 0.0, literal.y, | |
| ; CM-NEXT: BFE_INT T0.Z, T1.Y, 0.0, literal.y, | |
| ; CM-NEXT: LSHR * T1.W, T0.Y, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 12(1.681558e-44), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T50.X, T21.X, literal.x, | |
| ; CM-NEXT: BFE_INT * T0.Y, PV.W, 0.0, literal.y, | |
| ; CM-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <64 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <64 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <64 x i8> %load to <64 x i32> | |
| store <64 x i32> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_i8_to_i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_i8_to_i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_i8_to_i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ubyte v0, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_dwordx2 v[2:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_i8_to_i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v0, 0xffff, v0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_i8_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) | |
| ; | |
| ; CM-LABEL: global_zextload_i8_to_i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 2, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: MOV * T0.Y, 0.0, | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %a = load i8, ptr addrspace(1) %in | |
| %ext = zext i8 %a to i64 | |
| store i64 %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG Why 31 not 7 ? | |
| define amdgpu_kernel void @global_sextload_i8_to_i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_i8_to_i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_i8_to_i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_sbyte v0, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx2 v[2:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_i8_to_i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v0, v0, 0, 16 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_i8_to_i64: | |
| ; 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 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, literal.x, | |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 2(2.802597e-45) | |
| ; EG-NEXT: ASHR * T0.Y, PV.X, literal.x, | |
| ; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_sextload_i8_to_i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 4, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: BFE_INT * T0.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: ASHR * T0.Y, PV.X, literal.y, | |
| ; CM-NEXT: 2(2.802597e-45), 31(4.344025e-44) | |
| %a = load i8, ptr addrspace(1) %in | |
| %ext = sext i8 %a to i64 | |
| store i64 %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v1i8_to_v1i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v1i8_to_v1i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v1i8_to_v1i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ubyte v0, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_dwordx2 v[2:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v1i8_to_v1i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v1i8_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) | |
| ; | |
| ; CM-LABEL: global_zextload_v1i8_to_v1i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 2, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: MOV * T0.Y, 0.0, | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %load = load <1 x i8>, ptr addrspace(1) %in | |
| %ext = zext <1 x i8> %load to <1 x i64> | |
| store <1 x i64> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; TODO: EG Why 31 not 7 ? | |
| define amdgpu_kernel void @global_sextload_v1i8_to_v1i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v1i8_to_v1i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v1i8_to_v1i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_sbyte v0, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx2 v[2:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v1i8_to_v1i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v0, v0, 0, 16 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v1i8_to_v1i64: | |
| ; 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 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, literal.x, | |
| ; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 2(2.802597e-45) | |
| ; EG-NEXT: ASHR * T0.Y, PV.X, literal.x, | |
| ; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_sextload_v1i8_to_v1i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 4, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: BFE_INT * T0.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: ASHR * T0.Y, PV.X, literal.y, | |
| ; CM-NEXT: 2(2.802597e-45), 31(4.344025e-44) | |
| %load = load <1 x i8>, ptr addrspace(1) %in | |
| %ext = sext <1 x i8> %load to <1 x i64> | |
| store <1 x i64> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v2i8_to_v2i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v2i8_to_v2i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, v1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v2, 8, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xff, v0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v2i8_to_v2i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ushort v0, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, v1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v2, 8, v0 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v0, 0xff, v0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v2i8_to_v2i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, v1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_sdwa v2, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v0, 0xff, v0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v2i8_to_v2i64: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 1, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 14, @10, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T4.XYZW, T5.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: PAD | |
| ; EG-NEXT: Fetch clause starting at 6: | |
| ; EG-NEXT: VTX_READ_16 T4.X, T4.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 8: | |
| ; EG-NEXT: MOV * T0.Y, T2.X, | |
| ; EG-NEXT: MOV * T4.X, KC0[2].Z, | |
| ; EG-NEXT: ALU clause starting at 10: | |
| ; EG-NEXT: AND_INT T0.W, T4.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.Y, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T2.X, PV.W, | |
| ; EG-NEXT: MOV T0.Y, PV.X, | |
| ; EG-NEXT: MOV * T1.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT * T4.Z, PV.Y, literal.x, PV.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T4.X, T0.W, literal.x, | |
| ; EG-NEXT: MOV T4.Y, 0.0, | |
| ; EG-NEXT: MOV T4.W, 0.0, | |
| ; EG-NEXT: LSHR * T5.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 255(3.573311e-43), 2(2.802597e-45) | |
| ; | |
| ; CM-LABEL: global_zextload_v2i8_to_v2i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 1, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 15, @10, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T4, T5.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_16 T4.X, T4.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.Y, T2.X, | |
| ; CM-NEXT: MOV * T4.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 10: | |
| ; CM-NEXT: AND_INT T0.Z, T4.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; CM-NEXT: OR_INT * T0.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T2.X, PV.W, | |
| ; CM-NEXT: MOV T0.Y, PV.X, | |
| ; CM-NEXT: MOV * T1.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT * T4.Z, PV.Y, literal.x, PV.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T4.X, T0.W, literal.x, | |
| ; CM-NEXT: MOV T4.Y, 0.0, | |
| ; CM-NEXT: MOV * T4.W, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR * T5.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %load = load <2 x i8>, ptr addrspace(1) %in | |
| %ext = zext <2 x i8> %load to <2 x i64> | |
| store <2 x i64> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v2i8_to_v2i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v2i8_to_v2i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s7 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v2, 8, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v2, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v2i8_to_v2i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ushort v0, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v2, 8, v0 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v2, v2, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v2i8_to_v2i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s7 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v2, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v2, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v2i8_to_v2i64: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 1, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 15, @10, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T4.XYZW, T5.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: PAD | |
| ; EG-NEXT: Fetch clause starting at 6: | |
| ; EG-NEXT: VTX_READ_16 T4.X, T4.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 8: | |
| ; EG-NEXT: MOV * T0.Y, T2.X, | |
| ; EG-NEXT: MOV * T4.X, KC0[2].Z, | |
| ; EG-NEXT: ALU clause starting at 10: | |
| ; EG-NEXT: AND_INT T0.W, T4.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.Y, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T2.X, PV.W, | |
| ; EG-NEXT: MOV * T0.Y, PV.X, | |
| ; EG-NEXT: BFE_INT * T4.X, T0.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ASHR T4.Y, PV.X, literal.x, | |
| ; EG-NEXT: LSHR * T0.W, T0.Y, literal.y, | |
| ; EG-NEXT: 31(4.344025e-44), 8(1.121039e-44) | |
| ; EG-NEXT: BFE_INT * T4.Z, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: LSHR T5.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: ASHR * T4.W, PV.Z, literal.y, | |
| ; EG-NEXT: 2(2.802597e-45), 31(4.344025e-44) | |
| ; | |
| ; CM-LABEL: global_sextload_v2i8_to_v2i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 1, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 15, @10, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T4, T5.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_16 T4.X, T4.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.Y, T2.X, | |
| ; CM-NEXT: MOV * T4.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 10: | |
| ; CM-NEXT: AND_INT T0.Z, T4.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; CM-NEXT: OR_INT * T0.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T2.X, PV.W, | |
| ; CM-NEXT: MOV * T0.Y, PV.X, | |
| ; CM-NEXT: BFE_INT * T4.X, T0.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ASHR T4.Y, PV.X, literal.x, | |
| ; CM-NEXT: LSHR * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 31(4.344025e-44), 8(1.121039e-44) | |
| ; CM-NEXT: BFE_INT * T4.Z, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR T5.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: ASHR * T4.W, PV.Z, literal.y, | |
| ; CM-NEXT: 2(2.802597e-45), 31(4.344025e-44) | |
| %load = load <2 x i8>, ptr addrspace(1) %in | |
| %ext = sext <2 x i8> %load to <2 x i64> | |
| store <2 x i64> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v4i8_to_v4i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v4i8_to_v4i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v0, v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, v1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v5, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v7, v1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v6, v0, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v2, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v4, 0xff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v0, v0, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v4i8_to_v4i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dword v0, v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v11, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v10, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v7, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, s0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_bfe_u32 v6, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v2, 24, v0 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v4, 0xff, v0 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v0, v0, 16, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[10:11], v[0:3] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[8:9], v[4:7] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v4i8_to_v4i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dword v0, v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v7, v1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v2, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v4, 0xff, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v6, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v0, v0, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v4i8_to_v4i64: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 16, @11, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.XYZW, T7.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T6.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; 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: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 11: | |
| ; EG-NEXT: MOV * T0.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T5.X, T0.X, literal.x, PV.W, | |
| ; EG-NEXT: LSHR * T5.Z, T0.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T5.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T0.Z, T0.X, literal.x, T0.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T0.X, T0.X, literal.x, | |
| ; EG-NEXT: MOV T0.Y, 0.0, | |
| ; EG-NEXT: MOV T5.W, 0.0, | |
| ; EG-NEXT: MOV * T0.W, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: LSHR * T6.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T7.X, PV.X, literal.x, | |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_zextload_v4i8_to_v4i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 16, @11, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T5, T0.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T6, T7.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 11: | |
| ; CM-NEXT: MOV * T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT * T5.Z, T0.X, literal.x, PV.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T5.X, T0.X, literal.x, | |
| ; CM-NEXT: MOV * T5.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T6.X, T0.X, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T6.Z, T0.X, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T6.Y, 0.0, | |
| ; CM-NEXT: MOV * T5.W, 0.0, | |
| ; CM-NEXT: MOV * T6.W, 0.0, | |
| ; CM-NEXT: LSHR * T0.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T7.X, PV.X, literal.x, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <4 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <4 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <4 x i8> %load to <4 x i64> | |
| store <4 x i64> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v4i8_to_v4i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v4i8_to_v4i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v0, v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v3, 16, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v4, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v2, 8, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v6, v4, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v4, v3, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v2, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v4i8_to_v4i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dword v0, v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v11, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v10, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, s0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v3, 16, v0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v4, 24, v0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v2, 8, v0 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v4, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v3, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v2, v2, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[10:11], v[4:7] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[8:9], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v4i8_to_v4i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dword v0, v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 16, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v4, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v2, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v6, v4, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v4, v3, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v2, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v4i8_to_v4i64: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 18, @11, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.XYZW, T7.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T6.XYZW, T0.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; 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: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 11: | |
| ; EG-NEXT: ASHR * T5.W, T0.X, literal.x, | |
| ; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T6.X, T0.X, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T5.Z, T0.X, literal.y, | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T5.X, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T6.Y, PV.X, literal.y, | |
| ; EG-NEXT: LSHR T0.W, T0.X, literal.x, | |
| ; EG-NEXT: LSHR * T0.X, KC0[2].Y, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT * T6.Z, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T7.X, T0.X, literal.x, | |
| ; EG-NEXT: ASHR T5.Y, T5.X, literal.y, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: ASHR * T6.W, PV.Z, literal.y, | |
| ; EG-NEXT: 4(5.605194e-45), 31(4.344025e-44) | |
| ; | |
| ; CM-LABEL: global_sextload_v4i8_to_v4i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 17, @11, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T6, T0.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T5, T7.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 11: | |
| ; CM-NEXT: ASHR * T5.W, T0.X, literal.x, | |
| ; CM-NEXT: 31(4.344025e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T6.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T5.Z, T0.X, literal.y, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.z, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T5.X, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T6.Y, PV.X, literal.y, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: LSHR T0.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: BFE_INT * T6.Z, PV.W, 0.0, literal.y, | |
| ; CM-NEXT: 2(2.802597e-45), 8(1.121039e-44) | |
| ; CM-NEXT: ADD_INT T7.X, PV.X, literal.x, | |
| ; CM-NEXT: ASHR T5.Y, T5.X, literal.y, | |
| ; CM-NEXT: ASHR * T6.W, PV.Z, literal.y, | |
| ; CM-NEXT: 4(5.605194e-45), 31(4.344025e-44) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <4 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <4 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <4 x i8> %load to <4 x i64> | |
| store <4 x i64> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v8i8_to_v8i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v8i8_to_v8i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx2 v[16:17], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, v1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v5, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v7, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v9, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v11, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v13, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v15, v1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v2, 24, v17 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v0, v17, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v6, 24, v16 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v10, v17, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v14, v16, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v12, 0xff, v16 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v8, 0xff, v17 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v4, v16, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v8i8_to_v8i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx2 v[15:16], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 48 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, v1 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v6, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v10, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, v1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v2, 24, v16 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v0, v16, 16, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[17:18], v[0:3] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 32 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s2 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v5, 24, v15 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v3, v15, 16, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v20, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s0 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v9, v16, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v13, v15, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v11, 0xff, v15 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v7, 0xff, v16 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[17:18], v[3:6] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[19:20], v[7:10] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[11:14] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v8i8_to_v8i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx2 v[16:17], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v7, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v9, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v11, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v13, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v15, v1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v2, 24, v17 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v0, v17, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v6, 24, v16 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v12, 0xff, v17 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v14, 8, v17 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v8, 0xff, v16 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v10, 8, v16 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v4, v16, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v8i8_to_v8i64: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @10, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @8 | |
| ; EG-NEXT: ALU 30, @13, 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 T0.XYZW, T9.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: Fetch clause starting at 8: | |
| ; EG-NEXT: VTX_READ_64 T0.XY, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 10: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 13: | |
| ; EG-NEXT: MOV * T0.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T6.X, T0.Y, literal.x, PV.W, | |
| ; EG-NEXT: LSHR * T6.Z, T0.Y, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T6.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T7.Z, T0.Y, literal.x, T0.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T7.X, T0.Y, literal.x, | |
| ; EG-NEXT: MOV * T7.Y, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T8.X, T0.X, literal.x, T0.W, | |
| ; EG-NEXT: LSHR * T8.Z, T0.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T8.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T0.Z, T0.X, literal.x, T0.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T0.X, T0.X, literal.x, | |
| ; EG-NEXT: MOV T0.Y, 0.0, | |
| ; EG-NEXT: MOV T6.W, 0.0, | |
| ; EG-NEXT: MOV * T7.W, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: MOV T8.W, 0.0, | |
| ; EG-NEXT: MOV * T0.W, 0.0, | |
| ; EG-NEXT: LSHR * T9.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T10.X, PV.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T11.X, PV.X, literal.y, | |
| ; EG-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT * T12.X, T9.X, literal.x, | |
| ; EG-NEXT: 12(1.681558e-44), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_zextload_v8i8_to_v8i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @10, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @8 | |
| ; CM-NEXT: ALU 31, @13, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T6, T9.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T7, T12.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T8, T11.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T10.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 8: | |
| ; CM-NEXT: VTX_READ_64 T0.XY, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 10: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 13: | |
| ; CM-NEXT: MOV * T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT * T6.Z, T0.X, literal.x, PV.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T6.X, T0.X, literal.x, | |
| ; CM-NEXT: MOV * T6.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T7.X, T0.X, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T7.Z, T0.X, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T7.Y, 0.0, | |
| ; CM-NEXT: BFE_UINT * T8.Z, T0.Y, literal.x, T0.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T8.X, T0.Y, literal.x, | |
| ; CM-NEXT: MOV * T8.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T0.X, T0.Y, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T0.Z, T0.Y, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T0.Y, 0.0, | |
| ; CM-NEXT: MOV * T6.W, 0.0, | |
| ; CM-NEXT: MOV * T7.W, 0.0, | |
| ; CM-NEXT: MOV * T8.W, 0.0, | |
| ; CM-NEXT: MOV * T0.W, 0.0, | |
| ; CM-NEXT: LSHR * T9.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T10.X, PV.X, literal.x, | |
| ; CM-NEXT: 12(1.681558e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T11.X, T9.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T12.X, T9.X, literal.x, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <8 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <8 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <8 x i8> %load to <8 x i64> | |
| store <8 x i64> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v8i8_to_v8i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v8i8_to_v8i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v3, 16, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v6, 8, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v5, 16, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v7, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v2, 8, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v8, v3, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v4, v1, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v11, 31, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v10, 24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v2, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v14, v7, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v12, v5, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v6, v6, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v15, 31, v14 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v13, 31, v12 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v8i8_to_v8i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx2 v[4:5], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 48 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v7, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v6, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 32 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s0 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 16 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s2 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v0, 16, v5 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v9, 8, v5 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v3, 31, v5 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v2, 24, v5 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v8, 16, v4 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v10, 24, v4 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v11, 8, v4 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[6:7], v[0:3] | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v9, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v0, v4, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v5, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v2, v11, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v10, v10, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v8, v8, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v11, 31, v10 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[4:7] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[8:11] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[12:13], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v8i8_to_v8i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx2 v[1:2], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 16, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v8, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v7, 16, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v6, 8, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v5, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v8, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v8, v3, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v0, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v4, v1, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 31, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v10, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v2, v5, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v6, v6, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v12, v7, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v15, 31, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v13, 31, v12 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v8i8_to_v8i64: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @10, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @8 | |
| ; EG-NEXT: ALU 34, @13, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T12.XYZW, T0.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T6.XYZW, T9.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T10.XYZW, T8.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T11.XYZW, T7.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: Fetch clause starting at 8: | |
| ; EG-NEXT: VTX_READ_64 T0.XY, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 10: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 13: | |
| ; EG-NEXT: BFE_INT T6.X, T0.Y, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T7.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 2(2.802597e-45) | |
| ; EG-NEXT: ADD_INT T8.X, PS, literal.x, | |
| ; EG-NEXT: ASHR T6.Y, PV.X, literal.y, | |
| ; EG-NEXT: LSHR T0.W, T0.Y, literal.z, | |
| ; EG-NEXT: ADD_INT * T9.X, PS, literal.z, | |
| ; EG-NEXT: 4(5.605194e-45), 31(4.344025e-44) | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T6.Z, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR * T10.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: BFE_INT T11.X, T0.X, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T10.Z, T0.X, literal.y, | |
| ; EG-NEXT: LSHR T0.W, T0.X, literal.z, | |
| ; EG-NEXT: ASHR * T12.W, T0.Y, literal.w, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44) | |
| ; EG-NEXT: BFE_INT T10.X, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T11.Y, PV.X, literal.y, | |
| ; EG-NEXT: ASHR T12.Z, T0.Y, literal.z, | |
| ; EG-NEXT: LSHR T0.W, T0.X, literal.x, | |
| ; EG-NEXT: LSHR * T1.W, T0.Y, literal.w, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT T12.X, PS, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T10.Y, PV.X, literal.y, | |
| ; EG-NEXT: BFE_INT T11.Z, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: ADD_INT * T0.X, T7.X, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 12(1.681558e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ASHR T12.Y, PV.X, literal.x, | |
| ; EG-NEXT: ASHR T11.W, PV.Z, literal.x, | |
| ; EG-NEXT: ASHR * T6.W, T6.Z, literal.x, | |
| ; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_sextload_v8i8_to_v8i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @10, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @8 | |
