diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp index 1c1b283e7443..93ffa0d9a54c 100644 --- a/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp +++ b/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp @@ -863,10 +863,11 @@ RegBankLegalizeRules::RegBankLegalizeRules(const GCNSubtarget &_ST, .Div(S64, {{Vgpr64}, {Vgpr64, Vgpr64, SgprV4S32_WF, Vgpr32, Vgpr32, Sgpr32_WF}}); - addRulesForGOpcs({G_AMDGPU_BUFFER_ATOMIC_SWAP, G_AMDGPU_BUFFER_ATOMIC_UMAX, - G_AMDGPU_BUFFER_ATOMIC_UMIN, G_AMDGPU_BUFFER_ATOMIC_SMAX, - G_AMDGPU_BUFFER_ATOMIC_SMIN, G_AMDGPU_BUFFER_ATOMIC_FMAX, - G_AMDGPU_BUFFER_ATOMIC_FMIN}, + addRulesForGOpcs({G_AMDGPU_BUFFER_ATOMIC_ADD, G_AMDGPU_BUFFER_ATOMIC_FMAX, + G_AMDGPU_BUFFER_ATOMIC_FMIN, G_AMDGPU_BUFFER_ATOMIC_SMAX, + G_AMDGPU_BUFFER_ATOMIC_SMIN, G_AMDGPU_BUFFER_ATOMIC_SUB, + G_AMDGPU_BUFFER_ATOMIC_SWAP, G_AMDGPU_BUFFER_ATOMIC_UMAX, + G_AMDGPU_BUFFER_ATOMIC_UMIN}, Standard) .Div(S32, {{Vgpr32}, {Vgpr32, SgprV4S32_WF, Vgpr32, Vgpr32, Sgpr32_WF}}) .Div(S64, {{Vgpr64}, {Vgpr64, SgprV4S32_WF, Vgpr32, Vgpr32, Sgpr32_WF}}); diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.atomic.add.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.atomic.add.ll index 79a92918bfe8..9a9b5278b788 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.atomic.add.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.atomic.add.ll @@ -1,7 +1,7 @@ ; 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 -o - %s | FileCheck -check-prefixes=GFX8 %s -; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1200 -stop-after=instruction-select -o - %s | FileCheck --check-prefixes=GFX12,GFX1200 %s -; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1250 -stop-after=instruction-select -o - %s | FileCheck --check-prefixes=GFX12,GFX1250 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -stop-after=instruction-select -o - %s | FileCheck -check-prefixes=GFX8 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1200 -stop-after=instruction-select -o - %s | FileCheck --check-prefixes=GFX12,GFX1200 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1250 -stop-after=instruction-select -o - %s | FileCheck --check-prefixes=GFX12,GFX1250 %s ; Natural mapping define amdgpu_ps float @raw_buffer_atomic_add_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) { @@ -41,7 +41,7 @@ define amdgpu_ps float @raw_buffer_atomic_add_i32__vgpr_val__sgpr_rsrc__vgpr_vof ret float %cast } -define amdgpu_ps float @raw_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) { +define amdgpu_ps void @raw_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) { ; GFX8-LABEL: name: raw_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset ; GFX8: bb.1 (%ir-block.0): ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 @@ -54,9 +54,8 @@ define amdgpu_ps float @raw_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vg ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 ; GFX8-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 ; GFX8-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 - ; GFX8-NEXT: [[BUFFER_ATOMIC_ADD_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) - ; GFX8-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_OFFEN_RTN]] - ; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; GFX8-NEXT: BUFFER_ATOMIC_ADD_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: S_ENDPGM 0 ; ; GFX12-LABEL: name: raw_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset ; GFX12: bb.1 (%ir-block.0): @@ -70,12 +69,10 @@ define amdgpu_ps float @raw_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vg ; GFX12-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 ; GFX12-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 - ; GFX12-NEXT: [[BUFFER_ATOMIC_ADD_VBUFFER_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_VBUFFER_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) - ; GFX12-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_VBUFFER_OFFEN_RTN]] - ; GFX12-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; GFX12-NEXT: BUFFER_ATOMIC_ADD_VBUFFER_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX12-NEXT: S_ENDPGM 0 %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.add.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0) - %cast = bitcast i32 %ret to float - ret float %cast + ret void } define amdgpu_ps <2 x float> @raw_buffer_atomic_add_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i64 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) { diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.atomic.sub.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.atomic.sub.ll new file mode 100644 index 000000000000..ceba7ba6d57c --- /dev/null +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.atomic.sub.ll @@ -0,0 +1,624 @@ +; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -stop-after=instruction-select -o - %s | FileCheck -check-prefixes=GFX8 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1200 -stop-after=instruction-select -o - %s | FileCheck --check-prefixes=GFX12,GFX1200 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1250 -stop-after=instruction-select -o - %s | FileCheck --check-prefixes=GFX12,GFX1250 %s + +; Natural mapping +define amdgpu_ps float @raw_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) { + ; GFX8-LABEL: name: raw_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX8-NEXT: [[BUFFER_ATOMIC_SUB_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_OFFEN_RTN]] + ; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; + ; GFX12-LABEL: name: raw_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX12: bb.1 (%ir-block.0): + ; GFX12-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; GFX12-NEXT: {{ $}} + ; GFX12-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX12-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX12-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX12-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX12-NEXT: [[BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX12-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN]] + ; GFX12-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.sub.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i32 %ret to float + ret float %cast +} + +define amdgpu_ps void @raw_buffer_atomic_sub_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) { + ; GFX8-LABEL: name: raw_buffer_atomic_sub_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX8-NEXT: BUFFER_ATOMIC_SUB_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: S_ENDPGM 0 + ; + ; GFX12-LABEL: name: raw_buffer_atomic_sub_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX12: bb.1 (%ir-block.0): + ; GFX12-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; GFX12-NEXT: {{ $}} + ; GFX12-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX12-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX12-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX12-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX12-NEXT: BUFFER_ATOMIC_SUB_VBUFFER_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX12-NEXT: S_ENDPGM 0 + %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.sub.