maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.load.ll

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

819 lines
59 KiB
LLVM
Raw Normal View History

AMDGPU/GlobalISel: Select llvm.amdgcn.raw.buffer.load Use intermediate instructions, unlike with buffer stores. This is necessary because of the need to have an internal way to distinguish between signed and unsigned extloads. This introduces some duplication and near duplication with the buffer store selection path. The store handling should maybe be moved into legalization to match and eliminate the duplication.
2019-09-19 12:29:20 +08:00
; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -stop-after=instruction-select -verify-machineinstrs -o - %s | FileCheck %s
; FIXME: Test with SI when argument lowering not broken for f16
; Natural mapping
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret float %val
}
; Copies for VGPR arguments
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__sgpr_voffset__sgpr_soffset(<4 x i32> inreg %rsrc, i32 inreg %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__sgpr_voffset__sgpr_soffset
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $sgpr7
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr7
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY4]]
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY6]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret float %val
}
; Waterfall for rsrc
define amdgpu_ps float @raw_buffer_load_f32__vgpr_rsrc__vgpr_voffset__sgpr_soffset(<4 x i32> %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__vgpr_rsrc__vgpr_voffset__sgpr_soffset
; CHECK: bb.1 (%ir-block.0):
; CHECK: successors: %bb.2(0x80000000)
; CHECK: liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4
; CHECK: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr4
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[COPY6:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
; CHECK: [[COPY7:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3
; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
; CHECK: bb.2:
; CHECK: successors: %bb.3(0x40000000), %bb.2(0x40000000)
; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE1]], [[COPY6]], implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE2]], [[COPY7]], implicit $exec
; CHECK: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
; CHECK: [[REG_SEQUENCE3:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE3]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_]], implicit-def $exec, implicit-def $scc, implicit $exec
; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
; CHECK: S_CBRANCH_EXECNZ %bb.2, implicit $exec
; CHECK: bb.3:
; CHECK: successors: %bb.4(0x80000000)
; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
; CHECK: bb.4:
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret float %val
}
; Waterfall for rsrc and soffset
define amdgpu_ps float @raw_buffer_load_f32__vgpr_rsrc__vgpr_voffset__vgpr_soffset(<4 x i32> %rsrc, i32 %voffset, i32 %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__vgpr_rsrc__vgpr_voffset__vgpr_soffset
; CHECK: bb.1 (%ir-block.0):
; CHECK: successors: %bb.2(0x80000000)
; CHECK: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5
; CHECK: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr4
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr5
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[COPY6:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
; CHECK: [[COPY7:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3
; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
; CHECK: bb.2:
; CHECK: successors: %bb.3(0x40000000), %bb.2(0x40000000)
; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE1]], [[COPY6]], implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE2]], [[COPY7]], implicit $exec
; CHECK: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
; CHECK: [[REG_SEQUENCE3:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
; CHECK: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY5]], implicit $exec
; CHECK: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY5]], implicit $exec
; CHECK: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U32_e64_]], [[S_AND_B64_]], implicit-def $scc
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE3]], [[V_READFIRSTLANE_B32_4]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_1]], implicit-def $exec, implicit-def $scc, implicit $exec
; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
; CHECK: S_CBRANCH_EXECNZ %bb.2, implicit $exec
; CHECK: bb.3:
; CHECK: successors: %bb.4(0x80000000)
; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
; CHECK: bb.4:
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret float %val
}
; Natural mapping + glc
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_glc(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_glc
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 1, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 1)
ret float %val
}
; Natural mapping + slc
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 1, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 2)
ret float %val
}
; Natural mapping + dlc
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_dlc(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_dlc
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 0, 0, 1, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 4)
ret float %val
}
; Natural mapping + slc + dlc
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc_dlc(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc_dlc
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 1, 0, 1, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 6)
ret float %val
}
; Natural mapping + glc + dlc
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_glc_dlc(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_glc_dlc
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 1, 0, 0, 1, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 5)