| ; CM-NEXT: ALU 34, @13, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T11, T7.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T12.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T6, T9.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T10, T8.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 8: | |
| ; CM-NEXT: VTX_READ_64 T0.XY, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 10: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 13: | |
| ; CM-NEXT: BFE_INT * T6.X, T0.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR * T7.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T8.X, PV.X, literal.x, | |
| ; CM-NEXT: ASHR T6.Y, T6.X, literal.y, | |
| ; CM-NEXT: LSHR * T0.W, T0.Y, literal.z, | |
| ; CM-NEXT: 12(1.681558e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T9.X, T7.X, literal.x, | |
| ; CM-NEXT: LSHR T1.Y, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: BFE_INT T6.Z, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: ASHR * T10.W, T0.Y, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: BFE_INT T11.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T2.Y, T0.Y, literal.y, | |
| ; CM-NEXT: ASHR T10.Z, T0.Y, literal.z, | |
| ; CM-NEXT: ASHR * T0.W, T0.X, literal.w, | |
| ; CM-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 31(4.344025e-44) | |
| ; CM-NEXT: BFE_INT T10.X, PV.Y, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T11.Y, PV.X, literal.y, | |
| ; CM-NEXT: ASHR T0.Z, T0.X, literal.z, | |
| ; CM-NEXT: LSHR * T1.W, T0.X, literal.w, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; CM-NEXT: BFE_INT T0.X, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T10.Y, PV.X, literal.y, | |
| ; CM-NEXT: BFE_INT T11.Z, T1.Y, 0.0, literal.x, | |
| ; CM-NEXT: ASHR * T6.W, T6.Z, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: ADD_INT T12.X, T7.X, literal.x, | |
| ; CM-NEXT: ASHR T0.Y, PV.X, literal.y, | |
| ; CM-NEXT: ASHR * T11.W, PV.Z, literal.y, | |
| ; CM-NEXT: 4(5.605194e-45), 31(4.344025e-44) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <8 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <8 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <8 x i8> %load to <8 x i64> | |
| store <8 x i64> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v16i8_to_v16i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v16i8_to_v16i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[3:6], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v32, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v34, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v28, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v30, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v26, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v20, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v22, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v16, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v18, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v8, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v10, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v12, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v14, v1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v2, 24, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v9, v3, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v7, 0xff, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v29, 24, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v27, v3, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v0, v5, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v13, v4, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v11, 0xff, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v17, v6, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v15, 0xff, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v21, v5, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v19, 0xff, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v25, 24, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v23, v4, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v33, 24, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v31, v6, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[31:34], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[27:30], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[23:26], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[19:22], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[15:18], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[7:10], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[11:14], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v16i8_to_v16i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x50 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 0x70 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 16 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, v5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v11, v5 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, v5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, v5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, v5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v7, v5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, v5 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v10, 24, v3 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v8, v3, 16, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[15:16], v[8:11] | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v13, 24, v0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 48 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v11, v0, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[8:9], v[11:14] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, v5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 64 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v16, 24, v1 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v14, v1, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, v5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v11, v5 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[12:13], v[14:17] | |
| ; GCN-HSA-NEXT: v_bfe_u32 v10, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s4 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v8, 0xff, v2 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[15:16], v[8:11] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x60 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v6, 24, v2 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v4, v2, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, v5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, v5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, v5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v10, v5 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[15:16], v[4:7] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, v5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 32 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v20, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s2 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v13, v3, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v11, 0xff, v3 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v9, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v17, v1, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v15, 0xff, v1 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v7, 0xff, v0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[19:20], v[11:14] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[7:10] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[15:18] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v16i8_to_v16i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v29, 0 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v31, v29 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v25, v29 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v27, v29 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v21, v29 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v23, v29 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v17, v29 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v19, v29 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v13, v29 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v15, v29 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v9, v29 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v11, v29 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, v29 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v7, v29 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v30, 24, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v28, v3, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v4, 0xff, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v6, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v8, 0xff, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v10, 8, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v12, 0xff, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v14, 8, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v16, 0xff, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v18, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v22, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v20, v0, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v26, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v24, v2, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v2, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v0, v1, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, v29 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, v29 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[28:31], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[24:27], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[20:23], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v16i8_to_v16i64: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @14, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @12 | |
| ; EG-NEXT: ALU 58, @17, 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 T0.XYZW, T15.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: Fetch clause starting at 12: | |
| ; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 14: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 17: | |
| ; EG-NEXT: MOV * T1.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T8.X, T0.W, literal.x, PV.W, | |
| ; EG-NEXT: LSHR * T8.Z, T0.W, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T8.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T9.Z, T0.W, literal.x, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T9.X, T0.W, literal.x, | |
| ; EG-NEXT: MOV * T9.Y, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T10.X, T0.Z, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T10.Z, T0.Z, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T10.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T11.Z, T0.Z, literal.x, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T11.X, T0.Z, literal.x, | |
| ; EG-NEXT: MOV * T11.Y, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T12.X, T0.Y, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T12.Z, T0.Y, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T12.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T13.Z, T0.Y, literal.x, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T13.X, T0.Y, literal.x, | |
| ; EG-NEXT: MOV * T13.Y, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T14.X, T0.X, literal.x, T1.W, | |
| ; EG-NEXT: LSHR * T14.Z, T0.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T14.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T0.Z, T0.X, literal.x, T1.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T0.X, T0.X, literal.x, | |
| ; EG-NEXT: MOV T0.Y, 0.0, | |
| ; EG-NEXT: MOV T8.W, 0.0, | |
| ; EG-NEXT: MOV * T9.W, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; 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 * T0.W, 0.0, | |
| ; EG-NEXT: LSHR * T15.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T16.X, PV.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T17.X, PV.X, literal.y, | |
| ; EG-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T18.X, T15.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T19.X, T15.X, literal.y, | |
| ; EG-NEXT: 12(1.681558e-44), 16(2.242078e-44) | |
| ; EG-NEXT: ADD_INT T20.X, T15.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T21.X, T15.X, literal.y, | |
| ; EG-NEXT: 20(2.802597e-44), 24(3.363116e-44) | |
| ; EG-NEXT: ADD_INT * T22.X, T15.X, literal.x, | |
| ; EG-NEXT: 28(3.923636e-44), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_zextload_v16i8_to_v16i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @14, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @12 | |
| ; CM-NEXT: ALU 61, @17, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T8, T15.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T9, T22.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T10, T21.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T11, T20.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T12, T19.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T18.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T13, T17.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T14, T16.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 12: | |
| ; CM-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 14: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 17: | |
| ; CM-NEXT: MOV * T1.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT * T8.Z, T0.X, literal.x, PV.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T8.X, T0.X, literal.x, | |
| ; CM-NEXT: MOV * T8.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T9.X, T0.X, literal.x, T1.W, | |
| ; CM-NEXT: LSHR * T9.Z, T0.X, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T9.Y, 0.0, | |
| ; CM-NEXT: BFE_UINT * T10.Z, T0.Y, literal.x, T1.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T10.X, T0.Y, literal.x, | |
| ; CM-NEXT: MOV * T10.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T11.X, T0.Y, literal.x, T1.W, | |
| ; CM-NEXT: LSHR * T11.Z, T0.Y, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T11.Y, 0.0, | |
| ; CM-NEXT: BFE_UINT * T12.Z, T0.Z, literal.x, T1.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T12.X, T0.Z, literal.x, | |
| ; CM-NEXT: MOV * T12.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T0.X, T0.Z, literal.x, T1.W, | |
| ; CM-NEXT: LSHR * T0.Z, T0.Z, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T0.Y, 0.0, | |
| ; CM-NEXT: BFE_UINT * T13.Z, T0.W, literal.x, T1.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T13.X, T0.W, literal.x, | |
| ; CM-NEXT: MOV * T13.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T14.X, T0.W, literal.x, T1.W, | |
| ; CM-NEXT: LSHR * T14.Z, T0.W, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T14.Y, 0.0, | |
| ; CM-NEXT: MOV * T8.W, 0.0, | |
| ; CM-NEXT: MOV * T9.W, 0.0, | |
| ; CM-NEXT: MOV * T10.W, 0.0, | |
| ; CM-NEXT: MOV * T11.W, 0.0, | |
| ; CM-NEXT: MOV * T12.W, 0.0, | |
| ; CM-NEXT: MOV * T0.W, 0.0, | |
| ; CM-NEXT: MOV * T13.W, 0.0, | |
| ; CM-NEXT: MOV * T14.W, 0.0, | |
| ; CM-NEXT: LSHR * T15.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T16.X, PV.X, literal.x, | |
| ; CM-NEXT: 28(3.923636e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T17.X, T15.X, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T18.X, T15.X, literal.x, | |
| ; CM-NEXT: 20(2.802597e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T19.X, T15.X, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T20.X, T15.X, literal.x, | |
| ; CM-NEXT: 12(1.681558e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T21.X, T15.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T22.X, T15.X, literal.x, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <16 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <16 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <16 x i8> %load to <16 x i64> | |
| store <16 x i64> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v16i8_to_v16i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v16i8_to_v16i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[2:5], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v1, 16, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v6, 8, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v15, 16, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v16, 24, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v2, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v6, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v6, v1, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v10, 8, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v11, 16, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v9, 31, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v8, 24, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v12, 24, v4 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[6:9], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v6, v11, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v11, v10, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v9, v5, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v8, v12, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v10, 31, v9 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v12, 31, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v1, 8, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v13, 16, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v14, 8, v3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[9:12], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v17, v4, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v19, v1, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v5, v3, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[6:9], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v11, v16, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v9, v15, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v7, v14, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v13, v13, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v18, 31, v17 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v20, 31, v19 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v16, 31, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v15, 24, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v6, 31, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v12, 31, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v10, 31, v9 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v8, 31, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v14, 31, v13 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[17:20], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[13:16], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[5:8], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[9:12], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v16i8_to_v16i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x70 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x60 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v11, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v10, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x50 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 64 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 48 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 32 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s0 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 16 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v21, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v20, s2 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v23, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v22, s0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v4, 16, v3 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v4, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 31, v3 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v6, 24, v3 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[8:9], v[4:7] | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v8, 16, v0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v4, 8, v3 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v4, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v3, v3, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v4, 31, v3 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v6, 31, v5 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[10:11], v[3:6] | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v7, 8, v1 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v3, 16, v2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v4, 24, v2 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v4, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v3, v3, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v6, 31, v5 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v4, 31, v3 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[3:6] | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v9, 24, v0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v3, 8, v2 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v3, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v2, v2, 0, 8 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v6, 16, v1 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[2:5] | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v10, 8, v0 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v2, v6, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v5, 31, v1 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v4, 24, v1 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[18:19], v[2:5] | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v7, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v1, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v2, v10, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v10, v9, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v8, v8, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v11, 31, v10 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[20:21], v[4:7] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[22:23], v[8:11] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[12:13], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v16i8_to_v16i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v4, 16, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v4, v4, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v8, 16, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v9, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v7, 31, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v6, 24, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v10, 16, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v6, v9, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v4, v8, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v11, 16, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v4, v10, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v13, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v16, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v7, 31, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v6, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v14, 8, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v15, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v12, 8, v1 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v18, v13, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v4, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v2, v16, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v11, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v12, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v8, v1, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v12, v3, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v6, v15, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v14, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v19, 31, v18 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v17, 31, v16 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v13, 31, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 31, v10 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v15, 31, v14 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v16i8_to_v16i64: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @14, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @12 | |
| ; EG-NEXT: ALU 67, @17, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T21.XYZW, T22.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T13.XYZW, T16.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T15.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T12.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T11.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T14.XYZW, T10.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T17.XYZW, T9.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T18.XYZW, T8.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: Fetch clause starting at 12: | |
| ; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 14: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 17: | |
| ; EG-NEXT: LSHR * T8.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T9.X, PV.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T10.X, PV.X, literal.y, | |
| ; EG-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T11.X, T8.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T12.X, T8.X, literal.y, | |
| ; EG-NEXT: 12(1.681558e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT * T13.X, T0.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T14.X, T0.Y, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T13.Y, PV.X, literal.y, | |
| ; EG-NEXT: LSHR T1.W, T0.W, literal.x, | |
| ; EG-NEXT: ADD_INT * T15.X, T8.X, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 20(2.802597e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ASHR T14.Y, PV.X, literal.x, | |
| ; EG-NEXT: BFE_INT T13.Z, PV.W, 0.0, literal.y, | |
| ; EG-NEXT: LSHR * T1.W, T0.Y, literal.y, | |
| ; EG-NEXT: 31(4.344025e-44), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T16.X, T8.X, literal.x, | |
| ; EG-NEXT: BFE_INT T14.Z, PV.W, 0.0, literal.y, | |
| ; EG-NEXT: ASHR * T17.W, T0.X, literal.z, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 24(3.363116e-44), 8(1.121039e-44) | |
| ; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T18.X, T0.X, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T17.Z, T0.X, literal.y, | |
| ; EG-NEXT: LSHR T1.W, T0.X, literal.z, | |
| ; EG-NEXT: ASHR * T19.W, T0.Y, literal.w, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44) | |
| ; EG-NEXT: BFE_INT T17.X, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T18.Y, PV.X, literal.y, | |
| ; EG-NEXT: ASHR T19.Z, T0.Y, literal.z, | |
| ; EG-NEXT: LSHR T1.W, T0.X, literal.x, | |
| ; EG-NEXT: LSHR * T2.W, T0.Y, literal.w, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT T19.X, PS, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T17.Y, PV.X, literal.y, | |
| ; EG-NEXT: BFE_INT T18.Z, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR * T20.W, T0.Z, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: BFE_INT T0.X, T0.Z, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T19.Y, PV.X, literal.y, | |
| ; EG-NEXT: ASHR T20.Z, T0.Z, literal.z, | |
| ; EG-NEXT: LSHR T1.W, T0.Z, literal.w, | |
| ; EG-NEXT: ASHR * T21.W, T0.W, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT T20.X, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T0.Y, PV.X, literal.y, | |
| ; EG-NEXT: ASHR T21.Z, T0.W, literal.z, | |
| ; EG-NEXT: LSHR T1.W, T0.Z, literal.x, | |
| ; EG-NEXT: LSHR * T0.W, T0.W, literal.w, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT T21.X, PS, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T20.Y, PV.X, literal.y, | |
| ; EG-NEXT: BFE_INT T0.Z, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T18.W, T18.Z, literal.y, | |
| ; EG-NEXT: ASHR * T14.W, T14.Z, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: ADD_INT T22.X, T8.X, literal.x, | |
| ; EG-NEXT: ASHR T21.Y, PV.X, literal.y, | |
| ; EG-NEXT: ASHR T0.W, PV.Z, literal.y, | |
| ; EG-NEXT: ASHR * T13.W, T13.Z, literal.y, | |
| ; EG-NEXT: 28(3.923636e-44), 31(4.344025e-44) | |
| ; | |
| ; CM-LABEL: global_sextload_v16i8_to_v16i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @14, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @12 | |
| ; CM-NEXT: ALU 69, @17, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T21, T8.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T22.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T13, T16.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T20, T15.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T18, T12.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T19, T11.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T14, T10.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T17, T9.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 12: | |
| ; CM-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 14: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 17: | |
| ; CM-NEXT: LSHR * T8.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T9.X, PV.X, literal.x, | |
| ; CM-NEXT: 28(3.923636e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T10.X, T8.X, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T11.X, T8.X, literal.x, | |
| ; CM-NEXT: 20(2.802597e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T12.X, T8.X, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T13.X, T0.Y, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T1.Y, T0.X, literal.x, | |
| ; CM-NEXT: LSHR T1.Z, T0.X, literal.y, | |
| ; CM-NEXT: LSHR * T1.W, T0.Y, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; CM-NEXT: BFE_INT T14.X, T0.W, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T13.Y, PV.X, literal.y, | |
| ; CM-NEXT: LSHR T2.Z, T0.Z, literal.x, | |
| ; CM-NEXT: LSHR * T2.W, T0.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: ADD_INT T15.X, T8.X, literal.x, | |
| ; CM-NEXT: ASHR T14.Y, PV.X, literal.y, | |
| ; CM-NEXT: BFE_INT T13.Z, PV.W, 0.0, literal.z, | |
| ; CM-NEXT: LSHR * T2.W, T0.W, literal.z, | |
| ; CM-NEXT: 12(1.681558e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T16.X, T8.X, literal.x, | |
| ; CM-NEXT: LSHR T2.Y, T0.Z, literal.y, | |
| ; CM-NEXT: BFE_INT T14.Z, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: ASHR * T17.W, T0.W, literal.z, | |
| ; CM-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 31(4.344025e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T18.X, T0.Z, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T3.Y, T0.W, literal.y, | |
| ; CM-NEXT: ASHR T17.Z, T0.W, literal.z, | |
| ; CM-NEXT: ASHR * T19.W, T0.Z, literal.w, | |
| ; CM-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 31(4.344025e-44) | |
| ; CM-NEXT: BFE_INT T17.X, PV.Y, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T18.Y, PV.X, literal.y, | |
| ; CM-NEXT: ASHR T19.Z, T0.Z, literal.z, | |
| ; CM-NEXT: ASHR * T14.W, T14.Z, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T19.X, T2.Y, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T17.Y, PV.X, literal.y, | |
| ; CM-NEXT: BFE_INT T18.Z, T2.Z, 0.0, literal.x, | |
| ; CM-NEXT: ASHR * T20.W, T0.Y, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: BFE_INT T21.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T19.Y, PV.X, literal.y, | |
| ; CM-NEXT: ASHR T20.Z, T0.Y, literal.z, | |
| ; CM-NEXT: ASHR * T0.W, T0.X, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T20.X, T1.W, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T21.Y, PV.X, literal.y, | |
| ; CM-NEXT: ASHR T0.Z, T0.X, literal.z, | |
| ; CM-NEXT: ASHR * T18.W, T18.Z, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T0.X, T1.Z, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T20.Y, PV.X, literal.y, | |
| ; CM-NEXT: BFE_INT T21.Z, T1.Y, 0.0, literal.x, | |
| ; CM-NEXT: ASHR * T13.W, T13.Z, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: ADD_INT T22.X, T8.X, literal.x, | |
| ; CM-NEXT: ASHR T0.Y, PV.X, literal.y, | |
| ; CM-NEXT: ASHR * T21.W, PV.Z, literal.y, | |
| ; CM-NEXT: 4(5.605194e-45), 31(4.344025e-44) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <16 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <16 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <16 x i8> %load to <16 x i64> | |