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0) + ret void +} + +define amdgpu_ps <2 x float> @raw_buffer_atomic_sub_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i64 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) { + ; GFX8-LABEL: name: raw_buffer_atomic_sub_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX8-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX8-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX8-NEXT: [[BUFFER_ATOMIC_SUB_X2_OFFEN_RTN:%[0-9]+]]:vreg_64 = BUFFER_ATOMIC_SUB_X2_OFFEN_RTN [[REG_SEQUENCE]], [[COPY6]], [[REG_SEQUENCE1]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 8) + ; GFX8-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_OFFEN_RTN]].sub0 + ; GFX8-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_OFFEN_RTN]].sub1 + ; GFX8-NEXT: $vgpr0 = COPY [[COPY8]] + ; GFX8-NEXT: $vgpr1 = COPY [[COPY9]] + ; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0, implicit $vgpr1 + ; + ; GFX1200-LABEL: name: raw_buffer_atomic_sub_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX1200: bb.1 (%ir-block.0): + ; GFX1200-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1200-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1200-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; GFX1200-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1200-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1200-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1200-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1200-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1200-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1200-NEXT: [[BUFFER_ATOMIC_SUB_X2_VBUFFER_OFFEN_RTN:%[0-9]+]]:vreg_64 = BUFFER_ATOMIC_SUB_X2_VBUFFER_OFFEN_RTN [[REG_SEQUENCE]], [[COPY6]], [[REG_SEQUENCE1]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 8) + ; GFX1200-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_VBUFFER_OFFEN_RTN]].sub0 + ; GFX1200-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_VBUFFER_OFFEN_RTN]].sub1 + ; GFX1200-NEXT: $vgpr0 = COPY [[COPY8]] + ; GFX1200-NEXT: $vgpr1 = COPY [[COPY9]] + ; GFX1200-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0, implicit $vgpr1 + ; + ; GFX1250-LABEL: name: raw_buffer_atomic_sub_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX1250: bb.1 (%ir-block.0): + ; GFX1250-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1250-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1250-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; GFX1250-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1250-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1250-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1250-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1250-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1250-NEXT: [[BUFFER_ATOMIC_SUB_X2_VBUFFER_OFFEN_RTN:%[0-9]+]]:vreg_64_align2 = BUFFER_ATOMIC_SUB_X2_VBUFFER_OFFEN_RTN [[REG_SEQUENCE]], [[COPY6]], [[REG_SEQUENCE1]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 8) + ; GFX1250-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_VBUFFER_OFFEN_RTN]].sub0 + ; GFX1250-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_VBUFFER_OFFEN_RTN]].sub1 + ; GFX1250-NEXT: $vgpr0 = COPY [[COPY8]] + ; GFX1250-NEXT: $vgpr1 = COPY [[COPY9]] + ; GFX1250-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0, implicit $vgpr1 + %ret = call i64 @llvm.amdgcn.raw.buffer.atomic.sub.i64(i64 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i64 %ret to <2 x float> + ret <2 x float> %cast +} + +define amdgpu_ps void @raw_buffer_atomic_sub_i64_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i64 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) { + ; GFX8-LABEL: name: raw_buffer_atomic_sub_i64_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX8-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX8-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX8-NEXT: BUFFER_ATOMIC_SUB_X2_OFFEN [[REG_SEQUENCE]], [[COPY6]], [[REG_SEQUENCE1]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 8) + ; GFX8-NEXT: S_ENDPGM 0 + ; + ; GFX1200-LABEL: name: raw_buffer_atomic_sub_i64_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX1200: bb.1 (%ir-block.0): + ; GFX1200-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1200-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1200-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; GFX1200-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1200-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1200-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1200-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1200-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1200-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1200-NEXT: BUFFER_ATOMIC_SUB_X2_VBUFFER_OFFEN [[REG_SEQUENCE]], [[COPY6]], [[REG_SEQUENCE1]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 8) + ; GFX1200-NEXT: S_ENDPGM 0 + ; + ; GFX1250-LABEL: name: raw_buffer_atomic_sub_i64_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX1250: bb.1 (%ir-block.0): + ; GFX1250-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1250-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1250-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; GFX1250-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1250-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1250-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1250-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1250-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1250-NEXT: BUFFER_ATOMIC_SUB_X2_VBUFFER_OFFEN [[REG_SEQUENCE]], [[COPY6]], [[REG_SEQUENCE1]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 8) + ; GFX1250-NEXT: S_ENDPGM 0 + %ret = call i64 @llvm.amdgcn.raw.buffer.atomic.sub.i64(i64 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0) + ret void +} + +; All operands need regbank legalization +define amdgpu_ps float @raw_buffer_atomic_sub_i32__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset(i32 inreg %val, <4 x i32> %rsrc, i32 inreg %voffset, i32 %soffset) { + ; GFX8-LABEL: name: raw_buffer_atomic_sub_i32__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: successors: %bb.2(0x80000000) + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; GFX8-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; GFX8-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; GFX8-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.2: + ; GFX8-NEXT: successors: %bb.3(0x80000000) + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; GFX8-NEXT: [[REG_SEQUENCE1:%[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 + ; GFX8-NEXT: [[COPY9:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; GFX8-NEXT: [[COPY10:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; GFX8-NEXT: [[COPY11:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; GFX8-NEXT: [[COPY12:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; GFX8-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY11]], [[COPY9]], implicit $exec + ; GFX8-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY12]], [[COPY10]], implicit $exec + ; GFX8-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]], implicit $exec + ; GFX8-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY6]], implicit $exec + ; GFX8-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; GFX8-NEXT: [[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 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.3: + ; GFX8-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[BUFFER_ATOMIC_SUB_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_OFFEN_RTN [[COPY7]], [[COPY8]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc + ; GFX8-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.4: + ; GFX8-NEXT: successors: %bb.5(0x80000000) + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: $exec = S_MOV_B64_term [[S_MOV_B64_]] + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.5: + ; GFX8-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_OFFEN_RTN]] + ; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; + ; GFX1200-LABEL: name: raw_buffer_atomic_sub_i32__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; GFX1200: bb.1 (%ir-block.0): + ; GFX1200-NEXT: successors: %bb.2(0x80000000) + ; GFX1200-NEXT: liveins: $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1200-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1200-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1200-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1200-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX1200-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1200-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; GFX1200-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; GFX1200-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; GFX1200-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_MOV_B32 $exec_lo + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.2: + ; GFX1200-NEXT: successors: %bb.3(0x80000000) + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; GFX1200-NEXT: [[REG_SEQUENCE1:%[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 + ; GFX1200-NEXT: [[COPY9:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; GFX1200-NEXT: [[COPY10:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; GFX1200-NEXT: [[COPY11:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; GFX1200-NEXT: [[COPY12:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; GFX1200-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY11]], [[COPY9]], implicit $exec + ; GFX1200-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY12]], [[COPY10]], implicit $exec + ; GFX1200-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]], implicit $exec + ; GFX1200-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY6]], implicit $exec + ; GFX1200-NEXT: [[S_AND_B32_1:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[S_AND_B32_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; GFX1200-NEXT: [[S_AND_SAVEEXEC_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_SAVEEXEC_B32 killed [[S_AND_B32_1]], implicit-def $exec, implicit-def $scc, implicit $exec + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.3: + ; GFX1200-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN [[COPY7]], [[COPY8]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1200-NEXT: $exec_lo = S_XOR_B32_term $exec_lo, [[S_AND_SAVEEXEC_B32_]], implicit-def $scc + ; GFX1200-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.4: + ; GFX1200-NEXT: successors: %bb.5(0x80000000) + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: $exec_lo = S_MOV_B32_term [[S_MOV_B32_]] + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.5: + ; GFX1200-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN]] + ; GFX1200-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; + ; GFX1250-LABEL: name: raw_buffer_atomic_sub_i32__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; GFX1250: bb.1 (%ir-block.0): + ; GFX1250-NEXT: successors: %bb.2(0x80000000) + ; GFX1250-NEXT: liveins: $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1250-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1250-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1250-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1250-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX1250-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128_align2 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1250-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; GFX1250-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; GFX1250-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_MOV_B32 $exec_lo + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.2: + ; GFX1250-NEXT: successors: %bb.3(0x80000000) + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; GFX1250-NEXT: [[REG_SEQUENCE1:%[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 + ; GFX1250-NEXT: [[COPY9:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; GFX1250-NEXT: [[COPY10:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; GFX1250-NEXT: [[COPY11:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; GFX1250-NEXT: [[COPY12:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; GFX1250-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY11]], [[COPY9]], implicit $exec + ; GFX1250-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY12]], [[COPY10]], implicit $exec + ; GFX1250-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]], implicit $exec + ; GFX1250-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY6]], implicit $exec + ; GFX1250-NEXT: [[S_AND_B32_1:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[S_AND_B32_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; GFX1250-NEXT: [[S_AND_SAVEEXEC_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_SAVEEXEC_B32 killed [[S_AND_B32_1]], implicit-def $exec, implicit-def $scc, implicit $exec + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.3: + ; GFX1250-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN [[COPY7]], [[COPY8]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1250-NEXT: $exec_lo = S_XOR_B32_term $exec_lo, [[S_AND_SAVEEXEC_B32_]], implicit-def $scc + ; GFX1250-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.4: + ; GFX1250-NEXT: successors: %bb.5(0x80000000) + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: $exec_lo = S_MOV_B32_term [[S_MOV_B32_]] + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.5: + ; GFX1250-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN]] + ; GFX1250-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.sub.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i32 %ret to float + ret float %cast +} + +; All operands need regbank legalization +define amdgpu_ps void @raw_buffer_atomic_sub_i32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset(i32 inreg %val, <4 x i32> %rsrc, i32 inreg %voffset, i32 %soffset) { + ; GFX8-LABEL: name: raw_buffer_atomic_sub_i32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: successors: %bb.2(0x80000000) + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; GFX8-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; GFX8-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; GFX8-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.2: + ; GFX8-NEXT: successors: %bb.3(0x80000000) + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; GFX8-NEXT: [[REG_SEQUENCE1:%[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 + ; GFX8-NEXT: [[COPY9:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; GFX8-NEXT: [[COPY10:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; GFX8-NEXT: [[COPY11:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; GFX8-NEXT: [[COPY12:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; GFX8-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY11]], [[COPY9]], implicit $exec + ; GFX8-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY12]], [[COPY10]], implicit $exec + ; GFX8-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]], implicit $exec + ; GFX8-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY6]], implicit $exec + ; GFX8-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; GFX8-NEXT: [[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 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.3: + ; GFX8-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: BUFFER_ATOMIC_SUB_OFFEN [[COPY7]], [[COPY8]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc + ; GFX8-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.4: + ; GFX8-NEXT: successors: %bb.5(0x80000000) + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: $exec = S_MOV_B64_term [[S_MOV_B64_]] + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.5: + ; GFX8-NEXT: S_ENDPGM 0 + ; + ; GFX1200-LABEL: name: raw_buffer_atomic_sub_i32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; GFX1200: bb.1 (%ir-block.0): + ; GFX1200-NEXT: successors: %bb.2(0x80000000) + ; GFX1200-NEXT: liveins: $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1200-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1200-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1200-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1200-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX1200-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1200-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; GFX1200-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; GFX1200-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; GFX1200-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_MOV_B32 $exec_lo + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.2: + ; GFX1200-NEXT: successors: %bb.3(0x80000000) + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; GFX1200-NEXT: [[REG_SEQUENCE1:%[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 + ; GFX1200-NEXT: [[COPY9:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; GFX1200-NEXT: [[COPY10:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; GFX1200-NEXT: [[COPY11:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; GFX1200-NEXT: [[COPY12:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; GFX1200-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY11]], [[COPY9]], implicit $exec + ; GFX1200-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY12]], [[COPY10]], implicit $exec + ; GFX1200-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]], implicit $exec + ; GFX1200-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY6]], implicit $exec + ; GFX1200-NEXT: [[S_AND_B32_1:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[S_AND_B32_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; GFX1200-NEXT: [[S_AND_SAVEEXEC_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_SAVEEXEC_B32 killed [[S_AND_B32_1]], implicit-def $exec, implicit-def $scc, implicit $exec + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.3: + ; GFX1200-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: BUFFER_ATOMIC_SUB_VBUFFER_OFFEN [[COPY7]], [[COPY8]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1200-NEXT: $exec_lo = S_XOR_B32_term $exec_lo, [[S_AND_SAVEEXEC_B32_]], implicit-def $scc + ; GFX1200-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.4: + ; GFX1200-NEXT: successors: %bb.5(0x80000000) + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: $exec_lo = S_MOV_B32_term [[S_MOV_B32_]] + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.5: + ; GFX1200-NEXT: S_ENDPGM 0 + ; + ; GFX1250-LABEL: name: raw_buffer_atomic_sub_i32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; GFX1250: bb.1 (%ir-block.0): + ; GFX1250-NEXT: successors: %bb.2(0x80000000) + ; GFX1250-NEXT: liveins: $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1250-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1250-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1250-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1250-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX1250-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128_align2 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1250-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; GFX1250-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; GFX1250-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_MOV_B32 $exec_lo + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.2: + ; GFX1250-NEXT: successors: %bb.3(0x80000000) + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; GFX1250-NEXT: [[REG_SEQUENCE1:%[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 + ; GFX1250-NEXT: [[COPY9:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; GFX1250-NEXT: [[COPY10:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; GFX1250-NEXT: [[COPY11:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; GFX1250-NEXT: [[COPY12:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; GFX1250-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY11]], [[COPY9]], implicit $exec + ; GFX1250-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY12]], [[COPY10]], implicit $exec + ; GFX1250-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]], implicit $exec + ; GFX1250-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY6]], implicit $exec + ; GFX1250-NEXT: [[S_AND_B32_1:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[S_AND_B32_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; GFX1250-NEXT: [[S_AND_SAVEEXEC_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_SAVEEXEC_B32 killed [[S_AND_B32_1]], implicit-def $exec, implicit-def $scc, implicit $exec + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.3: + ; GFX1250-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: BUFFER_ATOMIC_SUB_VBUFFER_OFFEN [[COPY7]], [[COPY8]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1250-NEXT: $exec_lo = S_XOR_B32_term $exec_lo, [[S_AND_SAVEEXEC_B32_]], implicit-def $scc + ; GFX1250-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.4: + ; GFX1250-NEXT: successors: %bb.5(0x80000000) + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: $exec_lo = S_MOV_B32_term [[S_MOV_B32_]] + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.5: + ; GFX1250-NEXT: S_ENDPGM 0 + %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.sub.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0) + ret void +} + +define amdgpu_ps float @raw_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add4095(i32 %val, <4 x i32> inreg %rsrc, i32 %voffset.base, i32 inreg %soffset) { + ; GFX8-LABEL: name: raw_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add4095 + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX8-NEXT: [[BUFFER_ATOMIC_SUB_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 4095, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_OFFEN_RTN]] + ; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; + ; GFX1200-LABEL: name: raw_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add4095 + ; GFX1200: bb.1 (%ir-block.0): + ; GFX1200-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1200-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1200-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1200-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1200-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1200-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1200-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1200-NEXT: [[BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 4095, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1200-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN]] + ; GFX1200-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; + ; GFX1250-LABEL: name: raw_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add4095 + ; GFX1250: bb.1 (%ir-block.0): + ; GFX1250-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1250-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1250-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1250-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1250-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1250-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1250-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1250-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4095 + ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]] + ; GFX1250-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[COPY5]], [[COPY7]], 0, implicit $exec + ; GFX1250-NEXT: [[BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN [[COPY]], [[V_ADD_U32_e64_]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1250-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN]] + ; GFX1250-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %voffset = add i32 %voffset.base, 4095 + %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.sub.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i32 %ret to float + ret float %cast +} + +; Natural mapping + slc +define amdgpu_ps float @raw_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc(i32 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) { + ; GFX8-LABEL: name: raw_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX8-NEXT: [[BUFFER_ATOMIC_SUB_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 3, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_OFFEN_RTN]] + ; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; + ; GFX12-LABEL: name: raw_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc + ; GFX12: bb.1 (%ir-block.0): + ; GFX12-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; GFX12-NEXT: {{ $}} + ; GFX12-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX12-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX12-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX12-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX12-NEXT: [[BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 3, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX12-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_VBUFFER_OFFEN_RTN]] + ; GFX12-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.sub.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 2) + %cast = bitcast i32 %ret to float + ret float %cast +} + +declare i32 @llvm.amdgcn.raw.buffer.atomic.sub.i32(i32, <4 x i32>, i32, i32, i32 immarg) #0 +declare i64 @llvm.amdgcn.raw.buffer.atomic.sub.i64(i64, <4 x i32>, i32, i32, i32 immarg) #0 +attributes #0 = { nounwind } diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.ptr.buffer.atomic.add.