ret float %val
}
; Natural mapping + glc + slc + dlc
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_glc_slc_dlc(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_glc_slc_dlc
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 1, 1, 0, 1, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 7)
ret float %val
}
; Natural mapping
define amdgpu_ps <2 x float> @raw_buffer_load_v2f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_v2f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORDX2_OFFEN:%[0-9]+]]:vreg_64 = BUFFER_LOAD_DWORDX2_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 8 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_DWORDX2_OFFEN]].sub0
; CHECK: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_DWORDX2_OFFEN]].sub1
; CHECK: $vgpr0 = COPY [[COPY6]]
; CHECK: $vgpr1 = COPY [[COPY7]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0, implicit $vgpr1
%val = call <2 x float> @llvm.amdgcn.raw.buffer.load.v2f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret <2 x float> %val
}
define amdgpu_ps <3 x float> @raw_buffer_load_v3f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_v3f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORDX3_OFFEN:%[0-9]+]]:vreg_96 = BUFFER_LOAD_DWORDX3_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 12 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_DWORDX3_OFFEN]].sub0
; CHECK: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_DWORDX3_OFFEN]].sub1
; CHECK: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_DWORDX3_OFFEN]].sub2
; CHECK: $vgpr0 = COPY [[COPY6]]
; CHECK: $vgpr1 = COPY [[COPY7]]
; CHECK: $vgpr2 = COPY [[COPY8]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0, implicit $vgpr1, implicit $vgpr2
%val = call <3 x float> @llvm.amdgcn.raw.buffer.load.v3f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret <3 x float> %val
}
define amdgpu_ps <4 x float> @raw_buffer_load_v4f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_v4f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORDX4_OFFEN:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_DWORDX4_OFFEN]].sub0
; CHECK: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_DWORDX4_OFFEN]].sub1
; CHECK: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_DWORDX4_OFFEN]].sub2
; CHECK: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_DWORDX4_OFFEN]].sub3
; CHECK: $vgpr0 = COPY [[COPY6]]
; CHECK: $vgpr1 = COPY [[COPY7]]
; CHECK: $vgpr2 = COPY [[COPY8]]
; CHECK: $vgpr3 = COPY [[COPY9]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0, implicit $vgpr1, implicit $vgpr2, implicit $vgpr3
%val = call <4 x float> @llvm.amdgcn.raw.buffer.load.v4f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret <4 x float> %val
}
define amdgpu_ps half @raw_buffer_load_f16__sgpr_rsrc__vgpr_voffset__sgpr_soffset(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f16__sgpr_rsrc__vgpr_voffset__sgpr_soffset
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_USHORT_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 2 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_USHORT_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call half @llvm.amdgcn.raw.buffer.load.f16(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret half %val
}
define amdgpu_ps <2 x half> @raw_buffer_load_v2f16__sgpr_rsrc__vgpr_voffset__sgpr_soffset(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_v2f16__sgpr_rsrc__vgpr_voffset__sgpr_soffset
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call <2 x half> @llvm.amdgcn.raw.buffer.load.v2f16(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret <2 x half> %val
}
; FIXME: Crashes
; define amdgpu_ps <3 x half> @raw_buffer_load_v3f16__sgpr_rsrc__vgpr_voffset__sgpr_soffset(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; %val = call <3 x half> @llvm.amdgcn.raw.buffer.load.v3f16(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
; ret <3 x half> %val
; }
define amdgpu_ps <4 x half> @raw_buffer_load_v4f16__sgpr_rsrc__vgpr_voffset__sgpr_soffset(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_v4f16__sgpr_rsrc__vgpr_voffset__sgpr_soffset
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORDX2_OFFEN:%[0-9]+]]:vreg_64 = BUFFER_LOAD_DWORDX2_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 8 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_DWORDX2_OFFEN]].sub0
; CHECK: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_DWORDX2_OFFEN]].sub1
; CHECK: $vgpr0 = COPY [[COPY6]]
; CHECK: $vgpr1 = COPY [[COPY7]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0, implicit $vgpr1
%val = call <4 x half> @llvm.amdgcn.raw.buffer.load.v4f16(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret <4 x half> %val
}
define amdgpu_ps float @raw_buffer_load_i8__sgpr_rsrc__vgpr_voffset__sgpr_soffset_zext(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_i8__sgpr_rsrc__vgpr_voffset__sgpr_soffset_zext
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_UBYTE_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_UBYTE_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 1 from custom "TargetCustom7", addrspace 4)
[AMDGPU][GlobalISel] Compute known bits for zero-extending loads Implement computeKnownBitsForTargetInstr for G_AMDGPU_BUFFER_LOAD_UBYTE and G_AMDGPU_BUFFER_LOAD_USHORT. This allows generic combines to remove some unnecessary G_ANDs. Differential Revision: https://reviews.llvm.org/D89316
2020-10-13 21:40:27 +08:00
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_UBYTE_OFFEN]]
AMDGPU/GlobalISel: Select llvm.amdgcn.raw.buffer.load Use intermediate instructions, unlike with buffer stores. This is necessary because of the need to have an internal way to distinguish between signed and unsigned extloads. This introduces some duplication and near duplication with the buffer store selection path. The store handling should maybe be moved into legalization to match and eliminate the duplication.