| store <16 x i64> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v32i8_to_v32i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v32i8_to_v32i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[18:21], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[30:33], v[0:1], s[8:11], 0 addr64 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v57, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v59, v1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v44, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v46, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v54, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v41, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v51, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v38, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v28, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v35, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v48, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v26, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v22, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v16, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v12, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v14, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v8, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v10, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v5, v1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v45, 24, v21 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v43, v21, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v58, 24, v32 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v56, v32, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v42, 24, v20 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v40, v20, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v55, 24, v31 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v53, v31, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[56:59], off, s[0:3], 0 offset:208 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v56, v1 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[43:46], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v43, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v9, v19, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v7, 0xff, v19 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v17, v21, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v15, 0xff, v21 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v39, 24, v19 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v37, v19, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v21, v30, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v19, 0xff, v30 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v29, v32, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v27, 0xff, v32 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v52, 24, v30 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v50, v30, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[53:56], off, s[0:3], 0 offset:176 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v53, v1 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[40:43], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v40, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v30, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v6, v18, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v4, 0xff, v18 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v13, v20, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v11, 0xff, v20 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v36, 24, v18 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v34, v18, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v2, 24, v33 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v25, v31, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v23, 0xff, v31 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v49, v33, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v47, 0xff, v33 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_u32 v0, v33, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[50:53], off, s[0:3], 0 offset:144 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[37:40], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v37, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v50, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v20, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v18, v1 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[27:30], off, s[0:3], 0 offset:192 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[23:26], off, s[0:3], 0 offset:160 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[19:22], off, s[0:3], 0 offset:128 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:240 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[15:18], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[11:14], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[7:10], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v7, v1 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[34:37], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[47:50], off, s[0:3], 0 offset:224 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v32i8_to_v32i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[2:5], v[0:1] | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, 16, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[6:9], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x70 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 0x50 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 48 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v11, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, v1 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, v1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v12, 24, v4 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v10, v4, 16, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[10:13] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 16 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v12, 24, v3 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v10, v3, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[10:13] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 0xf0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v12, 24, v2 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v10, v2, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[10:13] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 0xd0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(3) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v12, 24, v9 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v10, v9, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[10:13] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 0xb0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v12, 24, v8 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v10, v8, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[10:13] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 0x90 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v12, 24, v7 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v10, v7, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[10:13] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 0x60 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v12, 24, v6 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v10, v6, 16, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[10:13] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 64 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v12, v5, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v10, 0xff, v5 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[10:13] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s5 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 32 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v12, v4, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v10, 0xff, v4 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[10:13] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s5 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v12, v3, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v10, 0xff, v3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, s4 | |
| ; GCN-HSA-NEXT: s_add_u32 s4, s0, 0xe0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[3:4], v[10:13] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v12, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v10, 0xff, v2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0 | |
| ; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[2:3], v[10:13] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v2, s4 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v11, v9, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v9, 0xff, v9 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v10, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, s5 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[2:3], v[9:12] | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v2, 24, v5 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v0, v5, 16, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xc0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, v1 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3] | |
| ; GCN-HSA-NEXT: v_bfe_u32 v16, v8, 8, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v2, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, s3 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xa0 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 0x80 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v14, 0xff, v8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, v1 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v9, v7, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_u32 v13, v6, 8, 8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v11, 0xff, v6 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v7, 0xff, v7 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[2:3], v[14:17] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v2, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v3, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[2:3], v[7:10] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[11:14] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v32i8_to_v32i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v2, vcc, 16, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[12:15], v[2:3] | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[38:41], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v6, 0xff, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v8, 8, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v33, 24, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v31, v14, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v14, 0 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v32, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v34, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v30, 24, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v28, v12, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[31:34], off, s[0:3], 0 offset:208 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v29, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v31, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v9, 0xff, v15 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v25, 0xff, v40 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v27, 8, v40 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[28:31], off, s[0:3], 0 offset:144 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v26, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v28, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v0, 0xff, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v2, 8, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v3, 0xff, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v5, 8, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v11, 8, v15 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v44, 24, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v42, v13, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v47, 24, v15 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v45, v15, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v17, 0xff, v38 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v19, 8, v38 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v21, 0xff, v39 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v23, 8, v39 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v15, 24, v41 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v13, v41, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v22, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v24, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v18, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v20, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v16, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v10, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v12, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v7, v14 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[25:28], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[21:24], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[17:20], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[13:16], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[9:12], off, s[0:3], 0 offset:224 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v35, 0xff, v41 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v9, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v37, 8, v41 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v51, 24, v38 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v49, v38, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v54, 24, v40 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v52, v40, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v40, 24, v39 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v38, v39, 16, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v39, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v41, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v46, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v48, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v53, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v55, v14 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[6:9], off, s[0:3], 0 offset:192 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v6, v14 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[38:41], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[45:48], off, s[0:3], 0 offset:240 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v43, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v45, v14 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[52:55], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v50, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v52, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v36, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v38, v14 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[3:6], off, s[0:3], 0 offset:160 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, v14 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[42:45], off, s[0:3], 0 offset:176 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[49:52], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[35:38], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:128 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v32i8_to_v32i64: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @26, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @22 | |
| ; EG-NEXT: ALU 12, @29, KC0[], KC1[] | |
| ; EG-NEXT: TEX 0 @24 | |
| ; EG-NEXT: ALU 102, @42, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T42.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T14.XYZW, T41.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T15.XYZW, T40.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T16.XYZW, T39.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T17.XYZW, T38.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T18.XYZW, T37.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T36.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T13.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 T12.XYZW, T27.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: Fetch clause starting at 22: | |
| ; EG-NEXT: VTX_READ_128 T13.XYZW, T12.X, 16, #1 | |
| ; EG-NEXT: Fetch clause starting at 24: | |
| ; EG-NEXT: VTX_READ_128 T12.XYZW, T12.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 26: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 5(7.006492e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T12.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 29: | |
| ; EG-NEXT: MOV * T0.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T0.X, T13.W, literal.x, PV.W, | |
| ; EG-NEXT: LSHR * T0.Z, T13.W, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T0.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T14.Z, T13.W, literal.x, T0.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T14.X, T13.W, literal.x, | |
| ; EG-NEXT: MOV * T14.Y, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT * T15.X, T13.Z, literal.x, T0.W, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ALU clause starting at 42: | |
| ; EG-NEXT: LSHR * T15.Z, T13.Z, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: MOV T15.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T16.Z, T13.Z, literal.x, T0.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T16.X, T13.Z, literal.x, | |
| ; EG-NEXT: MOV * T16.Y, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T17.X, T13.Y, literal.x, T0.W, | |
| ; EG-NEXT: LSHR * T17.Z, T13.Y, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T17.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T18.Z, T13.Y, literal.x, T0.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T18.X, T13.Y, literal.x, | |
| ; EG-NEXT: MOV * T18.Y, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T19.X, T13.X, literal.x, T0.W, | |
| ; EG-NEXT: LSHR * T19.Z, T13.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T19.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T13.Z, T13.X, literal.x, T0.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T13.X, T13.X, literal.x, | |
| ; EG-NEXT: MOV * T13.Y, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T20.X, T12.W, literal.x, T0.W, | |
| ; EG-NEXT: LSHR * T20.Z, T12.W, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T20.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T21.Z, T12.W, literal.x, T0.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T21.X, T12.W, literal.x, | |
| ; EG-NEXT: MOV * T21.Y, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T22.X, T12.Z, literal.x, T0.W, | |
| ; EG-NEXT: LSHR * T22.Z, T12.Z, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T22.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T23.Z, T12.Z, literal.x, T0.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T23.X, T12.Z, literal.x, | |
| ; EG-NEXT: MOV * T23.Y, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T24.X, T12.Y, literal.x, T0.W, | |
| ; EG-NEXT: LSHR * T24.Z, T12.Y, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T24.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T25.Z, T12.Y, literal.x, T0.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T25.X, T12.Y, literal.x, | |
| ; EG-NEXT: MOV * T25.Y, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T26.X, T12.X, literal.x, T0.W, | |
| ; EG-NEXT: LSHR * T26.Z, T12.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; EG-NEXT: MOV T26.Y, 0.0, | |
| ; EG-NEXT: BFE_UINT * T12.Z, T12.X, literal.x, T0.W, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T12.X, T12.X, literal.x, | |
| ; EG-NEXT: MOV T12.Y, 0.0, | |
| ; EG-NEXT: MOV T0.W, 0.0, | |
| ; EG-NEXT: MOV * T14.W, 0.0, | |
| ; EG-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; 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 * T13.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 * T12.W, 0.0, | |
| ; EG-NEXT: LSHR * T27.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T28.X, PV.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T29.X, PV.X, literal.y, | |
| ; EG-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T30.X, T27.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T31.X, T27.X, literal.y, | |
| ; EG-NEXT: 12(1.681558e-44), 16(2.242078e-44) | |
| ; EG-NEXT: ADD_INT T32.X, T27.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T33.X, T27.X, literal.y, | |
| ; EG-NEXT: 20(2.802597e-44), 24(3.363116e-44) | |
| ; EG-NEXT: ADD_INT T34.X, T27.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T35.X, T27.X, literal.y, | |
| ; EG-NEXT: 28(3.923636e-44), 32(4.484155e-44) | |
| ; EG-NEXT: ADD_INT T36.X, T27.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T37.X, T27.X, literal.y, | |
| ; EG-NEXT: 36(5.044674e-44), 40(5.605194e-44) | |
| ; EG-NEXT: ADD_INT T38.X, T27.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T39.X, T27.X, literal.y, | |
| ; EG-NEXT: 44(6.165713e-44), 48(6.726233e-44) | |
| ; EG-NEXT: ADD_INT T40.X, T27.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T41.X, T27.X, literal.y, | |
| ; EG-NEXT: 52(7.286752e-44), 56(7.847271e-44) | |
| ; EG-NEXT: ADD_INT * T42.X, T27.X, literal.x, | |
| ; EG-NEXT: 60(8.407791e-44), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_zextload_v32i8_to_v32i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @26, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @22 | |
| ; CM-NEXT: ALU 12, @29, KC0[], KC1[] | |
| ; CM-NEXT: TEX 0 @24 | |
| ; CM-NEXT: ALU 108, @42, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T27.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T14, T42.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T12, T41.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T16, T40.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T17, T39.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T13, T38.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T18, T37.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T19, T36.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T20, T35.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T21, T34.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T22, T33.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T23, T32.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T24, T31.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T15, T30.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T25, T29.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T26, T28.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 22: | |
| ; CM-NEXT: VTX_READ_128 T13.XYZW, T12.X, 0, #1 | |
| ; CM-NEXT: Fetch clause starting at 24: | |
| ; CM-NEXT: VTX_READ_128 T15.XYZW, T12.X, 16, #1 | |
| ; CM-NEXT: ALU clause starting at 26: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 5(7.006492e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T12.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 29: | |
| ; CM-NEXT: MOV * T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT * T0.Z, T13.X, literal.x, PV.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.X, T13.X, literal.x, | |
| ; CM-NEXT: MOV * T0.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T14.X, T13.X, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T14.Z, T13.X, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T14.Y, 0.0, | |
| ; CM-NEXT: BFE_UINT * T12.Z, T13.Y, literal.x, T0.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ALU clause starting at 42: | |
| ; CM-NEXT: AND_INT T12.X, T13.Y, literal.x, | |
| ; CM-NEXT: MOV * T12.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T16.X, T13.Y, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T16.Z, T13.Y, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T16.Y, 0.0, | |
| ; CM-NEXT: BFE_UINT * T17.Z, T13.Z, literal.x, T0.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T17.X, T13.Z, literal.x, | |
| ; CM-NEXT: MOV * T17.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T13.X, T13.Z, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T13.Z, T13.Z, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T13.Y, 0.0, | |
| ; CM-NEXT: BFE_UINT * T18.Z, T13.W, literal.x, T0.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T18.X, T13.W, literal.x, | |
| ; CM-NEXT: MOV * T18.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T19.X, T13.W, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T19.Z, T13.W, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T19.Y, 0.0, | |
| ; CM-NEXT: BFE_UINT * T20.Z, T15.X, literal.x, T0.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T20.X, T15.X, literal.x, | |
| ; CM-NEXT: MOV * T20.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T21.X, T15.X, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T21.Z, T15.X, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T21.Y, 0.0, | |
| ; CM-NEXT: BFE_UINT * T22.Z, T15.Y, literal.x, T0.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T22.X, T15.Y, literal.x, | |
| ; CM-NEXT: MOV * T22.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T23.X, T15.Y, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T23.Z, T15.Y, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T23.Y, 0.0, | |
| ; CM-NEXT: BFE_UINT * T24.Z, T15.Z, literal.x, T0.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T24.X, T15.Z, literal.x, | |
| ; CM-NEXT: MOV * T24.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T15.X, T15.Z, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T15.Z, T15.Z, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T15.Y, 0.0, | |
| ; CM-NEXT: BFE_UINT * T25.Z, T15.W, literal.x, T0.W, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T25.X, T15.W, literal.x, | |
| ; CM-NEXT: MOV * T25.Y, 0.0, | |
| ; CM-NEXT: 255(3.573311e-43), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_UINT T26.X, T15.W, literal.x, T0.W, | |
| ; CM-NEXT: LSHR * T26.Z, T15.W, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 24(3.363116e-44) | |
| ; CM-NEXT: MOV T26.Y, 0.0, | |
| ; CM-NEXT: MOV * T0.W, 0.0, | |
| ; CM-NEXT: MOV * T14.W, 0.0, | |
| ; CM-NEXT: MOV * T12.W, 0.0, | |
| ; CM-NEXT: MOV * T16.W, 0.0, | |
| ; CM-NEXT: MOV * T17.W, 0.0, | |
| ; CM-NEXT: MOV * T13.W, 0.0, | |
| ; CM-NEXT: MOV * T18.W, 0.0, | |
| ; CM-NEXT: MOV * T19.W, 0.0, | |
| ; CM-NEXT: MOV * T20.W, 0.0, | |
| ; CM-NEXT: MOV * T21.W, 0.0, | |
| ; CM-NEXT: MOV * T22.W, 0.0, | |
| ; CM-NEXT: MOV * T23.W, 0.0, | |
| ; CM-NEXT: MOV * T24.W, 0.0, | |
| ; CM-NEXT: MOV * T15.W, 0.0, | |
| ; CM-NEXT: MOV * T25.W, 0.0, | |
| ; CM-NEXT: MOV * T26.W, 0.0, | |
| ; CM-NEXT: LSHR * T27.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T28.X, PV.X, literal.x, | |
| ; CM-NEXT: 60(8.407791e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T29.X, T27.X, literal.x, | |
| ; CM-NEXT: 56(7.847271e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T30.X, T27.X, literal.x, | |
| ; CM-NEXT: 52(7.286752e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T31.X, T27.X, literal.x, | |
| ; CM-NEXT: 48(6.726233e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T32.X, T27.X, literal.x, | |
| ; CM-NEXT: 44(6.165713e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T33.X, T27.X, literal.x, | |
| ; CM-NEXT: 40(5.605194e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T34.X, T27.X, literal.x, | |
| ; CM-NEXT: 36(5.044674e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T35.X, T27.X, literal.x, | |
| ; CM-NEXT: 32(4.484155e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T36.X, T27.X, literal.x, | |
| ; CM-NEXT: 28(3.923636e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T37.X, T27.X, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T38.X, T27.X, literal.x, | |
| ; CM-NEXT: 20(2.802597e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T39.X, T27.X, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T40.X, T27.X, literal.x, | |
| ; CM-NEXT: 12(1.681558e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T41.X, T27.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T42.X, T27.X, literal.x, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <32 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <32 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <32 x i8> %load to <32 x i64> | |