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.ptr.buffer.atomic.add.ll index 7760a8d22f79..1edd74af01f0 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.ptr.buffer.atomic.add.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.ptr.buffer.atomic.add.ll @@ -1,5 +1,5 @@ ; 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 -o - %s | FileCheck %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -stop-after=instruction-select -o - %s | FileCheck %s ; Natural mapping define amdgpu_ps float @raw_ptr_buffer_atomic_add_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) { @@ -23,7 +23,7 @@ define amdgpu_ps float @raw_ptr_buffer_atomic_add_i32__vgpr_val__sgpr_rsrc__vgpr ret float %cast } -define amdgpu_ps float @raw_ptr_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) { +define amdgpu_ps void @raw_ptr_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) { ; CHECK-LABEL: name: raw_ptr_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset ; CHECK: bb.1 (%ir-block.0): ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 @@ -36,12 +36,10 @@ define amdgpu_ps float @raw_ptr_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc ; CHECK-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 ; CHECK-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 - ; CHECK-NEXT: [[BUFFER_ATOMIC_ADD_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) - ; CHECK-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_OFFEN_RTN]] - ; CHECK-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; CHECK-NEXT: BUFFER_ATOMIC_ADD_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: S_ENDPGM 0 %ret = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.add.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0) - %cast = bitcast i32 %ret to float - ret float %cast + ret void } define amdgpu_ps <2 x float> @raw_ptr_buffer_atomic_add_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i64 %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) { diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.ptr.buffer.atomic.sub.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.ptr.buffer.atomic.sub.ll new file mode 100644 index 000000000000..6a45fecc89c5 --- /dev/null +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.ptr.buffer.atomic.sub.ll @@ -0,0 +1,255 @@ +; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -stop-after=instruction-select -o - %s | FileCheck %s + +; Natural mapping +define amdgpu_ps float @raw_ptr_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) { + ; CHECK-LABEL: name: raw_ptr_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; CHECK-NEXT: [[BUFFER_ATOMIC_SUB_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_OFFEN_RTN]] + ; CHECK-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %ret = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.sub.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i32 %ret to float + ret float %cast +} + +define amdgpu_ps void @raw_ptr_buffer_atomic_sub_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) { + ; CHECK-LABEL: name: raw_ptr_buffer_atomic_sub_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; CHECK-NEXT: BUFFER_ATOMIC_SUB_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: S_ENDPGM 0 + %ret = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.sub.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0) + ret void +} + +define amdgpu_ps <2 x float> @raw_ptr_buffer_atomic_sub_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i64 %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) { + ; CHECK-LABEL: name: raw_ptr_buffer_atomic_sub_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; CHECK-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; CHECK-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; CHECK-NEXT: [[BUFFER_ATOMIC_SUB_X2_OFFEN_RTN:%[0-9]+]]:vreg_64 = BUFFER_ATOMIC_SUB_X2_OFFEN_RTN [[REG_SEQUENCE]], [[COPY6]], [[REG_SEQUENCE1]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_OFFEN_RTN]].sub0 + ; CHECK-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_OFFEN_RTN]].sub1 + ; CHECK-NEXT: $vgpr0 = COPY [[COPY8]] + ; CHECK-NEXT: $vgpr1 = COPY [[COPY9]] + ; CHECK-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0, implicit $vgpr1 + %ret = call i64 @llvm.amdgcn.raw.ptr.buffer.atomic.sub.i64(i64 %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i64 %ret to <2 x float> + ret <2 x float> %cast +} + +define amdgpu_ps void @raw_ptr_buffer_atomic_sub_i64_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i64 %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) { + ; CHECK-LABEL: name: raw_ptr_buffer_atomic_sub_i64_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; CHECK-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; CHECK-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; CHECK-NEXT: BUFFER_ATOMIC_SUB_X2_OFFEN [[REG_SEQUENCE]], [[COPY6]], [[REG_SEQUENCE1]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: S_ENDPGM 0 + %ret = call i64 @llvm.amdgcn.raw.ptr.buffer.atomic.sub.i64(i64 %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0) + ret void +} + +; All operands need regbank legalization +define amdgpu_ps float @raw_ptr_buffer_atomic_sub_i32__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset(i32 inreg %val, ptr addrspace(8) %rsrc, i32 inreg %voffset, i32 %soffset) { + ; CHECK-LABEL: name: raw_ptr_buffer_atomic_sub_i32__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: successors: %bb.2(0x80000000) + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; CHECK-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; CHECK-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; CHECK-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.2: + ; CHECK-NEXT: successors: %bb.3(0x80000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; CHECK-NEXT: [[REG_SEQUENCE1:%[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-NEXT: [[COPY9:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; CHECK-NEXT: [[COPY10:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; CHECK-NEXT: [[COPY11:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; CHECK-NEXT: [[COPY12:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; CHECK-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY11]], [[COPY9]], implicit $exec + ; CHECK-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY12]], [[COPY10]], implicit $exec + ; CHECK-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]], implicit $exec + ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY6]], implicit $exec + ; CHECK-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; CHECK-NEXT: [[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-NEXT: {{ $}} + ; CHECK-NEXT: bb.3: + ; CHECK-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[BUFFER_ATOMIC_SUB_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_OFFEN_RTN [[COPY7]], [[COPY8]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc + ; CHECK-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.4: + ; CHECK-NEXT: successors: %bb.5(0x80000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: $exec = S_MOV_B64_term [[S_MOV_B64_]] + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.5: + ; CHECK-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_OFFEN_RTN]] + ; CHECK-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %ret = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.sub.