2019-09-19 12:29:20 +08:00
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call i8 @llvm.amdgcn.raw.buffer.load.i8(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
%zext = zext i8 %val to i32
%cast = bitcast i32 %zext to float
ret float %cast
}
define amdgpu_ps float @raw_buffer_load_i8__sgpr_rsrc__vgpr_voffset__sgpr_soffset_sext(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_i8__sgpr_rsrc__vgpr_voffset__sgpr_soffset_sext
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_UBYTE_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_UBYTE_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 1 from custom "TargetCustom7", addrspace 4)
AMDGPU/GlobalISel: Select G_SEXT_INREG
2020-01-25 08:38:53 +08:00
; CHECK: [[V_BFE_I32_:%[0-9]+]]:vgpr_32 = V_BFE_I32 [[BUFFER_LOAD_UBYTE_OFFEN]], 0, 8, implicit $exec
; CHECK: $vgpr0 = COPY [[V_BFE_I32_]]
AMDGPU/GlobalISel: Select llvm.amdgcn.raw.buffer.load Use intermediate instructions, unlike with buffer stores. This is necessary because of the need to have an internal way to distinguish between signed and unsigned extloads. This introduces some duplication and near duplication with the buffer store selection path. The store handling should maybe be moved into legalization to match and eliminate the duplication.
2019-09-19 12:29:20 +08:00
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call i8 @llvm.amdgcn.raw.buffer.load.i8(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
%zext = sext i8 %val to i32
%cast = bitcast i32 %zext to float
ret float %cast
}
; Waterfall for rsrc
define amdgpu_ps half @raw_buffer_load_f16__vgpr_rsrc__vgpr_voffset__sgpr_soffset(<4 x i32> %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f16__vgpr_rsrc__vgpr_voffset__sgpr_soffset
; CHECK: bb.1 (%ir-block.0):
; CHECK: successors: %bb.2(0x80000000)
; CHECK: liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4
; CHECK: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr4
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[COPY6:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
; CHECK: [[COPY7:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3
; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
; CHECK: bb.2:
; CHECK: successors: %bb.3(0x40000000), %bb.2(0x40000000)
; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE1]], [[COPY6]], implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE2]], [[COPY7]], implicit $exec
; CHECK: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
; CHECK: [[REG_SEQUENCE3:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_USHORT_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[COPY4]], [[REG_SEQUENCE3]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 2 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_]], implicit-def $exec, implicit-def $scc, implicit $exec
; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
; CHECK: S_CBRANCH_EXECNZ %bb.2, implicit $exec
; CHECK: bb.3:
; CHECK: successors: %bb.4(0x80000000)
; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
; CHECK: bb.4:
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_USHORT_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call half @llvm.amdgcn.raw.buffer.load.f16(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret half %val
}
; Waterfall for rsrc
define amdgpu_ps float @raw_buffer_load_i8__vgpr_rsrc__vgpr_voffset__sgpr_soffset(<4 x i32> %rsrc, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_i8__vgpr_rsrc__vgpr_voffset__sgpr_soffset
; CHECK: bb.1 (%ir-block.0):
; CHECK: successors: %bb.2(0x80000000)
; CHECK: liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4
; CHECK: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr4
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[COPY6:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
; CHECK: [[COPY7:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3
; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
; CHECK: bb.2:
; CHECK: successors: %bb.3(0x40000000), %bb.2(0x40000000)
; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE1]], [[COPY6]], implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE2]], [[COPY7]], implicit $exec
; CHECK: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
; CHECK: [[REG_SEQUENCE3:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_UBYTE_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_UBYTE_OFFEN [[COPY4]], [[REG_SEQUENCE3]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 1 from custom "TargetCustom7", addrspace 4)
; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_]], implicit-def $exec, implicit-def $scc, implicit $exec
; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
; CHECK: S_CBRANCH_EXECNZ %bb.2, implicit $exec
; CHECK: bb.3:
; CHECK: successors: %bb.4(0x80000000)
; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
; CHECK: bb.4:
[AMDGPU][GlobalISel] Compute known bits for zero-extending loads Implement computeKnownBitsForTargetInstr for G_AMDGPU_BUFFER_LOAD_UBYTE and G_AMDGPU_BUFFER_LOAD_USHORT. This allows generic combines to remove some unnecessary G_ANDs. Differential Revision: https://reviews.llvm.org/D89316
2020-10-13 21:40:27 +08:00
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_UBYTE_OFFEN]]
AMDGPU/GlobalISel: Select llvm.amdgcn.raw.buffer.load Use intermediate instructions, unlike with buffer stores. This is necessary because of the need to have an internal way to distinguish between signed and unsigned extloads. This introduces some duplication and near duplication with the buffer store selection path. The store handling should maybe be moved into legalization to match and eliminate the duplication.
2019-09-19 12:29:20 +08:00
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call i8 @llvm.amdgcn.raw.buffer.load.i8(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
%zext = zext i8 %val to i32
%cast = bitcast i32 %zext to float
ret float %cast
}
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vdpr_voffset__sgpr_soffset__voffset0(<4 x i32> inreg %rsrc, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vdpr_voffset__sgpr_soffset__voffset0
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFSET:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFSET [[REG_SEQUENCE]], [[COPY4]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFSET]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 0, i32 %soffset, i32 0)
ret float %val
}
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset__voffset4095(<4 x i32> inreg %rsrc, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset__voffset4095
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFSET:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFSET [[REG_SEQUENCE]], [[COPY4]], 4095, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7" + 4095, align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFSET]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 4095, i32 %soffset, i32 0)
ret float %val
}
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset__voffset4096(<4 x i32> inreg %rsrc, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset__voffset4096
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4096
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY5]], [[REG_SEQUENCE]], [[COPY4]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7" + 4096, align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 4096, i32 %soffset, i32 0)
ret float %val
}
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add16(<4 x i32> inreg %rsrc, i32 %voffset.base, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add16
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 16, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7" + 16, align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%voffset = add i32 %voffset.base, 16
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret float %val
}
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset__voffset_add4095(<4 x i32> inreg %rsrc, i32 %voffset.base, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset__voffset_add4095
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 4095, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7" + 4095, align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%voffset = add i32 %voffset.base, 4095
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret float %val
}
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset__voffset_add4096(<4 x i32> inreg %rsrc, i32 %voffset.base, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset__voffset_add4096
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4096
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
AMDGPU: Rename add/sub with carry out instructions The hardware has created a real mess in the naming for add/sub, which have been renamed basically every generation. Switch the carry out pseudos to have the gfx9/gfx10 names. We were using the original SI/CI v_add_i32/v_sub_i32 names. Later targets reintroduced these names as carryless instructions with a saturating clamp bit, which we do not define. Do this rename so we can unambiguously add these missing instructions. The carry-in versions should also be renamed, but at least those had a consistent _u32 name to begin with. The 16-bit instructions were also renamed, but aren't ambiguous. This does regress assembler error message quality in some cases. In mismatched wave32/wave64 situations, this will switch from "unsupported instruction" to "invalid operand", with the error pointing at the wrong position. I couldn't quite follow how the assembler selects these, but the previous behavior seemed accidental to me. It looked like there was a partial attempt to handle this which was never completed (i.e. there is an AMDGPUOperand::isBoolReg but it isn't used for anything).