| store <32 x i64> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v32i8_to_v32i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v32i8_to_v32i64: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[8:11], 0 addr64 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[0:3], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v12, 16, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v12, v12, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v16, 8, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v15, 31, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v14, 24, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v13, 31, v12 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:240 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v14, v16, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v12, v7, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v17, 16, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v18, 24, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v13, 31, v12 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v15, 31, v14 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:224 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v14, v18, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v12, v17, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v19, 8, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v15, 31, v14 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v13, 31, v12 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v11, 16, v5 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:208 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v14, v19, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v12, v6, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v13, 31, v12 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v15, 31, v14 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v11, v11, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v10, 8, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v8, 16, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v9, 24, v4 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:192 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v14, 31, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v13, 24, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v12, 31, v11 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[11:14], off, s[0:3], 0 offset:176 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v12, v10, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v10, v5, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v7, v9, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v5, v8, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v20, 8, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v8, 31, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v6, 31, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v11, 31, v10 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v13, 31, v12 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[5:8], off, s[0:3], 0 offset:144 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v7, v4, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v9, v20, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(6) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v16, 16, v3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[10:13], off, s[0:3], 0 offset:160 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v8, 31, v7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v10, 31, v9 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[7:10], off, s[0:3], 0 offset:128 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v7, v16, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v17, 8, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v11, 16, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v10, 31, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v9, 24, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v8, 31, v7 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[7:10], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v8, v11, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v13, v17, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v11, v3, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v18, 16, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v15, 24, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v12, 31, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v14, 31, v13 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[11:14], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v13, v15, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v11, v18, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v19, 16, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v14, 31, v13 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v12, 31, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v21, 8, v1 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[11:14], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v12, v1, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v16, v19, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v19, 31, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v18, 24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v1, 8, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v5, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v6, 8, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v4, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v2, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v1, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v6, v6, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v10, v5, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v14, v21, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v1, 31, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v3, 31, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v11, 31, v10 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v13, 31, v12 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v15, 31, v14 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v17, 31, v16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v32i8_to_v32i64: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[4:7], v[0:1] | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, 16, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x70 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x60 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x50 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 64 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v19, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v18, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 48 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v8, 16, v7 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v8, v8, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v11, 31, v7 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v10, 24, v7 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[12:13], v[8:11] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s3 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v8, 8, v7 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v9, v8, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v7, v7, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v8, 31, v7 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v10, 31, v9 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[7:10] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v11, s2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v7, 16, v6 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v8, 24, v6 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v9, v8, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v7, v7, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v10, 31, v9 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v8, 31, v7 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[7:10] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 32 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v7, 8, v6 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v8, v7, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v6, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[18:19], v[6:9] | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v10, 8, v5 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v6, 16, v5 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v6, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v9, 31, v5 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v8, 24, v5 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[11:12], v[6:9] | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v5, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v7, v10, 0, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v17, 16, v4 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v9, 24, v4 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v6, 31, v5 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v8, 31, v7 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[13:14], v[5:8] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s3 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v7, v9, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v17, 0, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v11, 8, v4 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v8, 31, v7 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v6, 31, v5 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v4, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[15:16], v[5:8] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xf0 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v11, 0, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(7) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v10, 16, v3 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, s0 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[8:9], v[4:7] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, s3 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v10, 0, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xe0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v12, 8, v3 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v5, 31, v4 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 31, v3 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v6, 24, v3 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[8:9], v[4:7] | |
| ; GCN-HSA-NEXT: v_bfe_i32 v3, v3, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v12, 0, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v7, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xd0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v13, 16, v2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v14, 24, v2 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v4, 31, v3 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v6, 31, v5 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[7:8], v[3:6] | |
| ; GCN-HSA-NEXT: v_bfe_i32 v7, v14, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v13, 0, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v10, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xc0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v17, 8, v2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v15, 16, v0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v16, 24, v0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v8, 31, v7 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v6, 31, v5 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[9:10], v[5:8] | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v16, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v7, v2, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v11, v15, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v9, v17, 0, 8 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xb0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v18, 16, v1 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v8, 31, v7 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v10, 31, v9 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v19, 8, v1 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v20, 8, v0 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v3, v0, 0, 8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[15:16], v[7:10] | |
| ; GCN-HSA-NEXT: v_bfe_i32 v15, v18, 0, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v7, v1, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v18, 31, v1 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v17, 24, v1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0xa0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v16, 31, v15 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[15:18] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x90 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v9, v19, 0, 8 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v8, 31, v7 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v10, 31, v9 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 0x80 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v20, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v14, 31, v13 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v12, 31, v11 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[7:10] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v4, 31, v3 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v6, 31, v5 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[11:14] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[3:6] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v32i8_to_v32i64: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v4, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[4:5] | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v4, vcc, 16, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[4:7], v[4:5] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v10, 16, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v10, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v14, 16, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v15, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v13, 31, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v12, 24, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 31, v10 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[10:13], off, s[0:3], 0 offset:112 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v16, 16, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v12, v15, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v14, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v13, 31, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 31, v10 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[10:13], off, s[0:3], 0 offset:80 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v17, 16, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v16, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v18, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v13, 31, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v12, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 31, v10 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[10:13], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v19, 8, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v13, v18, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v11, v17, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v14, 31, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v12, 31, v11 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[11:14], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v9, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v11, v3, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v13, v19, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v12, 31, v11 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v14, 31, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v8, 8, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[11:14], off, s[0:3], 0 offset:96 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v8, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v1, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(5) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v1, 16, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v12, v9, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 31, v10 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v13, 31, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v8, v1, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v2, 16, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 24, v6 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[10:13], off, s[0:3], 0 offset:64 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 31, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v10, 24, v7 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:240 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v12, 16, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v3, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v8, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 31, v10 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:208 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v13, 16, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v8, v12, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v18, 24, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v15, 31, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v17, 31, v16 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 31, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v10, 24, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v2, 8, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v20, 8, v5 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[14:17], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:176 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v16, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v18, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v8, v13, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v2, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v19, 8, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v14, 8, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 31, v10 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v12, v20, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v18, v0, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v20, v2, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v1, v4, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:144 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v6, v6, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v3, v14, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v8, v19, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v10, v5, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_i32 v14, v7, 0, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v19, 31, v18 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v21, 31, v20 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v2, 31, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v7, 31, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v4, 31, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v11, 31, v10 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v13, 31, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v9, 31, v8 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v15, 31, v14 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v17, 31, v16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[18:21], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[14:17], off, s[0:3], 0 offset:224 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[6:9], off, s[0:3], 0 offset:192 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[10:13], off, s[0:3], 0 offset:160 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[1:4], off, s[0:3], 0 offset:128 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v32i8_to_v32i64: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 2, @26, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 1 @22 | |
| ; EG-NEXT: ALU 83, @29, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: ALU 50, @113, KC0[], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T41.XYZW, T42.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T26.XYZW, T31.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T40.XYZW, T30.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T25.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T39.XYZW, T24.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T27.XYZW, T23.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T12.XYZW, T22.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T28.XYZW, T21.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T37.XYZW, T20.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T29.XYZW, T19.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T35.XYZW, T18.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T33.XYZW, T17.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T34.XYZW, T16.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T38.XYZW, T15.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T32.XYZW, T14.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T36.XYZW, T13.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: PAD | |
| ; EG-NEXT: Fetch clause starting at 22: | |
| ; EG-NEXT: VTX_READ_128 T12.XYZW, T0.X, 0, #1 | |
| ; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 16, #1 | |
| ; EG-NEXT: ALU clause starting at 26: | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 5(7.006492e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 29: | |
| ; EG-NEXT: LSHR * T13.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T14.X, PV.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T15.X, PV.X, literal.y, | |
| ; EG-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T16.X, T13.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T17.X, T13.X, literal.y, | |
| ; EG-NEXT: 12(1.681558e-44), 16(2.242078e-44) | |
| ; EG-NEXT: ADD_INT T18.X, T13.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T19.X, T13.X, literal.y, | |
| ; EG-NEXT: 20(2.802597e-44), 24(3.363116e-44) | |
| ; EG-NEXT: ADD_INT T20.X, T13.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T21.X, T13.X, literal.y, | |
| ; EG-NEXT: 28(3.923636e-44), 32(4.484155e-44) | |
| ; EG-NEXT: ADD_INT T22.X, T13.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T23.X, T13.X, literal.y, | |
| ; EG-NEXT: 36(5.044674e-44), 40(5.605194e-44) | |
| ; EG-NEXT: ADD_INT T24.X, T13.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T25.X, T13.X, literal.y, | |
| ; EG-NEXT: 44(6.165713e-44), 48(6.726233e-44) | |
| ; EG-NEXT: BFE_INT * T26.X, T0.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T27.X, T0.Y, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T26.Y, PV.X, literal.y, | |
| ; EG-NEXT: LSHR * T1.W, T0.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: BFE_INT T28.X, T0.X, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T27.Y, PV.X, literal.y, | |
| ; EG-NEXT: BFE_INT T26.Z, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR * T1.W, T0.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: BFE_INT T29.X, T12.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T28.Y, PV.X, literal.y, | |
| ; EG-NEXT: BFE_INT T27.Z, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T1.W, T0.X, literal.x, | |
| ; EG-NEXT: ADD_INT * T30.X, T13.X, literal.z, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 52(7.286752e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ASHR T29.Y, PV.X, literal.x, | |
| ; EG-NEXT: BFE_INT T28.Z, PV.W, 0.0, literal.y, | |
| ; EG-NEXT: LSHR * T1.W, T12.W, literal.y, | |
| ; EG-NEXT: 31(4.344025e-44), 8(1.121039e-44) | |
| ; EG-NEXT: ADD_INT T31.X, T13.X, literal.x, | |
| ; EG-NEXT: BFE_INT T29.Z, PV.W, 0.0, literal.y, | |
| ; EG-NEXT: ASHR * T32.W, T12.X, literal.z, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 56(7.847271e-44), 8(1.121039e-44) | |
| ; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T33.X, T12.Z, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T1.Y, T0.Z, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: ASHR T32.Z, T12.X, literal.y, | |
| ; EG-NEXT: LSHR T1.W, T12.X, literal.z, | |
| ; EG-NEXT: ASHR * T34.W, T12.Y, literal.w, | |
| ; EG-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44) | |
| ; EG-NEXT: BFE_INT T32.X, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T33.Y, PV.X, literal.y, | |
| ; EG-NEXT: ASHR T34.Z, T12.Y, literal.z, | |
| ; EG-NEXT: LSHR T1.W, T12.Z, literal.x, | |
| ; EG-NEXT: LSHR * T2.W, T12.Y, literal.w, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT T34.X, PS, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T32.Y, PV.X, literal.y, | |
| ; EG-NEXT: BFE_INT T33.Z, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: LSHR T1.W, T0.W, literal.z, | |
| ; EG-NEXT: ASHR * T35.W, T12.Z, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T36.X, T12.X, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T34.Y, PV.X, literal.y, | |
| ; EG-NEXT: ASHR T35.Z, T12.Z, literal.z, | |
| ; EG-NEXT: LSHR T2.W, T12.Z, literal.w, | |
| ; EG-NEXT: ASHR * T37.W, T12.W, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT T35.X, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T36.Y, PV.X, literal.y, | |
| ; EG-NEXT: ASHR T37.Z, T12.W, literal.z, | |
| ; EG-NEXT: LSHR T2.W, T12.X, literal.x, | |
| ; EG-NEXT: LSHR * T3.W, T12.W, literal.w, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT * T37.X, PS, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ALU clause starting at 113: | |
| ; EG-NEXT: ASHR T35.Y, T35.X, literal.x, | |
| ; EG-NEXT: BFE_INT T36.Z, T2.W, 0.0, literal.y, | |
| ; EG-NEXT: LSHR T2.W, T0.Z, literal.z, | |
| ; EG-NEXT: ASHR * T12.W, T0.X, literal.x, | |
| ; EG-NEXT: 31(4.344025e-44), 8(1.121039e-44) | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T38.X, T12.Y, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T37.Y, T37.X, literal.y, | |
| ; EG-NEXT: ASHR T12.Z, T0.X, literal.z, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR T3.W, T0.X, literal.w, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: ASHR * T39.W, T0.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT T12.X, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T38.Y, PV.X, literal.y, | |
| ; EG-NEXT: ASHR T39.Z, T0.Y, literal.z, | |
| ; EG-NEXT: LSHR T3.W, T12.Y, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: LSHR * T4.W, T0.Y, literal.w, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 16(2.242078e-44) | |
| ; EG-NEXT: BFE_INT T39.X, PS, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T12.Y, PV.X, literal.y, | |
| ; EG-NEXT: BFE_INT T38.Z, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T36.W, T36.Z, literal.y, | |
| ; EG-NEXT: ASHR * T40.W, T0.Z, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: BFE_INT T0.X, T0.Z, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T39.Y, PV.X, literal.y, | |
| ; EG-NEXT: ASHR T40.Z, T0.Z, literal.z, | |
| ; EG-NEXT: ASHR T38.W, PV.Z, literal.y, | |
| ; EG-NEXT: ASHR * T41.W, T0.W, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T40.X, T2.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T0.Y, PV.X, literal.y, | |
| ; EG-NEXT: ASHR T41.Z, T0.W, literal.z, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: ASHR T33.W, T33.Z, literal.y, | |
| ; EG-NEXT: ASHR * T29.W, T29.Z, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T41.X, T1.W, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T40.Y, PV.X, literal.y, | |
| ; EG-NEXT: BFE_INT T0.Z, T1.Y, 0.0, literal.x, | |
| ; EG-NEXT: ASHR T28.W, T28.Z, literal.y, | |
| ; EG-NEXT: ASHR * T27.W, T27.Z, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; EG-NEXT: ADD_INT T42.X, T13.X, literal.x, | |
| ; EG-NEXT: ASHR T41.Y, PV.X, literal.y, | |
| ; EG-NEXT: ASHR T0.W, PV.Z, literal.y, | |
| ; EG-NEXT: ASHR * T26.W, T26.Z, literal.y, | |
| ; EG-NEXT: 60(8.407791e-44), 31(4.344025e-44) | |
| ; | |
| ; CM-LABEL: global_sextload_v32i8_to_v32i64: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 2, @26, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 1 @22 | |
| ; CM-NEXT: ALU 86, @29, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: ALU 52, @116, KC0[], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T41, T13.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T42.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T28, T31.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T40, T30.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T38, T25.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T39, T24.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T36, T23.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T37, T22.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T33, T21.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T12, T20.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T29, T19.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T35, T18.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T27, T17.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T34, T16.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T26, T15.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T32, T14.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 22: | |
| ; CM-NEXT: VTX_READ_128 T12.XYZW, T0.X, 16, #1 | |
| ; CM-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 26: | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 5(7.006492e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 29: | |
| ; CM-NEXT: LSHR * T13.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T14.X, PV.X, literal.x, | |
| ; CM-NEXT: 60(8.407791e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T15.X, T13.X, literal.x, | |
| ; CM-NEXT: 56(7.847271e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T16.X, T13.X, literal.x, | |
| ; CM-NEXT: 52(7.286752e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T17.X, T13.X, literal.x, | |
| ; CM-NEXT: 48(6.726233e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T18.X, T13.X, literal.x, | |
| ; CM-NEXT: 44(6.165713e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T19.X, T13.X, literal.x, | |
| ; CM-NEXT: 40(5.605194e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T20.X, T13.X, literal.x, | |
| ; CM-NEXT: 36(5.044674e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T21.X, T13.X, literal.x, | |
| ; CM-NEXT: 32(4.484155e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T22.X, T13.X, literal.x, | |
| ; CM-NEXT: 28(3.923636e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T23.X, T13.X, literal.x, | |
| ; CM-NEXT: LSHR T1.Z, T0.X, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHR * T1.W, T0.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 24(3.363116e-44), 8(1.121039e-44) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T24.X, T13.X, literal.x, | |
| ; CM-NEXT: LSHR T1.Y, T0.Y, literal.y, | |
| ; CM-NEXT: LSHR T2.Z, T0.Z, literal.z, | |
| ; CM-NEXT: LSHR * T2.W, T0.Z, literal.y, | |
| ; CM-NEXT: 20(2.802597e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T25.X, T13.X, literal.x, | |
| ; CM-NEXT: LSHR T2.Y, T0.W, literal.x, | |
| ; CM-NEXT: LSHR T3.Z, T0.W, literal.y, | |
| ; CM-NEXT: LSHR * T3.W, T12.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 16(2.242078e-44), 8(1.121039e-44) | |
| ; CM-NEXT: BFE_INT T26.X, T12.W, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T3.Y, T12.Y, literal.y, | |
| ; CM-NEXT: LSHR T4.Z, T12.X, literal.x, | |
| ; CM-NEXT: LSHR * T4.W, T12.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; CM-NEXT: BFE_INT T27.X, T12.Z, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T26.Y, PV.X, literal.y, | |
| ; CM-NEXT: LSHR T5.Z, T0.Y, literal.x, | |
| ; CM-NEXT: LSHR * T5.W, T12.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: BFE_INT T28.X, T0.Y, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T27.Y, PV.X, literal.y, | |
| ; CM-NEXT: BFE_INT T26.Z, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: LSHR * T5.W, T12.Z, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: BFE_INT T29.X, T12.Y, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T28.Y, PV.X, literal.y, | |
| ; CM-NEXT: BFE_INT T27.Z, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: ASHR * T26.W, PV.Z, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: ADD_INT T30.X, T13.X, literal.x, | |
| ; CM-NEXT: ASHR T29.Y, PV.X, literal.y, | |
| ; CM-NEXT: BFE_INT T28.Z, T5.Z, 0.0, literal.z, | |
| ; CM-NEXT: ASHR * T27.W, PV.Z, literal.y, | |
| ; CM-NEXT: 12(1.681558e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T31.X, T13.X, literal.x, | |
| ; CM-NEXT: LSHR T4.Y, T12.Z, literal.y, | |
| ; CM-NEXT: BFE_INT T29.Z, T4.W, 0.0, literal.x, | |
| ; CM-NEXT: ASHR * T32.W, T12.W, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 31(4.344025e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T33.X, T12.X, 0.0, literal.x, | |
| ; CM-NEXT: LSHR T5.Y, T12.W, literal.y, | |
| ; CM-NEXT: ASHR T32.Z, T12.W, literal.z, | |
| ; CM-NEXT: ASHR * T34.W, T12.Z, literal.w, | |
| ; CM-NEXT: 8(1.121039e-44), 16(2.242078e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 31(4.344025e-44) | |
| ; CM-NEXT: BFE_INT T32.X, PV.Y, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T33.Y, PV.X, literal.y, | |
| ; CM-NEXT: ASHR T34.Z, T12.Z, literal.z, | |
| ; CM-NEXT: ASHR * T29.W, T29.Z, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T34.X, T4.Y, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T32.Y, PV.X, literal.y, | |
| ; CM-NEXT: BFE_INT T33.Z, T4.Z, 0.0, literal.x, | |
| ; CM-NEXT: ASHR * T35.W, T12.Y, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: BFE_INT * T36.X, T0.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ALU clause starting at 116: | |
| ; CM-NEXT: ASHR T34.Y, T34.X, literal.x, | |
| ; CM-NEXT: ASHR T35.Z, T12.Y, literal.y, | |
| ; CM-NEXT: ASHR * T12.W, T12.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 31(4.344025e-44), 24(3.363116e-44) | |
| ; CM-NEXT: BFE_INT T35.X, T3.Y, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T36.Y, T36.X, literal.y, | |
| ; CM-NEXT: ASHR T12.Z, T12.X, literal.z, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: ASHR * T33.W, T33.Z, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T12.X, T3.W, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T35.Y, PV.X, literal.y, | |
| ; CM-NEXT: BFE_INT T36.Z, T3.Z, 0.0, literal.x, | |
| ; CM-NEXT: ASHR * T37.W, T0.W, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: BFE_INT T38.X, T0.Z, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T12.Y, PV.X, literal.y, | |
| ; CM-NEXT: ASHR T37.Z, T0.W, literal.z, | |
| ; CM-NEXT: ASHR * T39.W, T0.Z, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T37.X, T2.Y, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T38.Y, PV.X, literal.y, | |
| ; CM-NEXT: ASHR T39.Z, T0.Z, literal.z, | |
| ; CM-NEXT: ASHR * T36.W, T36.Z, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T39.X, T2.W, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T37.Y, PV.X, literal.y, | |
| ; CM-NEXT: BFE_INT T38.Z, T2.Z, 0.0, literal.x, | |
| ; CM-NEXT: ASHR * T40.W, T0.Y, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: BFE_INT T41.X, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T39.Y, PV.X, literal.y, | |
| ; CM-NEXT: ASHR T40.Z, T0.Y, literal.z, | |
| ; CM-NEXT: ASHR * T0.W, T0.X, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T40.X, T1.Y, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T41.Y, PV.X, literal.y, | |
| ; CM-NEXT: ASHR T0.Z, T0.X, literal.z, | |
| ; CM-NEXT: ASHR * T38.W, T38.Z, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T0.X, T1.W, 0.0, literal.x, | |
| ; CM-NEXT: ASHR T40.Y, PV.X, literal.y, | |
| ; CM-NEXT: BFE_INT T41.Z, T1.Z, 0.0, literal.x, | |
| ; CM-NEXT: ASHR * T28.W, T28.Z, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 31(4.344025e-44) | |
| ; CM-NEXT: ADD_INT T42.X, T13.X, literal.x, | |
| ; CM-NEXT: ASHR T0.Y, PV.X, literal.y, | |
| ; CM-NEXT: ASHR * T41.W, PV.Z, literal.y, | |
| ; CM-NEXT: 4(5.605194e-45), 31(4.344025e-44) | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <32 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <32 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <32 x i8> %load to <32 x i64> | |
| store <32 x i64> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; XFUNC-LABEL: {{^}}global_zextload_v64i8_to_v64i64: | |
| ; define amdgpu_kernel void @global_zextload_v64i8_to_v64i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; %load = load <64 x i8>, ptr addrspace(1) %in | |
| ; %ext = zext <64 x i8> %load to <64 x i64> | |
| ; store <64 x i64> %ext, ptr addrspace(1) %out | |
| ; ret void | |
| ; } | |
| ; XFUNC-LABEL: {{^}}global_sextload_v64i8_to_v64i64: | |
| ; define amdgpu_kernel void @global_sextload_v64i8_to_v64i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; %load = load <64 x i8>, ptr addrspace(1) %in | |
| ; %ext = sext <64 x i8> %load to <64 x i64> | |
| ; store <64 x i64> %ext, ptr addrspace(1) %out | |
| ; ret void | |
| ; } | |