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i32 %ret to float + ret float %cast +} + +; All operands need regbank legalization +define amdgpu_ps void @raw_ptr_buffer_atomic_sub_i32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset(i32 inreg %val, ptr addrspace(8) %rsrc, i32 inreg %voffset, i32 %soffset) { + ; CHECK-LABEL: name: raw_ptr_buffer_atomic_sub_i32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: successors: %bb.2(0x80000000) + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; CHECK-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; CHECK-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; CHECK-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.2: + ; CHECK-NEXT: successors: %bb.3(0x80000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; CHECK-NEXT: [[REG_SEQUENCE1:%[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-NEXT: [[COPY9:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; CHECK-NEXT: [[COPY10:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; CHECK-NEXT: [[COPY11:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; CHECK-NEXT: [[COPY12:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; CHECK-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY11]], [[COPY9]], implicit $exec + ; CHECK-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY12]], [[COPY10]], implicit $exec + ; CHECK-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]], implicit $exec + ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY6]], implicit $exec + ; CHECK-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; CHECK-NEXT: [[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-NEXT: {{ $}} + ; CHECK-NEXT: bb.3: + ; CHECK-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: BUFFER_ATOMIC_SUB_OFFEN [[COPY7]], [[COPY8]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc + ; CHECK-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.4: + ; CHECK-NEXT: successors: %bb.5(0x80000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: $exec = S_MOV_B64_term [[S_MOV_B64_]] + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.5: + ; CHECK-NEXT: S_ENDPGM 0 + %ret = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.sub.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0) + ret void +} + +define amdgpu_ps float @raw_ptr_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add4095(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %voffset.base, i32 inreg %soffset) { + ; CHECK-LABEL: name: raw_ptr_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add4095 + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; CHECK-NEXT: [[BUFFER_ATOMIC_SUB_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 4095, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_OFFEN_RTN]] + ; CHECK-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %voffset = add i32 %voffset.base, 4095 + %ret = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.sub.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i32 %ret to float + ret float %cast +} + +; Natural mapping + slc +define amdgpu_ps float @raw_ptr_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) { + ; CHECK-LABEL: name: raw_ptr_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; CHECK-NEXT: [[BUFFER_ATOMIC_SUB_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 3, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_OFFEN_RTN]] + ; CHECK-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %ret = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.sub.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 2) + %cast = bitcast i32 %ret to float + ret float %cast +} + +declare i32 @llvm.amdgcn.raw.ptr.buffer.atomic.sub.i32(i32, ptr addrspace(8), i32, i32, i32 immarg) #0 +declare i64 @llvm.amdgcn.raw.ptr.buffer.atomic.sub.i64(i64, ptr addrspace(8), i32, i32, i32 immarg) #0 + +attributes #0 = { nounwind } diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.buffer.atomic.add.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.buffer.atomic.add.ll index 484639a23d28..3cfd1102c855 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.buffer.atomic.add.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.buffer.atomic.add.ll @@ -1,7 +1,7 @@ ; 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 -o - %s | FileCheck -check-prefix=GFX8 %s -; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1200 -stop-after=instruction-select -o - %s | FileCheck -check-prefix=GFX1200 %s -; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1250 -stop-after=instruction-select -o - %s | FileCheck -check-prefix=GFX1250 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -stop-after=instruction-select -o - %s | FileCheck -check-prefix=GFX8 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1200 -stop-after=instruction-select -o - %s | FileCheck -check-prefix=GFX1200 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1250 -stop-after=instruction-select -o - %s | FileCheck -check-prefix=GFX1250 %s ; Natural mapping define amdgpu_ps float @struct_buffer_atomic_add_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { @@ -63,7 +63,7 @@ define amdgpu_ps float @struct_buffer_atomic_add_i32__vgpr_val__sgpr_rsrc__vgpr_ ret float %cast } -define amdgpu_ps float @struct_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { +define amdgpu_ps void @struct_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { ; GFX8-LABEL: name: struct_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset ; GFX8: bb.1 (%ir-block.0): ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 @@ -78,9 +78,8 @@ define amdgpu_ps float @struct_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc_ ; GFX8-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 ; GFX8-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 ; GFX8-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 - ; GFX8-NEXT: [[BUFFER_ATOMIC_ADD_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) - ; GFX8-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_BOTHEN_RTN]] - ; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; GFX8-NEXT: BUFFER_ATOMIC_ADD_BOTHEN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: S_ENDPGM 0 ; ; GFX1200-LABEL: name: struct_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset ; GFX1200: bb.1 (%ir-block.0): @@ -96,9 +95,8 @@ define amdgpu_ps float @struct_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc_ ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 ; GFX1200-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 ; GFX1200-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 - ; GFX1200-NEXT: [[BUFFER_ATOMIC_ADD_VBUFFER_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_VBUFFER_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) - ; GFX1200-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_VBUFFER_BOTHEN_RTN]] - ; GFX1200-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; GFX1200-NEXT: BUFFER_ATOMIC_ADD_VBUFFER_BOTHEN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1200-NEXT: S_ENDPGM 0 ; ; GFX1250-LABEL: name: struct_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset ; GFX1250: bb.1 (%ir-block.0): @@ -114,12 +112,10 @@ define amdgpu_ps float @struct_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc_ ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 ; GFX1250-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 - ; GFX1250-NEXT: [[BUFFER_ATOMIC_ADD_VBUFFER_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_VBUFFER_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) - ; GFX1250-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_VBUFFER_BOTHEN_RTN]] - ; GFX1250-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; GFX1250-NEXT: BUFFER_ATOMIC_ADD_VBUFFER_BOTHEN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1250-NEXT: S_ENDPGM 0 %ret = call i32 @llvm.amdgcn.struct.buffer.atomic.add.i32(i32 %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) - %cast = bitcast i32 %ret to float - ret float %cast + ret void } define amdgpu_ps <2 x float> @struct_buffer_atomic_add_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i64 %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.buffer.atomic.sub.