2020-07-14 21:18:36 +08:00
; CHECK: %10:vgpr_32, dead %15:sreg_64_xexec = V_ADD_CO_U32_e64 [[COPY4]], [[COPY6]], 0, implicit $exec
AMDGPU/GlobalISel: Select llvm.amdgcn.raw.buffer.load Use intermediate instructions, unlike with buffer stores. This is necessary because of the need to have an internal way to distinguish between signed and unsigned extloads. This introduces some duplication and near duplication with the buffer store selection path. The store handling should maybe be moved into legalization to match and eliminate the duplication.
2019-09-19 12:29:20 +08:00
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN %10, [[REG_SEQUENCE]], [[COPY5]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7" + 4096, align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%voffset = add i32 %voffset.base, 4096
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret float %val
}
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_soffset4095(<4 x i32> inreg %rsrc, i32 %voffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_soffset4095
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4095
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[S_MOV_B32_]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 4095, i32 0)
ret float %val
}
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_soffset4096(<4 x i32> inreg %rsrc, i32 %voffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_soffset4096
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4096
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[S_MOV_B32_]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 4096, i32 0)
ret float %val
}
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_soffset_add16(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset.base) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_soffset_add16
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 16
AMDGPU/GlobalISel: Fix tablegen selection for scalar bin ops Fixes selection for scalar G_SMULH/G_UMULH. Also switches to using tablegen selected add/sub, which switch to the signed version of the opcode. This matches the current DAG behavior. We can't drop the manual selection for add/sub yet, because it's still both for VALU add/sub and for G_PTR_ADD.
2020-01-25 03:01:09 +08:00
; CHECK: [[S_ADD_I32_:%[0-9]+]]:sreg_32 = S_ADD_I32 [[COPY5]], [[S_MOV_B32_]], implicit-def $scc
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[S_ADD_I32_]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
AMDGPU/GlobalISel: Select llvm.amdgcn.raw.buffer.load Use intermediate instructions, unlike with buffer stores. This is necessary because of the need to have an internal way to distinguish between signed and unsigned extloads. This introduces some duplication and near duplication with the buffer store selection path. The store handling should maybe be moved into legalization to match and eliminate the duplication.
2019-09-19 12:29:20 +08:00
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%soffset = add i32 %soffset.base, 16
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret float %val
}
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_soffset_add4095(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset.base) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_soffset_add4095
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4095
AMDGPU/GlobalISel: Fix tablegen selection for scalar bin ops Fixes selection for scalar G_SMULH/G_UMULH. Also switches to using tablegen selected add/sub, which switch to the signed version of the opcode. This matches the current DAG behavior. We can't drop the manual selection for add/sub yet, because it's still both for VALU add/sub and for G_PTR_ADD.
2020-01-25 03:01:09 +08:00
; CHECK: [[S_ADD_I32_:%[0-9]+]]:sreg_32 = S_ADD_I32 [[COPY5]], [[S_MOV_B32_]], implicit-def $scc
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[S_ADD_I32_]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
AMDGPU/GlobalISel: Select llvm.amdgcn.raw.buffer.load Use intermediate instructions, unlike with buffer stores. This is necessary because of the need to have an internal way to distinguish between signed and unsigned extloads. This introduces some duplication and near duplication with the buffer store selection path. The store handling should maybe be moved into legalization to match and eliminate the duplication.
2019-09-19 12:29:20 +08:00
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%soffset = add i32 %soffset.base, 4095
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret float %val
}
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_soffset_add4096(<4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset.base) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_soffset_add4096
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4096
AMDGPU/GlobalISel: Fix tablegen selection for scalar bin ops Fixes selection for scalar G_SMULH/G_UMULH. Also switches to using tablegen selected add/sub, which switch to the signed version of the opcode. This matches the current DAG behavior. We can't drop the manual selection for add/sub yet, because it's still both for VALU add/sub and for G_PTR_ADD.