| define amdgpu_kernel void @global_zextload_i8_to_i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_i8_to_i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_short v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_i8_to_i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ubyte v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_short v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_i8_to_i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_short v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_i8_to_i16: | |
| ; 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: 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) | |
| ; | |
| ; CM-LABEL: global_zextload_i8_to_i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 10, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT MSKOR T0.XW, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: AND_INT * T0.W, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.x, | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: LSHL T0.X, T0.X, PV.W, | |
| ; CM-NEXT: LSHL * T0.W, literal.x, PV.W, | |
| ; CM-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; CM-NEXT: MOV T0.Y, 0.0, | |
| ; CM-NEXT: MOV * T0.Z, 0.0, | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %a = load i8, ptr addrspace(1) %in | |
| %ext = zext i8 %a to i16 | |
| store i16 %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_i8_to_i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_i8_to_i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_short v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_i8_to_i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_sbyte v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_short v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_i8_to_i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_short v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_i8_to_i16: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 12, @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: BFE_INT T0.W, T0.X, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 3(4.203895e-45) | |
| ; EG-NEXT: AND_INT T0.W, PV.W, literal.x, | |
| ; EG-NEXT: LSHL * T1.W, PS, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 3(4.203895e-45) | |
| ; EG-NEXT: LSHL T0.X, PV.W, PS, | |
| ; EG-NEXT: LSHL * T0.W, literal.x, PS, | |
| ; 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) | |
| ; | |
| ; CM-LABEL: global_sextload_i8_to_i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 12, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT MSKOR T0.XW, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: AND_INT T0.Z, KC0[2].Y, literal.x, | |
| ; CM-NEXT: BFE_INT * T0.W, T0.X, 0.0, literal.y, | |
| ; CM-NEXT: 3(4.203895e-45), 8(1.121039e-44) | |
| ; CM-NEXT: AND_INT T1.Z, PV.W, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.Z, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 3(4.203895e-45) | |
| ; CM-NEXT: LSHL T0.X, PV.Z, PV.W, | |
| ; CM-NEXT: LSHL * T0.W, literal.x, PV.W, | |
| ; CM-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; CM-NEXT: MOV T0.Y, 0.0, | |
| ; CM-NEXT: MOV * T0.Z, 0.0, | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %a = load i8, ptr addrspace(1) %in | |
| %ext = sext i8 %a to i16 | |
| store i16 %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v1i8_to_v1i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v1i8_to_v1i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_short v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v1i8_to_v1i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ubyte v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_short v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v1i8_to_v1i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_short v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v1i8_to_v1i16: | |
| ; 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: 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) | |
| ; | |
| ; CM-LABEL: global_zextload_v1i8_to_v1i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 10, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT MSKOR T0.XW, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: AND_INT * T0.W, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.x, | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: LSHL T0.X, T0.X, PV.W, | |
| ; CM-NEXT: LSHL * T0.W, literal.x, PV.W, | |
| ; CM-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; CM-NEXT: MOV T0.Y, 0.0, | |
| ; CM-NEXT: MOV * T0.Z, 0.0, | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %load = load <1 x i8>, ptr addrspace(1) %in | |
| %ext = zext <1 x i8> %load to <1 x i16> | |
| store <1 x i16> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v1i8_to_v1i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v1i8_to_v1i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_short v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v1i8_to_v1i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_sbyte v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: flat_store_short v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v1i8_to_v1i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_short v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v1i8_to_v1i16: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 12, @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: BFE_INT T0.W, T0.X, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 3(4.203895e-45) | |
| ; EG-NEXT: AND_INT T0.W, PV.W, literal.x, | |
| ; EG-NEXT: LSHL * T1.W, PS, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 3(4.203895e-45) | |
| ; EG-NEXT: LSHL T0.X, PV.W, PS, | |
| ; EG-NEXT: LSHL * T0.W, literal.x, PS, | |
| ; 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) | |
| ; | |
| ; CM-LABEL: global_sextload_v1i8_to_v1i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 12, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT MSKOR T0.XW, T1.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: AND_INT T0.Z, KC0[2].Y, literal.x, | |
| ; CM-NEXT: BFE_INT * T0.W, T0.X, 0.0, literal.y, | |
| ; CM-NEXT: 3(4.203895e-45), 8(1.121039e-44) | |
| ; CM-NEXT: AND_INT T1.Z, PV.W, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.Z, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 3(4.203895e-45) | |
| ; CM-NEXT: LSHL T0.X, PV.Z, PV.W, | |
| ; CM-NEXT: LSHL * T0.W, literal.x, PV.W, | |
| ; CM-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; CM-NEXT: MOV T0.Y, 0.0, | |
| ; CM-NEXT: MOV * T0.Z, 0.0, | |
| ; CM-NEXT: LSHR * T1.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %load = load <1 x i8>, ptr addrspace(1) %in | |
| %ext = sext <1 x i8> %load to <1 x i16> | |
| store <1 x i16> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v2i8_to_v2i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v2i8_to_v2i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v1, 8, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v0, v0, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xff00ff, v0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v2i8_to_v2i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ushort v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v3, 8, v2 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v2, v2, v3 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v2, 0xff00ff, v2 | |
| ; GCN-HSA-NEXT: flat_store_dword v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v2i8_to_v2i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s0, 0xc060c00 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v1, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v0, v1, v0, s0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v2i8_to_v2i16: | |
| ; 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 T5.X, T6.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: PAD | |
| ; EG-NEXT: Fetch clause starting at 6: | |
| ; EG-NEXT: VTX_READ_16 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: LSHL * T0.W, T5.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T0.W, PV.W, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T5.X, literal.y, | |
| ; EG-NEXT: 16711680(2.341805e-38), 255(3.573311e-43) | |
| ; EG-NEXT: OR_INT T5.X, PS, PV.W, | |
| ; EG-NEXT: LSHR * T6.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_zextload_v2i8_to_v2i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 7, @9, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T5.X, T6.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_16 T5.X, T5.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T5.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 9: | |
| ; CM-NEXT: LSHL * T0.W, T5.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.W, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, T5.X, literal.y, | |
| ; CM-NEXT: 16711680(2.341805e-38), 255(3.573311e-43) | |
| ; CM-NEXT: OR_INT * T5.X, PV.W, PV.Z, | |
| ; CM-NEXT: LSHR * T6.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %load = load <2 x i8>, ptr addrspace(1) %in | |
| %ext = zext <2 x i8> %load to <2 x i16> | |
| store <2 x i16> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v2i8_to_v2i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v2i8_to_v2i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v1, v0, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v0, v0, v1 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v2i8_to_v2i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_load_ushort v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_bfe_i32 v3, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v2, v2, 0, 8 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v3, 16, v3 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v2, 0xffff, v2 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v2, v2, v3 | |
| ; GCN-HSA-NEXT: flat_store_dword v[0:1], v2 | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v2i8_to_v2i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s7, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s6 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s11, s7 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s9, s3 | |
| ; GCN-NOHSA-VI-NEXT: buffer_load_ushort v0, off, s[8:11], 0 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 0xffff | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, 8 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s5, s1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v1, v1, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_sdwa v0, v2, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v0, v1, v0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dword v0, off, s[4:7], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v2i8_to_v2i16: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 1, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 16, @10, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.X, T6.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: PAD | |
| ; EG-NEXT: Fetch clause starting at 6: | |
| ; EG-NEXT: VTX_READ_16 T5.X, T5.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 8: | |
| ; EG-NEXT: MOV * T0.Y, T2.X, | |
| ; EG-NEXT: MOV * T5.X, KC0[2].Z, | |
| ; EG-NEXT: ALU clause starting at 10: | |
| ; EG-NEXT: AND_INT T0.W, T5.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.Y, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T2.X, PV.W, | |
| ; EG-NEXT: MOV * T0.Y, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, PV.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T0.Z, T0.W, 0.0, literal.x, | |
| ; EG-NEXT: BFE_INT * T0.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: LSHL T0.W, PV.W, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.Z, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: OR_INT T5.X, PS, PV.W, | |
| ; EG-NEXT: LSHR * T6.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; | |
| ; CM-LABEL: global_sextload_v2i8_to_v2i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 1, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 16, @10, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T5.X, T6.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_16 T5.X, T5.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV * T0.Y, T2.X, | |
| ; CM-NEXT: MOV * T5.X, KC0[2].Z, | |
| ; CM-NEXT: ALU clause starting at 10: | |
| ; CM-NEXT: AND_INT T0.Z, T5.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; CM-NEXT: OR_INT * T0.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T2.X, PV.W, | |
| ; CM-NEXT: MOV * T0.Y, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, PV.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT T0.Z, T0.W, 0.0, literal.x, | |
| ; CM-NEXT: BFE_INT * T0.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHL T1.Z, PV.W, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.Z, literal.y, | |
| ; CM-NEXT: 16(2.242078e-44), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T5.X, PV.W, PV.Z, | |
| ; CM-NEXT: LSHR * T6.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| %load = load <2 x i8>, ptr addrspace(1) %in | |
| %ext = sext <2 x i8> %load to <2 x i16> | |
| store <2 x i16> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v4i8_to_v4i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v4i8_to_v4i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v0, v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v1, 0xff00, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v2, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v3, 0xff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v0, v2, v0, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v2, 8, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v1, 0xff00ff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v0, v3, v2 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v4i8_to_v4i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dword v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v3, 0xff00, v2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v4, 24, v2 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v5, 0xff, v2 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v2, v4, v2, 16 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v4, 8, v3 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v3, 0xff00ff, v2 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v2, v5, v4 | |
| ; GCN-HSA-NEXT: flat_store_dwordx2 v[0:1], v[2:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v4i8_to_v4i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, 0x5040c00 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dword v1, v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v2, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v0, v1, 8, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v0, v0, v1, s4 | |
| ; GCN-NOHSA-VI-NEXT: v_alignbit_b32 v1, v2, v1, 16 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v1, 0xff00ff, v1 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v4i8_to_v4i16: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 3, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 31, @12, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T8.XY, T0.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.Y, T4.X, | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 12: | |
| ; EG-NEXT: AND_INT T0.W, T0.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.Y, literal.y, | |
| ; EG-NEXT: 255(3.573311e-43), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T4.X, PV.W, | |
| ; EG-NEXT: MOV T0.Y, PV.X, | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T0.W, PV.W, PS, | |
| ; EG-NEXT: MOV T4.X, PV.W, | |
| ; EG-NEXT: MOV T0.Y, T5.X, | |
| ; EG-NEXT: MOV * T0.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T0.W, T0.X, literal.x, PV.W, | |
| ; EG-NEXT: AND_INT * T1.W, PV.Y, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T5.X, PV.W, | |
| ; EG-NEXT: MOV T0.Y, PV.X, | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: LSHR T0.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: OR_INT * T8.Y, PV.W, PS, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: MOV T5.X, PV.Y, | |
| ; EG-NEXT: MOV * T8.X, T4.X, | |
| ; | |
| ; CM-LABEL: global_zextload_v4i8_to_v4i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 3, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 31, @12, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T8, T0.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV T0.Y, T4.X, | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 12: | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 255(3.573311e-43), -65536(nan) | |
| ; CM-NEXT: OR_INT * T0.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T4.X, PV.W, | |
| ; CM-NEXT: MOV T0.Y, PV.X, | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T4.X, PV.W, | |
| ; CM-NEXT: MOV T0.Y, T5.X, | |
| ; CM-NEXT: MOV * T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T0.W, T0.X, literal.y, PV.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T5.X, PV.W, | |
| ; CM-NEXT: MOV T0.Y, PV.X, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: LSHR T0.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: OR_INT * T8.Y, PV.Z, PV.W, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: MOV * T5.X, PV.Y, | |
| ; CM-NEXT: MOV * T8.X, T4.X, | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <4 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <4 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <4 x i8> %load to <4 x i16> | |
| store <4 x i16> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v4i8_to_v4i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v4i8_to_v4i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dword v0, v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v1, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v0, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v3, v0, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v1, v2, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v0, v0, v3 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v4i8_to_v4i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dword v2, v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v3, 24, v2 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v2, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v2, v2, 0, 8 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v3, 16, v3 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v4, 0xffff, v4 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v5, 16, v5 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v2, 0xffff, v2 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v3, v4, v3 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v2, v2, v5 | |
| ; GCN-HSA-NEXT: flat_store_dwordx2 v[0:1], v[2:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v4i8_to_v4i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, 8 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dword v0, v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 0xffff | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 16, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v4, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v5, v1, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v0, v2, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v2, 16, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v1, v1, sext(v3) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v1, v1, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v0, v5, v0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v4i8_to_v4i16: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 3, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 37, @12, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T8.XY, T0.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.Y, T4.X, | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 12: | |
| ; EG-NEXT: BFE_INT * T0.W, T0.X, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T0.W, PV.W, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.Y, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T4.X, PV.W, | |
| ; EG-NEXT: MOV T0.Y, PV.X, | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T0.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, T1.W, PV.W, | |
| ; EG-NEXT: MOV T4.X, PV.W, | |
| ; EG-NEXT: MOV T0.Y, T5.X, | |
| ; EG-NEXT: LSHR * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T0.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, T1.W, PV.W, | |
| ; EG-NEXT: MOV * T5.X, PV.W, | |
| ; EG-NEXT: MOV T0.Y, PV.X, | |
| ; EG-NEXT: ASHR * T0.W, T0.X, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, PV.Y, literal.x, | |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: LSHR T0.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: OR_INT * T8.Y, PV.W, PS, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: MOV T5.X, PV.Y, | |
| ; EG-NEXT: MOV * T8.X, T4.X, | |
| ; | |
| ; CM-LABEL: global_sextload_v4i8_to_v4i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 3, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 37, @12, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T8, T0.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV T0.Y, T4.X, | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 12: | |
| ; CM-NEXT: BFE_INT * T0.W, T0.X, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.W, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; CM-NEXT: OR_INT * T0.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T4.X, PV.W, | |
| ; CM-NEXT: MOV T0.Y, PV.X, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T0.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.Y, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T4.X, PV.W, | |
| ; CM-NEXT: MOV T0.Y, T5.X, | |
| ; CM-NEXT: LSHR * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T0.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T5.X, PV.W, | |
| ; CM-NEXT: MOV T0.Y, PV.X, | |
| ; CM-NEXT: ASHR * T0.W, T0.X, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.Y, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: LSHR T0.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: OR_INT * T8.Y, PV.Z, PV.W, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: MOV * T5.X, PV.Y, | |
| ; CM-NEXT: MOV * T8.X, T4.X, | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <4 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <4 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <4 x i8> %load to <4 x i16> | |
| store <4 x i16> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v8i8_to_v8i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v8i8_to_v8i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v2, 0xff00, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v3, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v4, 0xff00, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v5, 24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v6, 0xff, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v7, 0xff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v1, v5, v1, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v4, 8, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v0, v3, v0, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v5, 8, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v3, 0xff00ff, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v2, v6, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v1, 0xff00ff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v0, v7, v5 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v8i8_to_v8i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx2 v[0:1], v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v2, 0xff00, v0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v3, 24, v0 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v6, 0xff00, v1 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v7, 24, v1 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v8, 0xff, v1 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v9, 0xff, v0 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v1, v7, v1, 16 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v6, 8, v6 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v0, v3, v0, 16 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v7, 8, v2 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v3, 0xff00ff, v1 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v2, v8, v6 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v1, 0xff00ff, v0 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v0, v9, v7 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v8i8_to_v8i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, 0x5040c00 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx2 v[1:2], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v4, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v0, 8, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v5, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 16, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v6, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v0, v0, v1, s4 | |
| ; GCN-NOHSA-VI-NEXT: v_alignbit_b32 v1, v4, v1, 16 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v3, v6, v3, s4 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v2, v5, v2, s4 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v1, 0xff00ff, v1 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v8i8_to_v8i16: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 3, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 60, @12, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T12.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: PAD | |
| ; EG-NEXT: Fetch clause starting at 6: | |
| ; EG-NEXT: VTX_READ_64 T12.XY, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 8: | |
| ; EG-NEXT: MOV T0.Y, T8.X, | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 12: | |
| ; EG-NEXT: AND_INT T0.W, T12.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.Y, literal.y, | |
| ; EG-NEXT: 255(3.573311e-43), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T8.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHL * T0.W, T12.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T0.W, PV.W, PS, | |
| ; EG-NEXT: MOV T8.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T9.X, | |
| ; EG-NEXT: MOV * T0.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T1.W, T12.X, literal.x, PV.W, | |
| ; EG-NEXT: AND_INT * T2.W, T0.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), -65536(nan) | |
| ; EG-NEXT: OR_INT * T1.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T9.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T12.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T0.Y, PV.W, PS, | |
| ; EG-NEXT: MOV T9.X, PV.Y, | |
| ; EG-NEXT: MOV * T0.X, T4.X, | |
| ; EG-NEXT: AND_INT T1.W, PS, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, T12.Y, literal.y, | |
| ; EG-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T4.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHL * T1.W, T12.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV T4.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T5.X, | |
| ; EG-NEXT: BFE_UINT T0.W, T12.Y, literal.x, T0.W, | |
| ; EG-NEXT: AND_INT * T1.W, PS, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T5.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T0.W, T12.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: LSHR T12.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: OR_INT * T0.W, PV.W, PS, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: MOV T5.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T8.X, | |
| ; EG-NEXT: MOV * T0.Z, T4.X, | |
| ; | |
| ; CM-LABEL: global_zextload_v8i8_to_v8i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 3, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 60, @12, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T12.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_64 T12.XY, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV T0.Y, T8.X, | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 12: | |
| ; CM-NEXT: AND_INT T0.Z, T12.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 255(3.573311e-43), -65536(nan) | |
| ; CM-NEXT: OR_INT * T0.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T8.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHL * T0.W, T12.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T8.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T9.X, | |
| ; CM-NEXT: MOV * T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T1.W, T12.X, literal.y, PV.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T9.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T12.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T0.Y, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T9.X, PV.Y, | |
| ; CM-NEXT: MOV * T0.X, T4.X, | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, T12.Y, literal.y, | |
| ; CM-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T4.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHL * T1.W, T12.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T4.X, PV.W, | |
| ; CM-NEXT: MOV * T0.X, T5.X, | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T0.W, T12.Y, literal.y, T0.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T5.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T0.W, T12.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: LSHR T12.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: MOV * T5.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T8.X, | |
| ; CM-NEXT: MOV * T0.Z, T4.X, BS:VEC_120/SCL_212 | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <8 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <8 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <8 x i8> %load to <8 x i16> | |
| store <8 x i16> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v8i8_to_v8i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v8i8_to_v8i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v2, 24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v3, v1, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v4, v1, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v1, v1, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v5, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v6, v0, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v7, v0, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v3, 0xffff, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v6, 0xffff, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v3, v3, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v2, v1, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v1, v6, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v0, v0, v7 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v8i8_to_v8i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx2 v[0:1], v[0:1] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v2, 24, v1 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v3, v1, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v1, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v1, v1, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 24, v0 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v8, v0, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v9, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v2, 16, v2 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v3, 0xffff, v3 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v6, 16, v6 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v1, 0xffff, v1 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v7, 16, v7 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v8, 0xffff, v8 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v9, 16, v9 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v0, 0xffff, v0 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v3, v3, v2 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v2, v1, v6 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v1, v8, v7 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v0, v0, v9 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v8i8_to_v8i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, 0xffff | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, 8 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx2 v[0:1], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v4, 16, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v5, 16, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v6, v3, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v7, v2, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v8, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v9, v3, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v10, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v0, v6, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v6, 16, v8 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v5, v3, sext(v5) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v2, v9, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v1, 16, v10 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v4, v3, sext(v4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v3, v5, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v1, v4, v1 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v8i8_to_v8i16: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 3, @8, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @6 | |