2020-01-25 03:01:09 +08:00
; CHECK: [[S_ADD_I32_:%[0-9]+]]:sreg_32 = S_ADD_I32 [[COPY5]], [[S_MOV_B32_]], implicit-def $scc
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE]], [[S_ADD_I32_]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
AMDGPU/GlobalISel: Select llvm.amdgcn.raw.buffer.load Use intermediate instructions, unlike with buffer stores. This is necessary because of the need to have an internal way to distinguish between signed and unsigned extloads. This introduces some duplication and near duplication with the buffer store selection path. The store handling should maybe be moved into legalization to match and eliminate the duplication.
2019-09-19 12:29:20 +08:00
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%soffset = add i32 %soffset.base, 4096
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret float %val
}
; An add of the offset is necessary, with a waterfall loop. Make sure the add is done outside of the waterfall loop.
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_soffset_add5000(<4 x i32> %rsrc, i32 %voffset, i32 inreg %soffset.base) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_soffset_add5000
; CHECK: bb.1 (%ir-block.0):
; CHECK: successors: %bb.2(0x80000000)
; CHECK: liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4
; CHECK: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr4
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 5000
AMDGPU/GlobalISel: Fix tablegen selection for scalar bin ops Fixes selection for scalar G_SMULH/G_UMULH. Also switches to using tablegen selected add/sub, which switch to the signed version of the opcode. This matches the current DAG behavior. We can't drop the manual selection for add/sub yet, because it's still both for VALU add/sub and for G_PTR_ADD.
2020-01-25 03:01:09 +08:00
; CHECK: [[S_ADD_I32_:%[0-9]+]]:sreg_32 = S_ADD_I32 [[COPY5]], [[S_MOV_B32_]], implicit-def $scc
AMDGPU/GlobalISel: Select llvm.amdgcn.raw.buffer.load Use intermediate instructions, unlike with buffer stores. This is necessary because of the need to have an internal way to distinguish between signed and unsigned extloads. This introduces some duplication and near duplication with the buffer store selection path. The store handling should maybe be moved into legalization to match and eliminate the duplication.
2019-09-19 12:29:20 +08:00
; CHECK: [[COPY6:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
; CHECK: [[COPY7:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3
; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
; CHECK: bb.2:
; CHECK: successors: %bb.3(0x40000000), %bb.2(0x40000000)
; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE1]], [[COPY6]], implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE2]], [[COPY7]], implicit $exec
; CHECK: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
; CHECK: [[REG_SEQUENCE3:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
AMDGPU/GlobalISel: Fix tablegen selection for scalar bin ops Fixes selection for scalar G_SMULH/G_UMULH. Also switches to using tablegen selected add/sub, which switch to the signed version of the opcode. This matches the current DAG behavior. We can't drop the manual selection for add/sub yet, because it's still both for VALU add/sub and for G_PTR_ADD.
2020-01-25 03:01:09 +08:00
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN [[COPY4]], [[REG_SEQUENCE3]], [[S_ADD_I32_]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7", align 1, addrspace 4)
AMDGPU/GlobalISel: Select llvm.amdgcn.raw.buffer.load Use intermediate instructions, unlike with buffer stores. This is necessary because of the need to have an internal way to distinguish between signed and unsigned extloads. This introduces some duplication and near duplication with the buffer store selection path. The store handling should maybe be moved into legalization to match and eliminate the duplication.
2019-09-19 12:29:20 +08:00
; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_]], implicit-def $exec, implicit-def $scc, implicit $exec
; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
; CHECK: S_CBRANCH_EXECNZ %bb.2, implicit $exec
; CHECK: bb.3:
; CHECK: successors: %bb.4(0x80000000)
; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
; CHECK: bb.4:
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%soffset = add i32 %soffset.base, 5000
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret float %val
}
; An add of the offset is necessary, with a waterfall loop. Make sure the add is done outside of the waterfall loop.