| ; EG-NEXT: ALU 74, @12, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T12.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: PAD | |
| ; EG-NEXT: Fetch clause starting at 6: | |
| ; EG-NEXT: VTX_READ_64 T12.XY, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 8: | |
| ; EG-NEXT: MOV T0.Y, T8.X, | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 12: | |
| ; EG-NEXT: BFE_INT * T0.W, T12.X, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T0.W, PV.W, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.Y, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T8.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T0.W, T12.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T0.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, T1.W, PV.W, | |
| ; EG-NEXT: MOV T8.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T9.X, | |
| ; EG-NEXT: LSHR * T0.W, T12.X, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T0.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, T1.W, PV.W, | |
| ; EG-NEXT: MOV * T9.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: ASHR * T0.W, T12.X, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, PV.X, literal.x, | |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: OR_INT * T0.Y, PV.W, PS, | |
| ; EG-NEXT: MOV T9.X, PV.Y, | |
| ; EG-NEXT: MOV * T0.X, T4.X, | |
| ; EG-NEXT: BFE_INT T0.W, T12.Y, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PS, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, T1.W, PV.W, | |
| ; EG-NEXT: MOV * T4.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T0.W, T12.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T0.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, T1.W, PV.W, | |
| ; EG-NEXT: MOV T4.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T5.X, | |
| ; EG-NEXT: LSHR * T0.W, T12.Y, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T0.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, T1.W, PV.W, | |
| ; EG-NEXT: MOV * T5.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: ASHR * T0.W, T12.Y, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, PV.X, literal.x, | |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: LSHR T12.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: OR_INT * T0.W, PV.W, PS, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: MOV T5.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T8.X, | |
| ; EG-NEXT: MOV * T0.Z, T4.X, | |
| ; | |
| ; CM-LABEL: global_sextload_v8i8_to_v8i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 3, @8, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @6 | |
| ; CM-NEXT: ALU 74, @12, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T12.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 6: | |
| ; CM-NEXT: VTX_READ_64 T12.XY, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 8: | |
| ; CM-NEXT: MOV T0.Y, T8.X, | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 3(4.203895e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 12: | |
| ; CM-NEXT: BFE_INT * T0.W, T12.X, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.W, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; CM-NEXT: OR_INT * T0.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T8.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T0.W, T12.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T0.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T8.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T9.X, | |
| ; CM-NEXT: LSHR * T0.W, T12.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T0.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T9.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: ASHR * T0.W, T12.X, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T0.Y, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T9.X, PV.Y, | |
| ; CM-NEXT: MOV T0.X, T4.X, | |
| ; CM-NEXT: BFE_INT * T0.W, T12.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T4.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T0.W, T12.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T0.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T4.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T5.X, | |
| ; CM-NEXT: LSHR * T0.W, T12.Y, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T0.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T5.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: ASHR * T0.W, T12.Y, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: LSHR T12.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: MOV * T5.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T8.X, | |
| ; CM-NEXT: MOV * T0.Z, T4.X, BS:VEC_120/SCL_212 | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <8 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <8 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <8 x i8> %load to <8 x i16> | |
| store <8 x i16> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v16i8_to_v16i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v16i8_to_v16i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[0:3], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v4, 0xff00, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v5, 24, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v6, 0xff00, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v7, 24, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v8, 0xff00, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v9, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v10, 0xff00, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v11, 24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v14, 0xff, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v15, 0xff, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v7, v7, v3, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v6, 8, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v5, v5, v2, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v4, 8, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v12, 0xff, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v13, 0xff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v1, v11, v1, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v10, 8, v10 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v0, v9, v0, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v8, 8, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v7, 0xff00ff, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v6, v14, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v5, 0xff00ff, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v4, v15, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v3, 0xff00ff, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v2, v12, v10 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v1, 0xff00ff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v0, v13, v8 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v16i8_to_v16i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v11, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v10, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, s0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v4, 0xff00, v2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v5, 24, v2 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v6, 0xff00, v3 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v7, 24, v3 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v12, 0xff00, v0 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v13, 24, v0 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v14, 0xff00, v1 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v15, 24, v1 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v18, 0xff, v3 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v19, 0xff, v2 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v7, v7, v3, 16 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v6, 8, v6 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v5, v5, v2, 16 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v4, 8, v4 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v16, 0xff, v1 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v17, 0xff, v0 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v1, v15, v1, 16 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v14, 8, v14 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v0, v13, v0, 16 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v12, 8, v12 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v7, 0xff00ff, v7 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v6, v18, v6 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v5, 0xff00ff, v5 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v4, v19, v4 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v3, 0xff00ff, v1 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v2, v16, v14 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v1, 0xff00ff, v0 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v0, v17, v12 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[10:11], v[4:7] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[8:9], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v16i8_to_v16i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, 0x5040c00 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v5, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v6, v1, 8, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v8, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v9, v3, 8, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v10, 16, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v11, 24, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v12, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v4, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v7, 16, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v13, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v6, v6, v1, s4 | |
| ; GCN-NOHSA-VI-NEXT: v_alignbit_b32 v1, v5, v2, 16 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v4, v4, v0, s4 | |
| ; GCN-NOHSA-VI-NEXT: v_alignbit_b32 v0, v12, v0, 16 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v11, v11, v10, s4 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v10, v9, v3, s4 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v8, v8, v2, s4 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v9, 0xff00ff, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v7, v13, v7, s4 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v5, 0xff00ff, v0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v16i8_to_v16i16: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 3, @10, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @8 | |
| ; EG-NEXT: ALU 101, @14, KC0[], KC1[] | |
| ; EG-NEXT: ALU 18, @116, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T21.XYZW, T20.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T22.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: PAD | |
| ; EG-NEXT: Fetch clause starting at 8: | |
| ; EG-NEXT: VTX_READ_128 T20.XYZW, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 10: | |
| ; EG-NEXT: MOV T0.Y, T16.X, | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 14: | |
| ; EG-NEXT: AND_INT T0.W, T20.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.Y, literal.y, | |
| ; EG-NEXT: 255(3.573311e-43), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T16.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHL * T0.W, T20.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T0.W, PV.W, PS, | |
| ; EG-NEXT: MOV T16.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T17.X, | |
| ; EG-NEXT: MOV * T0.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T1.W, T20.X, literal.x, PV.W, | |
| ; EG-NEXT: AND_INT * T2.W, T0.X, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), -65536(nan) | |
| ; EG-NEXT: OR_INT * T1.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T17.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T20.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T21.Y, PV.W, PS, | |
| ; EG-NEXT: MOV T17.X, PV.Y, | |
| ; EG-NEXT: MOV * T0.X, T12.X, | |
| ; EG-NEXT: AND_INT T1.W, PS, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, T20.Y, literal.y, | |
| ; EG-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T12.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHL * T1.W, T20.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV T12.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T13.X, | |
| ; EG-NEXT: BFE_UINT T1.W, T20.Y, literal.x, T0.W, | |
| ; EG-NEXT: AND_INT * T2.W, PS, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), -65536(nan) | |
| ; EG-NEXT: OR_INT * T1.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T13.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T20.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T21.W, PV.W, PS, | |
| ; EG-NEXT: MOV T13.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T8.X, | |
| ; EG-NEXT: AND_INT T1.W, PS, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, T20.Z, literal.y, | |
| ; EG-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T8.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHL * T1.W, T20.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV T8.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T9.X, | |
| ; EG-NEXT: BFE_UINT T1.W, T20.Z, literal.x, T0.W, | |
| ; EG-NEXT: AND_INT * T2.W, PS, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), -65536(nan) | |
| ; EG-NEXT: OR_INT * T1.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T9.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T20.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T0.Y, PV.W, PS, | |
| ; EG-NEXT: MOV T9.X, PV.Y, | |
| ; EG-NEXT: MOV * T0.X, T4.X, | |
| ; EG-NEXT: AND_INT T1.W, PS, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, T20.W, literal.y, | |
| ; EG-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T4.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHL * T1.W, T20.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV T4.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T5.X, | |
| ; EG-NEXT: BFE_UINT * T0.W, T20.W, literal.x, T0.W, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ALU clause starting at 116: | |
| ; EG-NEXT: AND_INT * T1.W, T0.X, literal.x, | |
| ; EG-NEXT: -65536(nan), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, PV.W, T0.W, | |
| ; EG-NEXT: MOV * T5.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR T0.W, T20.W, literal.x, | |
| ; EG-NEXT: LSHR * T20.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 2(2.802597e-45) | |
| ; EG-NEXT: AND_INT T1.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: ADD_INT T22.X, T20.X, literal.x, | |
| ; EG-NEXT: OR_INT * T0.W, PV.W, PS, | |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; EG-NEXT: MOV T5.X, PV.W, | |
| ; EG-NEXT: MOV * T21.X, T16.X, | |
| ; EG-NEXT: MOV * T21.Z, T12.X, | |
| ; EG-NEXT: MOV T0.X, T8.X, | |
| ; EG-NEXT: MOV * T0.Z, T4.X, BS:VEC_120/SCL_212 | |
| ; | |
| ; CM-LABEL: global_zextload_v16i8_to_v16i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 3, @10, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @8 | |
| ; CM-NEXT: ALU 101, @14, KC0[], KC1[] | |
| ; CM-NEXT: ALU 19, @116, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T22.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T21, T20.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 8: | |
| ; CM-NEXT: VTX_READ_128 T20.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 10: | |
| ; CM-NEXT: MOV T0.Y, T16.X, | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 14: | |
| ; CM-NEXT: AND_INT T0.Z, T20.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 255(3.573311e-43), -65536(nan) | |
| ; CM-NEXT: OR_INT * T0.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T16.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHL * T0.W, T20.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T16.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T17.X, | |
| ; CM-NEXT: MOV * T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T1.W, T20.X, literal.y, PV.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T17.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T20.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T21.Y, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T17.X, PV.Y, | |
| ; CM-NEXT: MOV * T0.X, T12.X, | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, T20.Y, literal.y, | |
| ; CM-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T12.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHL * T1.W, T20.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T12.X, PV.W, | |
| ; CM-NEXT: MOV * T0.X, T13.X, | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T1.W, T20.Y, literal.y, T0.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T13.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T20.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T21.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T13.X, PV.W, | |
| ; CM-NEXT: MOV * T0.X, T8.X, | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, T20.Z, literal.y, | |
| ; CM-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T8.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHL * T1.W, T20.Z, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T8.X, PV.W, | |
| ; CM-NEXT: MOV * T0.X, T9.X, | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T1.W, T20.Z, literal.y, T0.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T9.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T20.Z, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T0.Y, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T9.X, PV.Y, | |
| ; CM-NEXT: MOV * T0.X, T4.X, | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, T20.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T4.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHL * T1.W, T20.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T4.X, PV.W, | |
| ; CM-NEXT: MOV * T0.X, T5.X, | |
| ; CM-NEXT: AND_INT * T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: -65536(nan), 0(0.000000e+00) | |
| ; CM-NEXT: ALU clause starting at 116: | |
| ; CM-NEXT: BFE_UINT * T0.W, T20.W, literal.x, T0.W, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: OR_INT * T0.W, T0.Z, PV.W, | |
| ; CM-NEXT: MOV * T5.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T0.W, T20.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR T20.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.y, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.z, | |
| ; CM-NEXT: 2(2.802597e-45), 65535(9.183409e-41) | |
| ; CM-NEXT: 16711680(2.341805e-38), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T22.X, PV.X, literal.x, | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; CM-NEXT: MOV * T5.X, PV.W, | |
| ; CM-NEXT: MOV T21.X, T16.X, | |
| ; CM-NEXT: MOV * T21.Z, T12.X, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: MOV T0.X, T8.X, | |
| ; CM-NEXT: MOV * T0.Z, T4.X, BS:VEC_120/SCL_212 | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <16 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <16 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <16 x i8> %load to <16 x i16> | |
| store <16 x i16> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v16i8_to_v16i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v16i8_to_v16i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[0:3], v[0:1], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v4, 24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v5, v1, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v6, v1, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v1, v1, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v7, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v8, v0, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v10, 24, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v11, v3, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v12, v3, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v3, v3, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v13, 24, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v14, v2, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v15, v2, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v2, v2, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v9, v0, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v5, 0xffff, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v8, 0xffff, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v11, 0xffff, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v16, 0xffff, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v14, 0xffff, v14 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v17, 0xffff, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v3, v5, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v2, v1, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v1, v8, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v7, v11, v10 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v6, v16, v12 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v5, v14, v13 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v4, v17, v15 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v0, v0, v9 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v16i8_to_v16i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v0, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v11, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v10, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v9, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v8, s0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v4, 24, v1 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v5, v1, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v6, v1, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v1, v1, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v7, 24, v0 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v12, v0, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v14, 24, v3 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v15, v3, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v16, v3, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v17, v3, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v18, 24, v2 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v19, v2, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v20, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v21, v2, 0, 8 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v2, 16, v4 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v3, 0xffff, v5 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v4, 16, v6 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v1, 0xffff, v1 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v3, v3, v2 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v2, v1, v4 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v1, 16, v7 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v4, 0xffff, v12 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v5, 16, v13 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v0, 0xffff, v0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v6, 16, v14 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v7, 0xffff, v15 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v12, 16, v16 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v13, 0xffff, v17 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v14, 16, v18 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v15, 0xffff, v19 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v16, 16, v20 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v17, 0xffff, v21 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v1, v4, v1 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v0, v0, v5 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v7, v7, v6 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v6, v13, v12 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v5, v15, v14 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v4, v17, v16 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[10:11], v[4:7] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[8:9], v[0:3] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v16i8_to_v16i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 4, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, 0xffff | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v7, 8 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[0:1] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v8, 16, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v4, 16, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v6, 16, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v9, 16, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v10, v5, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v1, v7, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v12, 24, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v11, v5, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v0, v7, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v13, v5, sext(v3) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v3, v7, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v14, v5, sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v15, v7, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v2, v10, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v1, v5, sext(v9) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v10, v5, sext(v6) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v12, 16, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v16, v5, sext(v4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v5, v5, sext(v8) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v8, v7, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v0, v11, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v9, v7, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v11, v7, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v6, v13, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v4, v14, v15 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v7, v16, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v5, v5, v8 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v3, v1, v9 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v1, v10, v11 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v16i8_to_v16i16: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 3, @10, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 0 @8 | |
| ; EG-NEXT: ALU 104, @14, KC0[], KC1[] | |
| ; EG-NEXT: ALU 44, @119, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T21.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T22.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: PAD | |
| ; EG-NEXT: Fetch clause starting at 8: | |
| ; EG-NEXT: VTX_READ_128 T20.XYZW, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 10: | |
| ; EG-NEXT: MOV T0.Y, T16.X, | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 14: | |
| ; EG-NEXT: BFE_INT * T0.W, T20.X, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T0.W, PV.W, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.Y, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; EG-NEXT: OR_INT * T0.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T16.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T0.W, T20.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T0.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, T1.W, PV.W, | |
| ; EG-NEXT: MOV T16.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T17.X, | |
| ; EG-NEXT: LSHR * T0.W, T20.X, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T0.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, T1.W, PV.W, | |
| ; EG-NEXT: MOV * T17.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: ASHR * T0.W, T20.X, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, PV.X, literal.x, | |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: OR_INT * T0.Y, PV.W, PS, | |
| ; EG-NEXT: MOV T17.X, PV.Y, | |
| ; EG-NEXT: MOV * T0.X, T12.X, | |
| ; EG-NEXT: BFE_INT T0.W, T20.Y, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PS, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, T1.W, PV.W, | |
| ; EG-NEXT: MOV * T12.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T0.W, T20.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T0.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, T1.W, PV.W, | |
| ; EG-NEXT: MOV T12.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T13.X, | |
| ; EG-NEXT: LSHR * T0.W, T20.Y, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T0.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T0.W, T1.W, PV.W, | |
| ; EG-NEXT: MOV * T13.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: ASHR * T0.W, T20.Y, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, PV.X, literal.x, | |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: OR_INT * T0.W, PV.W, PS, | |
| ; EG-NEXT: MOV T13.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T8.X, | |
| ; EG-NEXT: BFE_INT T1.W, T20.Z, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PS, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T8.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T20.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV T8.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T9.X, | |
| ; EG-NEXT: LSHR * T1.W, T20.Z, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T9.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: ASHR * T1.W, T20.Z, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.X, literal.x, | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: ALU clause starting at 119: | |
| ; EG-NEXT: OR_INT * T20.Y, T2.W, T1.W, | |
| ; EG-NEXT: MOV T9.X, PV.Y, | |
| ; EG-NEXT: MOV * T0.X, T4.X, | |
| ; EG-NEXT: BFE_INT T1.W, T20.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PS, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T4.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T20.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV T4.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T5.X, | |
| ; EG-NEXT: LSHR * T1.W, T20.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T5.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: ASHR T1.W, T20.W, literal.x, | |
| ; EG-NEXT: LSHR * T21.X, KC0[2].Y, literal.y, | |
| ; EG-NEXT: 24(3.363116e-44), 2(2.802597e-45) | |
| ; EG-NEXT: AND_INT T2.W, PV.X, literal.x, | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: ADD_INT T22.X, T21.X, literal.x, | |
| ; EG-NEXT: OR_INT * T20.W, PV.W, PS, | |
| ; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; EG-NEXT: MOV T5.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T16.X, | |
| ; EG-NEXT: MOV * T0.Z, T12.X, | |
| ; EG-NEXT: MOV T20.X, T8.X, | |
| ; EG-NEXT: MOV * T20.Z, T4.X, BS:VEC_120/SCL_212 | |
| ; | |
| ; CM-LABEL: global_sextload_v16i8_to_v16i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 3, @10, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 0 @8 | |
| ; CM-NEXT: ALU 104, @14, KC0[], KC1[] | |
| ; CM-NEXT: ALU 45, @119, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T20, T22.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T21.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: PAD | |
| ; CM-NEXT: Fetch clause starting at 8: | |
| ; CM-NEXT: VTX_READ_128 T20.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 10: | |
| ; CM-NEXT: MOV T0.Y, T16.X, | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 14: | |
| ; CM-NEXT: BFE_INT * T0.W, T20.X, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.W, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, T0.Y, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; CM-NEXT: OR_INT * T0.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T16.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T0.W, T20.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T0.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T16.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T17.X, | |
| ; CM-NEXT: LSHR * T0.W, T20.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T0.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T17.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: ASHR * T0.W, T20.X, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T0.Y, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T17.X, PV.Y, | |
| ; CM-NEXT: MOV T0.X, T12.X, | |