define amdgpu_ps float @raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add5000(<4 x i32> %rsrc, i32 %voffset.base, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_load_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add5000
; CHECK: bb.1 (%ir-block.0):
; CHECK: successors: %bb.2(0x80000000)
; CHECK: liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4
; CHECK: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr4
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4096
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
AMDGPU: Rename add/sub with carry out instructions The hardware has created a real mess in the naming for add/sub, which have been renamed basically every generation. Switch the carry out pseudos to have the gfx9/gfx10 names. We were using the original SI/CI v_add_i32/v_sub_i32 names. Later targets reintroduced these names as carryless instructions with a saturating clamp bit, which we do not define. Do this rename so we can unambiguously add these missing instructions. The carry-in versions should also be renamed, but at least those had a consistent _u32 name to begin with. The 16-bit instructions were also renamed, but aren't ambiguous. This does regress assembler error message quality in some cases. In mismatched wave32/wave64 situations, this will switch from "unsupported instruction" to "invalid operand", with the error pointing at the wrong position. I couldn't quite follow how the assembler selects these, but the previous behavior seemed accidental to me. It looked like there was a partial attempt to handle this which was never completed (i.e. there is an AMDGPUOperand::isBoolReg but it isn't used for anything).
2020-07-14 21:18:36 +08:00
; CHECK: %13:vgpr_32, dead %35:sreg_64_xexec = V_ADD_CO_U32_e64 [[COPY4]], [[COPY6]], 0, implicit $exec
AMDGPU/GlobalISel: Select llvm.amdgcn.raw.buffer.load Use intermediate instructions, unlike with buffer stores. This is necessary because of the need to have an internal way to distinguish between signed and unsigned extloads. This introduces some duplication and near duplication with the buffer store selection path. The store handling should maybe be moved into legalization to match and eliminate the duplication.
2019-09-19 12:29:20 +08:00
; CHECK: [[COPY7:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
; CHECK: [[COPY8:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3
; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
; CHECK: bb.2:
; CHECK: successors: %bb.3(0x40000000), %bb.2(0x40000000)
; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE1]], [[COPY7]], implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY8]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY8]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE2]], [[COPY8]], implicit $exec
; CHECK: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
; CHECK: [[REG_SEQUENCE3:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
; CHECK: [[BUFFER_LOAD_DWORD_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_DWORD_OFFEN %13, [[REG_SEQUENCE3]], [[COPY5]], 904, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from custom "TargetCustom7" + 5000, align 1, addrspace 4)
; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_]], implicit-def $exec, implicit-def $scc, implicit $exec
; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
; CHECK: S_CBRANCH_EXECNZ %bb.2, implicit $exec
; CHECK: bb.3:
; CHECK: successors: %bb.4(0x80000000)
; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
; CHECK: bb.4:
; CHECK: $vgpr0 = COPY [[BUFFER_LOAD_DWORD_OFFEN]]
; CHECK: SI_RETURN_TO_EPILOG implicit $vgpr0
%voffset = add i32 %voffset.base, 5000
%val = call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret float %val
}
declare float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32>, i32, i32, i32 immarg)
declare <2 x float> @llvm.amdgcn.raw.buffer.load.v2f32(<4 x i32>, i32, i32, i32 immarg)
declare <3 x float> @llvm.amdgcn.raw.buffer.load.v3f32(<4 x i32>, i32, i32, i32 immarg)
declare <4 x float> @llvm.amdgcn.raw.buffer.load.v4f32(<4 x i32>, i32, i32, i32 immarg)
declare half @llvm.amdgcn.raw.buffer.load.f16(<4 x i32>, i32, i32, i32 immarg)
declare <2 x half> @llvm.amdgcn.raw.buffer.load.v2f16(<4 x i32>, i32, i32, i32 immarg)
declare <3 x half> @llvm.amdgcn.raw.buffer.load.v3f16(<4 x i32>, i32, i32, i32 immarg)
declare <4 x half> @llvm.amdgcn.raw.buffer.load.v4f16(<4 x i32>, i32, i32, i32 immarg)
declare i8 @llvm.amdgcn.raw.buffer.load.i8(<4 x i32>, i32, i32, i32 immarg)