| ; CM-NEXT: BFE_INT * T0.W, T20.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T12.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T0.W, T20.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T0.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T12.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T13.X, | |
| ; CM-NEXT: LSHR * T0.W, T20.Y, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T0.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T13.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: ASHR * T0.W, T20.Y, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T13.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T8.X, | |
| ; CM-NEXT: BFE_INT * T1.W, T20.Z, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T8.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T20.Z, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T8.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T9.X, | |
| ; CM-NEXT: LSHR * T1.W, T20.Z, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T9.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: ASHR * T1.W, T20.Z, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: ALU clause starting at 119: | |
| ; CM-NEXT: OR_INT * T20.Y, T0.Z, T1.W, | |
| ; CM-NEXT: MOV * T9.X, PV.Y, | |
| ; CM-NEXT: MOV T0.X, T4.X, | |
| ; CM-NEXT: BFE_INT * T1.W, T20.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T4.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T20.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T4.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T5.X, | |
| ; CM-NEXT: LSHR * T1.W, T20.W, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T5.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: ASHR * T1.W, T20.W, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: LSHR T21.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.y, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.z, | |
| ; CM-NEXT: 2(2.802597e-45), 65535(9.183409e-41) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T22.X, PV.X, literal.x, | |
| ; CM-NEXT: OR_INT * T20.W, PV.Z, PV.W, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; CM-NEXT: MOV * T5.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T16.X, | |
| ; CM-NEXT: MOV * T0.Z, T12.X, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: MOV T20.X, T8.X, | |
| ; CM-NEXT: MOV * T20.Z, T4.X, BS:VEC_120/SCL_212 | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <16 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <16 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <16 x i8> %load to <16 x i16> | |
| store <16 x i16> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_zextload_v32i8_to_v32i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_zextload_v32i8_to_v32i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v7, 5, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v8, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v4, 0xff00, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v9, 0xff, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v4, 8, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v4, v9, v4 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[8:11], v[7:8], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v5, 0xff00, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v6, 0xff, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v5, 8, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v6, v6, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v5, 24, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v3, v5, v3, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v7, 0xff00ff, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v3, 24, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v2, v3, v2, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v5, 0xff00ff, v2 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v15, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v14, 24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v15, v15, v0, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v16, 0xff00, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v17, 0xff00, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v18, 0xff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v14, v14, v1, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v1, 0xff, v1 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v2, 24, v10 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v4, 24, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v6, 24, v9 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v3, 0xff00, v10 | |
| ; GCN-NOHSA-SI-NEXT: v_lshrrev_b32_e32 v5, 24, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v7, v6, v9, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v12, v4, v11, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v13, v2, v10, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v4, 0xff00, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v2, 0xff00, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v6, 0xff00, v9 | |
| ; GCN-NOHSA-SI-NEXT: v_alignbit_b32 v5, v5, v8, 16 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v9, 0xff, v9 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v8, 0xff, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v11, 0xff, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v10, 0xff, v10 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v0, 8, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v6, 8, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v4, 8, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v3, 8, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v2, v9, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v0, v8, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v6, v11, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v4, v10, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v3, 8, v17 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v8, 8, v16 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v10, v1, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v8, v18, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v11, 0xff00ff, v14 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v9, 0xff00ff, v15 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v3, 0xff00ff, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v1, 0xff00ff, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v7, 0xff00ff, v12 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v5, 0xff00ff, v13 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_zextload_v32i8_to_v32i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v4, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[4:5] | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v4, vcc, 16, v4 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[4:7], v[4:5] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 48 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v8, 0xff00, v2 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v9, 24, v2 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v10, 0xff00, v3 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v11, 24, v3 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v16, 0xff, v3 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v17, 0xff, v2 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v10, 8, v10 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v8, 8, v8 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v3, v11, v3, 16 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v2, v9, v2, 16 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v15, 24, v0 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v10, v16, v10 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v8, v17, v8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v16, 0xff00, v1 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v17, 24, v1 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v11, 0xff00ff, v3 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v9, 0xff00ff, v2 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v2, 24, v6 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v14, 0xff00, v0 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v3, 0xff, v1 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v18, 0xff, v0 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[12:13], v[8:11] | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v12, 24, v4 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v1, v17, v1, 16 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v0, v15, v0, 16 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v13, 24, v5 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v21, v2, v6, 16 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v2, 8, v16 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v8, 0xff00, v6 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v9, 0xff00, v7 | |
| ; GCN-HSA-NEXT: v_lshrrev_b32_e32 v10, 24, v7 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v11, 0xff00, v4 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v15, 0xff00, v5 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v17, 0xff, v5 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v19, 0xff, v4 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v5, v13, v5, 16 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v4, v12, v4, 16 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v2, v3, v2 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v3, 0xff00ff, v1 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v1, 0xff00ff, v0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v13, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v12, s0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 8, v14 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 32 | |
| ; GCN-HSA-NEXT: v_alignbit_b32 v20, v10, v7, 16 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v10, 0xff, v7 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v22, 0xff, v6 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v6, 8, v15 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v0, v18, v0 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v7, 0xff00ff, v5 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v5, 0xff00ff, v4 | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v4, 8, v11 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v9, 8, v9 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v8, 8, v8 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v6, v17, v6 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v15, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v14, s2 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v4, v19, v4 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v10, v10, v9 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v11, 0xff00ff, v20 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v9, 0xff00ff, v21 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s0 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v8, v22, v8 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[12:13], v[0:3] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[14:15], v[8:11] | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[4:7] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_zextload_v32i8_to_v32i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v4, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[4:5] | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v4, vcc, 16, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[4:7], v[4:5] | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, 0x5040c00 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v9, 24, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v12, v3, 8, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v13, 16, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v15, 24, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v8, 8, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v14, v12, v3, s2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 24, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_alignbit_b32 v20, v9, v2, 16 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v9, 8, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v11, 8, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v8, v8, v0, s2 | |
| ; GCN-NOHSA-VI-NEXT: v_alignbit_b32 v17, v3, v0, 16 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b16_e32 v0, 8, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v16, v9, v6, s2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v9, 24, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v10, v1, 8, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v12, v11, v2, s2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v2, v5, 8, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v0, v0, v4, s2 | |
| ; GCN-NOHSA-VI-NEXT: v_bfe_u32 v11, v7, 8, 8 | |
| ; GCN-NOHSA-VI-NEXT: v_alignbit_b32 v4, v9, v4, 16 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v9, 24, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v10, v10, v1, s2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 16, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v1, 24, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v2, v2, v5, s2 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v18, v11, v7, s2 | |
| ; GCN-NOHSA-VI-NEXT: v_alignbit_b32 v6, v9, v6, 16 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v9, 16, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v7, 24, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v19, 16, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v5, 24, v5 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v11, v1, v3, s2 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v15, v15, v13, s2 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v3, v5, v19, s2 | |
| ; GCN-NOHSA-VI-NEXT: v_perm_b32 v19, v7, v9, s2 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v13, 0xff00ff, v20 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v9, 0xff00ff, v17 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v1, 0xff00ff, v4 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v17, 0xff00ff, v6 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_zextload_v32i8_to_v32i16: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 3, @14, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 1 @10 | |
| ; EG-NEXT: ALU 103, @18, KC0[], KC1[] | |
| ; EG-NEXT: ALU 102, @122, KC0[], KC1[] | |
| ; EG-NEXT: ALU 36, @225, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T36.XYZW, T42.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T37.XYZW, T41.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T38.XYZW, T39.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T40.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: Fetch clause starting at 10: | |
| ; EG-NEXT: VTX_READ_128 T37.XYZW, T0.X, 16, #1 | |
| ; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 14: | |
| ; EG-NEXT: MOV T1.Y, T16.X, | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 5(7.006492e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 18: | |
| ; EG-NEXT: AND_INT T1.W, T37.X, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, T1.Y, literal.y, | |
| ; EG-NEXT: 255(3.573311e-43), -65536(nan) | |
| ; EG-NEXT: OR_INT * T1.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T16.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHL * T1.W, T37.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV T16.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, T17.X, | |
| ; EG-NEXT: MOV * T1.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_UINT T2.W, T37.X, literal.x, PV.W, | |
| ; EG-NEXT: AND_INT * T3.W, PV.Y, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), -65536(nan) | |
| ; EG-NEXT: OR_INT * T2.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T17.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHR * T2.W, T37.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T36.Y, PV.W, PS, | |
| ; EG-NEXT: MOV T17.X, PV.Y, | |
| ; EG-NEXT: MOV * T1.Y, T12.X, | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T3.W, T37.Y, literal.y, | |
| ; EG-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T12.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHL * T2.W, T37.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV T12.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, T13.X, | |
| ; EG-NEXT: BFE_UINT * T2.W, T37.Y, literal.x, T1.W, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT * T3.W, PV.Y, literal.x, | |
| ; EG-NEXT: -65536(nan), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, T2.W, | |
| ; EG-NEXT: MOV * T13.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHR * T2.W, T37.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T36.W, PV.W, PS, | |
| ; EG-NEXT: MOV T13.X, PV.W, | |
| ; EG-NEXT: MOV * T1.Y, T8.X, | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T3.W, T37.Z, literal.y, | |
| ; EG-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T8.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHL * T2.W, T37.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV T8.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, T9.X, | |
| ; EG-NEXT: BFE_UINT * T2.W, T37.Z, literal.x, T1.W, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT * T3.W, PV.Y, literal.x, | |
| ; EG-NEXT: -65536(nan), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, T2.W, | |
| ; EG-NEXT: MOV * T9.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHR * T2.W, T37.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T37.Y, PV.W, PS, | |
| ; EG-NEXT: MOV T9.X, PV.Y, | |
| ; EG-NEXT: MOV * T1.Y, T4.X, | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T3.W, T37.W, literal.y, | |
| ; EG-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T4.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHL * T2.W, T37.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV T4.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, T5.X, | |
| ; EG-NEXT: BFE_UINT * T2.W, T37.W, literal.x, T1.W, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ALU clause starting at 122: | |
| ; EG-NEXT: AND_INT * T3.W, T1.Y, literal.x, | |
| ; EG-NEXT: -65536(nan), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, T2.W, | |
| ; EG-NEXT: MOV * T5.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHR * T2.W, T37.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T37.W, PV.W, PS, | |
| ; EG-NEXT: MOV T5.X, PV.W, | |
| ; EG-NEXT: MOV * T1.Y, T32.X, | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T3.W, T0.X, literal.y, | |
| ; EG-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T32.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHL * T2.W, T0.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV T32.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, T33.X, | |
| ; EG-NEXT: BFE_UINT * T2.W, T0.X, literal.x, T1.W, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT * T3.W, PV.Y, literal.x, | |
| ; EG-NEXT: -65536(nan), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, T2.W, | |
| ; EG-NEXT: MOV * T33.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHR * T2.W, T0.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T38.Y, PV.W, PS, | |
| ; EG-NEXT: MOV T33.X, PV.Y, | |
| ; EG-NEXT: MOV * T0.X, T28.X, | |
| ; EG-NEXT: AND_INT T2.W, PS, literal.x, | |
| ; EG-NEXT: AND_INT * T3.W, T0.Y, literal.y, | |
| ; EG-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T28.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHL * T2.W, T0.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV T28.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T29.X, | |
| ; EG-NEXT: BFE_UINT T2.W, T0.Y, literal.x, T1.W, | |
| ; EG-NEXT: AND_INT * T3.W, PS, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), -65536(nan) | |
| ; EG-NEXT: OR_INT * T2.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T29.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T2.W, T0.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T38.W, PV.W, PS, | |
| ; EG-NEXT: MOV T29.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T24.X, | |
| ; EG-NEXT: AND_INT T2.W, PS, literal.x, | |
| ; EG-NEXT: AND_INT * T3.W, T0.Z, literal.y, | |
| ; EG-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T24.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHL * T2.W, T0.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV T24.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T25.X, | |
| ; EG-NEXT: BFE_UINT T2.W, T0.Z, literal.x, T1.W, | |
| ; EG-NEXT: AND_INT * T3.W, PS, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), -65536(nan) | |
| ; EG-NEXT: OR_INT * T2.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T25.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T2.W, T0.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T0.Y, PV.W, PS, | |
| ; EG-NEXT: MOV T25.X, PV.Y, | |
| ; EG-NEXT: MOV * T0.X, T20.X, | |
| ; EG-NEXT: AND_INT T2.W, PS, literal.x, | |
| ; EG-NEXT: AND_INT * T3.W, T0.W, literal.y, | |
| ; EG-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T20.X, PV.W, | |
| ; EG-NEXT: ALU clause starting at 225: | |
| ; EG-NEXT: MOV T0.X, T20.X, | |
| ; EG-NEXT: LSHL * T2.W, T0.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T3.W, PV.X, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: OR_INT * T2.W, PV.W, PS, | |
| ; EG-NEXT: MOV T20.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T21.X, | |
| ; EG-NEXT: BFE_UINT T1.W, T0.W, literal.x, T1.W, | |
| ; EG-NEXT: AND_INT * T2.W, PS, literal.y, | |
| ; EG-NEXT: 16(2.242078e-44), -65536(nan) | |
| ; EG-NEXT: OR_INT * T1.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T21.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T39.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T40.X, PS, literal.x, | |
| ; EG-NEXT: LSHR T0.W, T0.W, literal.y, | |
| ; EG-NEXT: ADD_INT * T41.X, PS, literal.z, | |
| ; EG-NEXT: 4(5.605194e-45), 8(1.121039e-44) | |
| ; EG-NEXT: 12(1.681558e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, T0.X, literal.x, | |
| ; EG-NEXT: AND_INT * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; EG-NEXT: ADD_INT T42.X, T39.X, literal.x, | |
| ; EG-NEXT: OR_INT * T0.W, PV.W, PS, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: MOV T21.X, PV.W, | |
| ; EG-NEXT: MOV * T36.X, T16.X, | |
| ; EG-NEXT: MOV * T36.Z, T12.X, | |
| ; EG-NEXT: MOV T37.X, T8.X, | |
| ; EG-NEXT: MOV T37.Z, T4.X, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: MOV * T38.X, T32.X, | |
| ; EG-NEXT: MOV * T38.Z, T28.X, | |
| ; EG-NEXT: MOV T0.X, T24.X, | |
| ; EG-NEXT: MOV * T0.Z, T20.X, BS:VEC_120/SCL_212 | |
| ; | |
| ; CM-LABEL: global_zextload_v32i8_to_v32i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 3, @14, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 1 @10 | |
| ; CM-NEXT: ALU 101, @18, KC0[], KC1[] | |
| ; CM-NEXT: ALU 101, @120, KC0[], KC1[] | |
| ; CM-NEXT: ALU 36, @222, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T42.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T38, T39.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T37, T41.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T36, T40.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 10: | |
| ; CM-NEXT: VTX_READ_128 T37.XYZW, T0.X, 16, #1 | |
| ; CM-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 14: | |
| ; CM-NEXT: MOV T1.Y, T16.X, | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 5(7.006492e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 18: | |
| ; CM-NEXT: AND_INT T1.Z, T37.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, T1.Y, literal.y, | |
| ; CM-NEXT: 255(3.573311e-43), -65536(nan) | |
| ; CM-NEXT: OR_INT * T1.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T16.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHL * T1.W, T37.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T16.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, T17.X, | |
| ; CM-NEXT: MOV * T1.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T2.W, T37.X, literal.y, PV.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T17.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHR * T2.W, T37.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T36.Y, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T17.X, PV.Y, | |
| ; CM-NEXT: MOV * T1.Y, T12.X, | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, T37.Y, literal.y, | |
| ; CM-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T12.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHL * T2.W, T37.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T12.X, PV.W, | |
| ; CM-NEXT: MOV * T1.Y, T13.X, | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T2.W, T37.Y, literal.y, T1.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T13.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHR * T2.W, T37.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T36.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T13.X, PV.W, | |
| ; CM-NEXT: MOV * T1.Y, T8.X, | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, T37.Z, literal.y, | |
| ; CM-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T8.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHL * T2.W, T37.Z, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T8.X, PV.W, | |
| ; CM-NEXT: MOV * T1.Y, T9.X, | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T2.W, T37.Z, literal.y, T1.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T9.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHR * T2.W, T37.Z, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T37.Y, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T9.X, PV.Y, | |
| ; CM-NEXT: MOV * T1.Y, T4.X, | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, T37.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T4.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHL * T2.W, T37.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T4.X, PV.W, | |
| ; CM-NEXT: MOV * T1.Y, T5.X, | |
| ; CM-NEXT: AND_INT * T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: -65536(nan), 0(0.000000e+00) | |
| ; CM-NEXT: ALU clause starting at 120: | |
| ; CM-NEXT: BFE_UINT * T2.W, T37.W, literal.x, T1.W, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: OR_INT * T2.W, T1.Z, PV.W, | |
| ; CM-NEXT: MOV * T5.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHR * T2.W, T37.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T37.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T5.X, PV.W, | |
| ; CM-NEXT: MOV * T1.Y, T32.X, | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, T0.X, literal.y, | |
| ; CM-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T32.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHL * T2.W, T0.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T32.X, PV.W, | |
| ; CM-NEXT: MOV * T1.Y, T33.X, | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: BFE_UINT * T2.W, T0.X, literal.y, T1.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T33.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHR * T2.W, T0.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T38.Y, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T33.X, PV.Y, | |
| ; CM-NEXT: MOV * T0.X, T28.X, | |
| ; CM-NEXT: AND_INT T1.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, T0.Y, literal.y, | |
| ; CM-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T28.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHL * T2.W, T0.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T28.X, PV.W, | |
| ; CM-NEXT: MOV * T0.X, T29.X, | |
| ; CM-NEXT: AND_INT T1.Z, PV.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T2.W, T0.Y, literal.y, T1.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T29.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T2.W, T0.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T38.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T29.X, PV.W, | |
| ; CM-NEXT: MOV * T0.X, T24.X, | |
| ; CM-NEXT: AND_INT T1.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, T0.Z, literal.y, | |
| ; CM-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T24.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHL * T2.W, T0.Z, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T24.X, PV.W, | |
| ; CM-NEXT: MOV * T0.X, T25.X, | |
| ; CM-NEXT: AND_INT T1.Z, PV.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T2.W, T0.Z, literal.y, T1.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T25.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T2.W, T0.Z, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T0.Y, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T25.X, PV.Y, | |
| ; CM-NEXT: MOV * T0.X, T20.X, | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, T0.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 255(3.573311e-43) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T20.X, PV.W, | |
| ; CM-NEXT: ALU clause starting at 222: | |
| ; CM-NEXT: MOV T0.X, T20.X, | |
| ; CM-NEXT: LSHL * T2.W, T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T2.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16711680(2.341805e-38) | |
| ; CM-NEXT: OR_INT * T2.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T20.X, PV.W, | |
| ; CM-NEXT: MOV * T0.X, T21.X, | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: BFE_UINT * T1.W, T0.W, literal.y, T1.W, | |
| ; CM-NEXT: -65536(nan), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T21.X, PV.W, | |
| ; CM-NEXT: MOV * T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T39.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T40.X, PV.X, literal.x, | |
| ; CM-NEXT: LSHR * T0.W, T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T41.X, T39.X, literal.x, | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: AND_INT * T0.W, PV.W, literal.z, | |
| ; CM-NEXT: 12(1.681558e-44), 65535(9.183409e-41) | |
| ; CM-NEXT: 16711680(2.341805e-38), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T42.X, T39.X, literal.x, | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; CM-NEXT: MOV * T21.X, PV.W, | |
| ; CM-NEXT: MOV T36.X, T16.X, | |
| ; CM-NEXT: MOV * T36.Z, T12.X, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: MOV T37.X, T8.X, | |
| ; CM-NEXT: MOV * T37.Z, T4.X, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: MOV T38.X, T32.X, | |
| ; CM-NEXT: MOV * T38.Z, T28.X, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: MOV T0.X, T24.X, | |
| ; CM-NEXT: MOV * T0.Z, T20.X, BS:VEC_120/SCL_212 | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <32 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <32 x i8>, ptr addrspace(1) %in.tid | |
| %ext = zext <32 x i8> %load to <32 x i16> | |
| store <32 x i16> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| define amdgpu_kernel void @global_sextload_v32i8_to_v32i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; GCN-NOHSA-SI-LABEL: global_sextload_v32i8_to_v32i16: | |
| ; GCN-NOHSA-SI: ; %bb.0: | |
| ; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, 0 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s3 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v4, 5, v0 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b64 s[8:9], s[6:7] | |
| ; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v5, 0 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[0:3], v[4:5], s[8:11], 0 addr64 | |
| ; GCN-NOHSA-SI-NEXT: buffer_load_dwordx4 v[4:7], v[4:5], s[8:11], 0 addr64 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s4 | |
| ; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s5 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v8, 24, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v9, v1, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v10, v1, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v1, v1, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v11, 24, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v12, v0, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v14, 24, v3 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v15, v3, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v16, v3, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v17, v3, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v18, 24, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v19, v2, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v20, v2, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v21, v2, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v31, 24, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v32, v6, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v33, v6, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v34, v6, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v2, 16, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v3, 0xffff, v9 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v6, 16, v10 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v13, v0, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v22, 24, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v23, v5, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v24, v5, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v25, 24, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v26, v4, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v27, v4, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v4, v4, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v28, 24, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v29, v7, 16, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v30, v7, 8, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v7, v7, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v3, v3, v2 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v2, v1, v6 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v1, 16, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v6, 0xffff, v12 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v10, 0xffff, v15 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v11, 16, v16 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v12, 0xffff, v17 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v15, 16, v20 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v16, 0xffff, v21 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_bfe_i32 v5, v5, 0, 8 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v8, 16, v13 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v9, 16, v14 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v13, 16, v18 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v14, 0xffff, v19 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v17, 16, v22 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v18, 0xffff, v23 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v19, 16, v24 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v21, 16, v25 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v22, 0xffff, v26 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v23, 16, v27 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v24, 0xffff, v4 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v25, 16, v28 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v26, 0xffff, v29 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v27, 16, v30 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v28, 0xffff, v7 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v29, 16, v31 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v30, 0xffff, v32 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v1, v6, v1 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v6, v12, v11 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v4, v16, v15 | |
| ; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v12, 16, v33 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v16, 0xffff, v34 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 | |
| ; GCN-NOHSA-SI-NEXT: v_and_b32_e32 v20, 0xffff, v5 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v5, v14, v13 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v15, v26, v25 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v14, v28, v27 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v13, v30, v29 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v12, v16, v12 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v0, v0, v8 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v7, v10, v9 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v11, v18, v17 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v10, v20, v19 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v9, v22, v21 | |
| ; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v8, v24, v23 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-SI-NEXT: s_endpgm | |
| ; | |
| ; GCN-HSA-LABEL: global_sextload_v32i8_to_v32i16: | |
| ; GCN-HSA: ; %bb.0: | |
| ; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-HSA-NEXT: s_add_i32 s12, s12, s17 | |
| ; GCN-HSA-NEXT: s_mov_b32 flat_scratch_lo, s13 | |
| ; GCN-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v4, vcc, s2, v0 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[4:5] | |
| ; GCN-HSA-NEXT: v_add_i32_e32 v4, vcc, 16, v4 | |
| ; GCN-HSA-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc | |
| ; GCN-HSA-NEXT: flat_load_dwordx4 v[4:7], v[4:5] | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 16 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v8, 24, v1 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v9, v1, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v10, v1, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v1, v1, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v11, 24, v0 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v12, v0, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v13, v0, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v0, v0, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v14, 24, v3 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v15, v3, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v16, v3, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v17, v3, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v18, 24, v2 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v19, v2, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v20, v2, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v21, v2, 0, 8 | |
| ; GCN-HSA-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v22, 24, v5 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v23, v5, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v24, v5, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v25, v5, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v26, 24, v4 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v27, v4, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v28, v4, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v4, v4, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v29, 24, v7 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v30, v7, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v31, v7, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v32, v7, 0, 8 | |
| ; GCN-HSA-NEXT: v_ashrrev_i32_e32 v33, 24, v6 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v34, v6, 16, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v35, v6, 8, 8 | |
| ; GCN-HSA-NEXT: v_bfe_i32 v36, v6, 0, 8 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v2, 16, v8 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v3, 0xffff, v9 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v5, 16, v10 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v1, 0xffff, v1 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v6, 16, v13 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v0, 0xffff, v0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v7, 16, v14 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v8, 0xffff, v15 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v9, 16, v16 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v10, 0xffff, v17 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v3, v3, v2 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v2, v1, v5 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v1, 16, v11 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v5, 0xffff, v12 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v11, 16, v18 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v12, 0xffff, v19 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v0, v0, v6 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v7, v8, v7 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v6, v10, v9 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v8, 16, v22 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v9, 0xffff, v23 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v15, 16, v28 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v4, 0xffff, v4 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v16, 16, v29 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v17, 0xffff, v30 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v1, v5, v1 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v5, v12, v11 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v11, v9, v8 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v8, v4, v15 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v15, v17, v16 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v4, 16, v20 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v16, 0xffff, v21 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v4, v16, v4 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v17, s3 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v16, s2 | |
| ; GCN-HSA-NEXT: s_add_u32 s2, s0, 48 | |
| ; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v10, 16, v24 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v12, 0xffff, v25 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v13, 16, v26 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v14, 0xffff, v27 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v18, 16, v31 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v19, 0xffff, v32 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v22, 16, v33 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v23, 0xffff, v34 | |
| ; GCN-HSA-NEXT: v_lshlrev_b32_e32 v24, 16, v35 | |
| ; GCN-HSA-NEXT: v_and_b32_e32 v25, 0xffff, v36 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[16:17], v[4:7] | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v10, v12, v10 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0 | |
| ; GCN-HSA-NEXT: s_add_u32 s0, s0, 32 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v9, v14, v13 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v14, v19, v18 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v13, v23, v22 | |
| ; GCN-HSA-NEXT: v_or_b32_e32 v12, v25, v24 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3] | |
| ; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[12:15] | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0 | |
| ; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1 | |
| ; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[8:11] | |
| ; GCN-HSA-NEXT: s_endpgm | |
| ; | |
| ; GCN-NOHSA-VI-LABEL: global_sextload_v32i8_to_v32i16: | |
| ; GCN-NOHSA-VI: ; %bb.0: | |
| ; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 5, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, 0xffff | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, 8 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3 | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v6, vcc, s2, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v1, vcc | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[0:3], v[6:7] | |
| ; GCN-NOHSA-VI-NEXT: v_add_u32_e32 v10, vcc, 16, v6 | |
| ; GCN-NOHSA-VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v7, vcc | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s3, 0xf000 | |
| ; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, -1 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0) | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v6, v4, sext(v3) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v7, v5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v9, v4, sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v12, v5, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v8, v6, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v6, v9, v12 | |
| ; GCN-NOHSA-VI-NEXT: flat_load_dwordx4 v[10:13], v[10:11] | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 16, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v7, v4, sext(v3) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v3, v5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v9, v7, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v3, v4, sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v2, v5, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v7, v3, v2 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v3, 16, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v2, v4, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v1, v5, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[6:9], off, s[0:3], 0 offset:16 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v2, v2, v1 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v7, 16, v0 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v1, v4, sext(v3) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v3, v5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v3, v1, v3 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v1, v4, sext(v7) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v7, v5, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v1, v1, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v6, v4, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v0, v5, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v0, v6, v0 | |
| ; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(1) | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v9, 16, v11 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v17, 24, v11 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v16, 16, v10 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v8, v4, sext(v11) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v11, v5, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v7, 16, v17 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v9, v4, sext(v9) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v14, 16, v12 | |
| ; GCN-NOHSA-VI-NEXT: v_lshrrev_b32_e32 v15, 16, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v19, v4, sext(v13) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v13, v5, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v8, v8, v11 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v9, v9, v7 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v7, v4, sext(v16) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v11, v5, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v18, v4, sext(v10) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v10, v5, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v20, v4, sext(v12) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v21, v5, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v12, v19, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v7, v7, v11 | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v11, v4, sext(v15) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v13, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_and_b32_sdwa v4, v4, sext(v14) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0 | |
| ; GCN-NOHSA-VI-NEXT: v_ashrrev_i16_sdwa v5, v5, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v6, v18, v10 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v10, v20, v21 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v13, v11, v13 | |
| ; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v11, v4, v5 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[10:13], off, s[0:3], 0 offset:48 | |
| ; GCN-NOHSA-VI-NEXT: buffer_store_dwordx4 v[6:9], off, s[0:3], 0 offset:32 | |
| ; GCN-NOHSA-VI-NEXT: s_endpgm | |
| ; | |
| ; EG-LABEL: global_sextload_v32i8_to_v32i16: | |
| ; EG: ; %bb.0: | |
| ; EG-NEXT: ALU 3, @14, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: TEX 1 @10 | |
| ; EG-NEXT: ALU 104, @18, KC0[], KC1[] | |
| ; EG-NEXT: ALU 104, @123, KC0[], KC1[] | |
| ; EG-NEXT: ALU 91, @228, KC0[CB0:0-32], KC1[] | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T36.XYZW, T42.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T37.XYZW, T41.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T38.XYZW, T39.X, 0 | |
| ; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T40.X, 1 | |
| ; EG-NEXT: CF_END | |
| ; EG-NEXT: Fetch clause starting at 10: | |
| ; EG-NEXT: VTX_READ_128 T37.XYZW, T0.X, 16, #1 | |
| ; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; EG-NEXT: ALU clause starting at 14: | |
| ; EG-NEXT: MOV T1.Y, T16.X, | |
| ; EG-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 5(7.006492e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; EG-NEXT: ALU clause starting at 18: | |
| ; EG-NEXT: BFE_INT * T1.W, T37.X, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, PV.W, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, T1.Y, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; EG-NEXT: OR_INT * T1.W, PS, PV.W, | |
| ; EG-NEXT: MOV * T16.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T37.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV T16.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, T17.X, | |
| ; EG-NEXT: LSHR * T1.W, T37.X, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T17.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: ASHR * T1.W, T37.X, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: OR_INT * T36.Y, PV.W, PS, | |
| ; EG-NEXT: MOV T17.X, PV.Y, | |
| ; EG-NEXT: MOV T1.Y, T12.X, | |
| ; EG-NEXT: BFE_INT * T1.W, T37.Y, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T12.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T37.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV T12.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, T13.X, | |
| ; EG-NEXT: LSHR * T1.W, T37.Y, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T13.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: ASHR * T1.W, T37.Y, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: OR_INT * T36.W, PV.W, PS, | |
| ; EG-NEXT: MOV T13.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, T8.X, | |
| ; EG-NEXT: BFE_INT * T1.W, T37.Z, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T8.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T37.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV T8.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, T9.X, | |
| ; EG-NEXT: LSHR * T1.W, T37.Z, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T9.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: ASHR * T1.W, T37.Z, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: ALU clause starting at 123: | |
| ; EG-NEXT: OR_INT * T37.Y, T2.W, T1.W, | |
| ; EG-NEXT: MOV T9.X, PV.Y, | |
| ; EG-NEXT: MOV T1.Y, T4.X, | |
| ; EG-NEXT: BFE_INT * T1.W, T37.W, 0.0, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T4.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T37.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV T4.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, T5.X, | |
| ; EG-NEXT: LSHR * T1.W, T37.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T5.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: ASHR * T1.W, T37.W, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: OR_INT * T37.W, PV.W, PS, | |
| ; EG-NEXT: MOV T5.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, T32.X, | |
| ; EG-NEXT: BFE_INT * T1.W, T0.X, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; EG-NEXT: OR_INT * T1.W, PV.W, PS, | |
| ; EG-NEXT: MOV * T32.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T0.X, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV T32.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, T33.X, | |
| ; EG-NEXT: LSHR * T1.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.Y, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T33.X, PV.W, | |
| ; EG-NEXT: MOV T1.Y, PV.X, | |
| ; EG-NEXT: ASHR * T1.W, T0.X, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.Y, literal.x, | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: OR_INT * T38.Y, PV.W, PS, | |
| ; EG-NEXT: MOV T33.X, PV.Y, | |
| ; EG-NEXT: MOV * T0.X, T28.X, | |
| ; EG-NEXT: BFE_INT T1.W, T0.Y, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PS, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T28.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T0.Y, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV T28.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T29.X, | |
| ; EG-NEXT: LSHR * T1.W, T0.Y, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T29.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: ASHR * T1.W, T0.Y, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: ALU clause starting at 228: | |
| ; EG-NEXT: AND_INT T2.W, T0.X, literal.x, | |
| ; EG-NEXT: LSHL * T1.W, T1.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: OR_INT * T38.W, PV.W, PS, | |
| ; EG-NEXT: MOV T29.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T24.X, | |
| ; EG-NEXT: BFE_INT T1.W, T0.Z, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PS, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T24.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T0.Z, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV T24.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T25.X, | |
| ; EG-NEXT: LSHR * T1.W, T0.Z, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T25.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: ASHR * T1.W, T0.Z, literal.x, | |
| ; EG-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T2.W, PV.X, literal.x, | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: OR_INT * T0.Y, PV.W, PS, | |
| ; EG-NEXT: MOV T25.X, PV.Y, | |
| ; EG-NEXT: MOV * T0.X, T20.X, | |
| ; EG-NEXT: BFE_INT T1.W, T0.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PS, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T20.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T1.W, T0.W, literal.x, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, PV.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), 65535(9.183409e-41) | |
| ; EG-NEXT: LSHL * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV T20.X, PV.W, | |
| ; EG-NEXT: MOV * T0.X, T21.X, | |
| ; EG-NEXT: LSHR * T1.W, T0.W, literal.x, | |
| ; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; EG-NEXT: BFE_INT T1.W, PV.W, 0.0, literal.x, | |
| ; EG-NEXT: AND_INT * T2.W, T0.X, literal.y, | |
| ; EG-NEXT: 8(1.121039e-44), -65536(nan) | |
| ; EG-NEXT: AND_INT * T1.W, PV.W, literal.x, | |
| ; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00) | |
| ; EG-NEXT: OR_INT * T1.W, T2.W, PV.W, | |
| ; EG-NEXT: MOV * T21.X, PV.W, | |
| ; EG-NEXT: MOV T0.X, PV.X, | |
| ; EG-NEXT: LSHR * T39.X, KC0[2].Y, literal.x, | |
| ; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; EG-NEXT: ADD_INT T40.X, PS, literal.x, | |
| ; EG-NEXT: ASHR T0.W, T0.W, literal.y, | |
| ; EG-NEXT: ADD_INT * T41.X, PS, literal.z, | |
| ; EG-NEXT: 4(5.605194e-45), 24(3.363116e-44) | |
| ; EG-NEXT: 12(1.681558e-44), 0(0.000000e+00) | |
| ; EG-NEXT: AND_INT T1.W, T0.X, literal.x, | |
| ; EG-NEXT: LSHL * T0.W, PV.W, literal.y, | |
| ; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; EG-NEXT: ADD_INT T42.X, T39.X, literal.x, | |
| ; EG-NEXT: OR_INT * T0.W, PV.W, PS, | |
| ; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; EG-NEXT: MOV T21.X, PV.W, | |
| ; EG-NEXT: MOV * T36.X, T16.X, | |
| ; EG-NEXT: MOV * T36.Z, T12.X, | |
| ; EG-NEXT: MOV T37.X, T8.X, | |
| ; EG-NEXT: MOV T37.Z, T4.X, BS:VEC_120/SCL_212 | |
| ; EG-NEXT: MOV * T38.X, T32.X, | |
| ; EG-NEXT: MOV * T38.Z, T28.X, | |
| ; EG-NEXT: MOV T0.X, T24.X, | |
| ; EG-NEXT: MOV * T0.Z, T20.X, BS:VEC_120/SCL_212 | |
| ; | |
| ; CM-LABEL: global_sextload_v32i8_to_v32i16: | |
| ; CM: ; %bb.0: | |
| ; CM-NEXT: ALU 3, @14, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: TEX 1 @10 | |
| ; CM-NEXT: ALU 104, @18, KC0[], KC1[] | |
| ; CM-NEXT: ALU 104, @123, KC0[], KC1[] | |
| ; CM-NEXT: ALU 91, @228, KC0[CB0:0-32], KC1[] | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0, T42.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T38, T39.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T37, T41.X | |
| ; CM-NEXT: MEM_RAT_CACHELESS STORE_DWORD T36, T40.X | |
| ; CM-NEXT: CF_END | |
| ; CM-NEXT: Fetch clause starting at 10: | |
| ; CM-NEXT: VTX_READ_128 T37.XYZW, T0.X, 16, #1 | |
| ; CM-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1 | |
| ; CM-NEXT: ALU clause starting at 14: | |
| ; CM-NEXT: MOV T1.Y, T16.X, | |
| ; CM-NEXT: LSHL * T0.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 5(7.006492e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT * T0.X, KC0[2].Z, PV.W, | |
| ; CM-NEXT: ALU clause starting at 18: | |
| ; CM-NEXT: BFE_INT * T1.W, T37.X, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.W, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, T1.Y, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), -65536(nan) | |
| ; CM-NEXT: OR_INT * T1.W, PV.W, PV.Z, | |
| ; CM-NEXT: MOV * T16.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T37.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T1.Y, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T16.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, T17.X, | |
| ; CM-NEXT: LSHR * T1.W, T37.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T1.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T17.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: ASHR * T1.W, T37.X, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T36.Y, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T17.X, PV.Y, | |
| ; CM-NEXT: MOV T1.Y, T12.X, | |
| ; CM-NEXT: BFE_INT * T1.W, T37.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T12.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T37.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T1.Y, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T12.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, T13.X, | |
| ; CM-NEXT: LSHR * T1.W, T37.Y, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T1.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T13.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: ASHR * T1.W, T37.Y, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T36.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T13.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, T8.X, | |
| ; CM-NEXT: BFE_INT * T1.W, T37.Z, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T8.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T37.Z, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T1.Y, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T8.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, T9.X, | |
| ; CM-NEXT: LSHR * T1.W, T37.Z, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T1.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T9.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: ASHR * T1.W, T37.Z, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: ALU clause starting at 123: | |
| ; CM-NEXT: OR_INT * T37.Y, T1.Z, T1.W, | |
| ; CM-NEXT: MOV T9.X, PV.Y, | |
| ; CM-NEXT: MOV T1.Y, T4.X, | |
| ; CM-NEXT: BFE_INT * T1.W, T37.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T4.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T37.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T1.Y, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T4.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, T5.X, | |
| ; CM-NEXT: LSHR * T1.W, T37.W, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T1.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T5.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: ASHR * T1.W, T37.W, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T37.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T5.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, T32.X, | |
| ; CM-NEXT: BFE_INT * T1.W, T0.X, 0.0, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T32.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T0.X, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T1.Y, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV T32.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, T33.X, | |
| ; CM-NEXT: LSHR * T1.W, T0.X, literal.x, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T1.Y, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T33.X, PV.W, | |
| ; CM-NEXT: MOV T1.Y, PV.X, | |
| ; CM-NEXT: ASHR * T1.W, T0.X, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.Y, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T38.Y, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T33.X, PV.Y, | |
| ; CM-NEXT: MOV T0.X, T28.X, | |
| ; CM-NEXT: BFE_INT * T1.W, T0.Y, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T28.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T0.Y, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T28.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T29.X, | |
| ; CM-NEXT: LSHR * T1.W, T0.Y, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T0.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T29.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: ASHR * T1.W, T0.Y, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ALU clause starting at 228: | |
| ; CM-NEXT: AND_INT T1.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, T1.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T38.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T29.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T24.X, | |
| ; CM-NEXT: BFE_INT * T1.W, T0.Z, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T24.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T0.Z, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T24.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T25.X, | |
| ; CM-NEXT: LSHR * T1.W, T0.Z, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T1.Z, T0.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T25.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: ASHR * T1.W, T0.Z, literal.x, | |
| ; CM-NEXT: 24(3.363116e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T0.Y, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T25.X, PV.Y, | |
| ; CM-NEXT: MOV T0.X, T20.X, | |
| ; CM-NEXT: BFE_INT * T1.W, T0.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, PV.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T20.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T1.W, T0.W, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: LSHL * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: 65535(9.183409e-41), 16(2.242078e-44) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T20.X, PV.W, | |
| ; CM-NEXT: MOV T0.X, T21.X, | |
| ; CM-NEXT: LSHR * T1.W, T0.W, literal.x, | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: BFE_INT * T1.W, PV.W, 0.0, literal.x, | |
| ; CM-NEXT: 8(1.121039e-44), 0(0.000000e+00) | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.x, | |
| ; CM-NEXT: AND_INT * T1.W, PV.W, literal.y, | |
| ; CM-NEXT: -65536(nan), 65535(9.183409e-41) | |
| ; CM-NEXT: OR_INT * T1.W, PV.Z, PV.W, | |
| ; CM-NEXT: MOV * T21.X, PV.W, | |
| ; CM-NEXT: MOV * T0.X, PV.X, | |
| ; CM-NEXT: LSHR * T39.X, KC0[2].Y, literal.x, | |
| ; CM-NEXT: 2(2.802597e-45), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T40.X, PV.X, literal.x, | |
| ; CM-NEXT: ASHR * T0.W, T0.W, literal.y, | |
| ; CM-NEXT: 8(1.121039e-44), 24(3.363116e-44) | |
| ; CM-NEXT: ADD_INT T41.X, T39.X, literal.x, | |
| ; CM-NEXT: AND_INT T0.Z, T0.X, literal.y, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: LSHL * T0.W, PV.W, literal.z, | |
| ; CM-NEXT: 12(1.681558e-44), 65535(9.183409e-41) | |
| ; CM-NEXT: 16(2.242078e-44), 0(0.000000e+00) | |
| ; CM-NEXT: ADD_INT T42.X, T39.X, literal.x, | |
| ; CM-NEXT: OR_INT * T0.W, PV.Z, PV.W, | |
| ; CM-NEXT: 4(5.605194e-45), 0(0.000000e+00) | |
| ; CM-NEXT: MOV * T21.X, PV.W, | |
| ; CM-NEXT: MOV T36.X, T16.X, | |
| ; CM-NEXT: MOV * T36.Z, T12.X, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: MOV T37.X, T8.X, | |
| ; CM-NEXT: MOV * T37.Z, T4.X, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: MOV T38.X, T32.X, | |
| ; CM-NEXT: MOV * T38.Z, T28.X, BS:VEC_120/SCL_212 | |
| ; CM-NEXT: MOV T0.X, T24.X, | |
| ; CM-NEXT: MOV * T0.Z, T20.X, BS:VEC_120/SCL_212 | |
| %tid = call i32 @llvm.amdgcn.workitem.id.x() | |
| %in.tid = getelementptr inbounds <32 x i8>, ptr addrspace(1) %in, i32 %tid | |
| %load = load <32 x i8>, ptr addrspace(1) %in.tid | |
| %ext = sext <32 x i8> %load to <32 x i16> | |
| store <32 x i16> %ext, ptr addrspace(1) %out | |
| ret void | |
| } | |
| ; XFUNC-LABEL: {{^}}global_zextload_v64i8_to_v64i16: | |
| ; define amdgpu_kernel void @global_zextload_v64i8_to_v64i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; %load = load <64 x i8>, ptr addrspace(1) %in | |
| ; %ext = zext <64 x i8> %load to <64 x i16> | |
| ; store <64 x i16> %ext, ptr addrspace(1) %out | |
| ; ret void | |
| ; } | |
| ; XFUNC-LABEL: {{^}}global_sextload_v64i8_to_v64i16: | |
| ; define amdgpu_kernel void @global_sextload_v64i8_to_v64i16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { | |
| ; %load = load <64 x i8>, ptr addrspace(1) %in | |
| ; %ext = sext <64 x i8> %load to <64 x i16> | |
| ; store <64 x i16> %ext, ptr addrspace(1) %out | |
| ; ret void | |
| ; } | |
| attributes #0 = { nounwind } |