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.buffer.atomic.sub.ll new file mode 100644 index 000000000000..6719e466e18c --- /dev/null +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.buffer.atomic.sub.ll @@ -0,0 +1,663 @@ +; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -stop-after=instruction-select -o - %s | FileCheck -check-prefix=GFX8 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1200 -stop-after=instruction-select -o - %s | FileCheck -check-prefix=GFX1200 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1250 -stop-after=instruction-select -o - %s | FileCheck -check-prefix=GFX1250 %s + +; Natural mapping +define amdgpu_ps float @struct_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { + ; GFX8-LABEL: name: struct_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX8-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX8-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 + ; GFX8-NEXT: [[BUFFER_ATOMIC_SUB_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_BOTHEN_RTN]] + ; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; + ; GFX1200-LABEL: name: struct_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX1200: bb.1 (%ir-block.0): + ; GFX1200-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1200-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1200-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1200-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1200-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1200-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1200-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1200-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1200-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 + ; GFX1200-NEXT: [[BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1200-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN]] + ; GFX1200-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; + ; GFX1250-LABEL: name: struct_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX1250: bb.1 (%ir-block.0): + ; GFX1250-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1250-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1250-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1250-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1250-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1250-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1250-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1250-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 + ; GFX1250-NEXT: [[BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1250-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN]] + ; GFX1250-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %ret = call i32 @llvm.amdgcn.struct.buffer.atomic.sub.i32(i32 %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i32 %ret to float + ret float %cast +} + +define amdgpu_ps void @struct_buffer_atomic_sub_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { + ; GFX8-LABEL: name: struct_buffer_atomic_sub_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX8-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX8-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 + ; GFX8-NEXT: BUFFER_ATOMIC_SUB_BOTHEN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: S_ENDPGM 0 + ; + ; GFX1200-LABEL: name: struct_buffer_atomic_sub_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX1200: bb.1 (%ir-block.0): + ; GFX1200-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1200-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1200-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1200-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1200-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1200-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1200-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1200-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1200-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 + ; GFX1200-NEXT: BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1200-NEXT: S_ENDPGM 0 + ; + ; GFX1250-LABEL: name: struct_buffer_atomic_sub_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX1250: bb.1 (%ir-block.0): + ; GFX1250-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1250-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1250-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1250-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1250-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1250-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1250-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1250-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 + ; GFX1250-NEXT: BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1250-NEXT: S_ENDPGM 0 + %ret = call i32 @llvm.amdgcn.struct.buffer.atomic.sub.i32(i32 %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) + ret void +} + +define amdgpu_ps <2 x float> @struct_buffer_atomic_sub_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i64 %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { + ; GFX8-LABEL: name: struct_buffer_atomic_sub_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2, $vgpr3 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX8-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX8-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX8-NEXT: [[COPY8:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX8-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY7]], %subreg.sub1 + ; GFX8-NEXT: [[BUFFER_ATOMIC_SUB_X2_BOTHEN_RTN:%[0-9]+]]:vreg_64 = BUFFER_ATOMIC_SUB_X2_BOTHEN_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[COPY8]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 8) + ; GFX8-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_BOTHEN_RTN]].sub0 + ; GFX8-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_BOTHEN_RTN]].sub1 + ; GFX8-NEXT: $vgpr0 = COPY [[COPY9]] + ; GFX8-NEXT: $vgpr1 = COPY [[COPY10]] + ; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0, implicit $vgpr1 + ; + ; GFX1200-LABEL: name: struct_buffer_atomic_sub_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX1200: bb.1 (%ir-block.0): + ; GFX1200-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2, $vgpr3 + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1200-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1200-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; GFX1200-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1200-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1200-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1200-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1200-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1200-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX1200-NEXT: [[COPY8:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1200-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY7]], %subreg.sub1 + ; GFX1200-NEXT: [[BUFFER_ATOMIC_SUB_X2_VBUFFER_BOTHEN_RTN:%[0-9]+]]:vreg_64 = BUFFER_ATOMIC_SUB_X2_VBUFFER_BOTHEN_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[COPY8]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 8) + ; GFX1200-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_VBUFFER_BOTHEN_RTN]].sub0 + ; GFX1200-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_VBUFFER_BOTHEN_RTN]].sub1 + ; GFX1200-NEXT: $vgpr0 = COPY [[COPY9]] + ; GFX1200-NEXT: $vgpr1 = COPY [[COPY10]] + ; GFX1200-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0, implicit $vgpr1 + ; + ; GFX1250-LABEL: name: struct_buffer_atomic_sub_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX1250: bb.1 (%ir-block.0): + ; GFX1250-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2, $vgpr3 + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1250-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1250-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; GFX1250-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1250-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1250-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1250-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1250-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX1250-NEXT: [[COPY8:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1250-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY7]], %subreg.sub1 + ; GFX1250-NEXT: [[BUFFER_ATOMIC_SUB_X2_VBUFFER_BOTHEN_RTN:%[0-9]+]]:vreg_64_align2 = BUFFER_ATOMIC_SUB_X2_VBUFFER_BOTHEN_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[COPY8]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 8) + ; GFX1250-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_VBUFFER_BOTHEN_RTN]].sub0 + ; GFX1250-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_VBUFFER_BOTHEN_RTN]].sub1 + ; GFX1250-NEXT: $vgpr0 = COPY [[COPY9]] + ; GFX1250-NEXT: $vgpr1 = COPY [[COPY10]] + ; GFX1250-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0, implicit $vgpr1 + %ret = call i64 @llvm.amdgcn.struct.buffer.atomic.sub.i64(i64 %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i64 %ret to <2 x float> + ret <2 x float> %cast +} + +define amdgpu_ps void @struct_buffer_atomic_sub_i64_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i64 %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { + ; GFX8-LABEL: name: struct_buffer_atomic_sub_i64_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2, $vgpr3 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX8-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX8-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX8-NEXT: [[COPY8:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX8-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY7]], %subreg.sub1 + ; GFX8-NEXT: BUFFER_ATOMIC_SUB_X2_BOTHEN [[REG_SEQUENCE]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[COPY8]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 8) + ; GFX8-NEXT: S_ENDPGM 0 + ; + ; GFX1200-LABEL: name: struct_buffer_atomic_sub_i64_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX1200: bb.1 (%ir-block.0): + ; GFX1200-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2, $vgpr3 + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1200-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1200-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; GFX1200-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1200-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1200-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1200-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1200-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1200-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX1200-NEXT: [[COPY8:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1200-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY7]], %subreg.sub1 + ; GFX1200-NEXT: BUFFER_ATOMIC_SUB_X2_VBUFFER_BOTHEN [[REG_SEQUENCE]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[COPY8]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 8) + ; GFX1200-NEXT: S_ENDPGM 0 + ; + ; GFX1250-LABEL: name: struct_buffer_atomic_sub_i64_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; GFX1250: bb.1 (%ir-block.0): + ; GFX1250-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2, $vgpr3 + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1250-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1250-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; GFX1250-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1250-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1250-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1250-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1250-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX1250-NEXT: [[COPY8:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1250-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY7]], %subreg.sub1 + ; GFX1250-NEXT: BUFFER_ATOMIC_SUB_X2_VBUFFER_BOTHEN [[REG_SEQUENCE]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[COPY8]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 8) + ; GFX1250-NEXT: S_ENDPGM 0 + %ret = call i64 @llvm.amdgcn.struct.buffer.atomic.sub.i64(i64 %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) + ret void +} + +; All register operands need legalization +define amdgpu_ps float @struct_buffer_atomic_sub_i32__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset(i32 inreg %val, <4 x i32> %rsrc, i32 inreg %vindex, i32 inreg %voffset, i32 %soffset) { + ; GFX8-LABEL: name: struct_buffer_atomic_sub_i32__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: successors: %bb.2(0x80000000) + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; GFX8-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; GFX8-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; GFX8-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[COPY6]] + ; GFX8-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.2: + ; GFX8-NEXT: successors: %bb.3(0x80000000) + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; GFX8-NEXT: [[REG_SEQUENCE1:%[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 + ; GFX8-NEXT: [[COPY11:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; GFX8-NEXT: [[COPY12:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; GFX8-NEXT: [[COPY13:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; GFX8-NEXT: [[COPY14:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; GFX8-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY13]], [[COPY11]], implicit $exec + ; GFX8-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY14]], [[COPY12]], implicit $exec + ; GFX8-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]], implicit $exec + ; GFX8-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY7]], implicit $exec + ; GFX8-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; GFX8-NEXT: [[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 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.3: + ; GFX8-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY10]], %subreg.sub1 + ; GFX8-NEXT: [[BUFFER_ATOMIC_SUB_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_BOTHEN_RTN [[COPY8]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc + ; GFX8-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.4: + ; GFX8-NEXT: successors: %bb.5(0x80000000) + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: $exec = S_MOV_B64_term [[S_MOV_B64_]] + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.5: + ; GFX8-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_BOTHEN_RTN]] + ; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; + ; GFX1200-LABEL: name: struct_buffer_atomic_sub_i32__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; GFX1200: bb.1 (%ir-block.0): + ; GFX1200-NEXT: successors: %bb.2(0x80000000) + ; GFX1200-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1200-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1200-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1200-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1200-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX1200-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1200-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1200-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; GFX1200-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; GFX1200-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; GFX1200-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[COPY6]] + ; GFX1200-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_MOV_B32 $exec_lo + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.2: + ; GFX1200-NEXT: successors: %bb.3(0x80000000) + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; GFX1200-NEXT: [[REG_SEQUENCE1:%[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 + ; GFX1200-NEXT: [[COPY11:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; GFX1200-NEXT: [[COPY12:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; GFX1200-NEXT: [[COPY13:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; GFX1200-NEXT: [[COPY14:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; GFX1200-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY13]], [[COPY11]], implicit $exec + ; GFX1200-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY14]], [[COPY12]], implicit $exec + ; GFX1200-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]], implicit $exec + ; GFX1200-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY7]], implicit $exec + ; GFX1200-NEXT: [[S_AND_B32_1:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[S_AND_B32_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; GFX1200-NEXT: [[S_AND_SAVEEXEC_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_SAVEEXEC_B32 killed [[S_AND_B32_1]], implicit-def $exec, implicit-def $scc, implicit $exec + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.3: + ; GFX1200-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY10]], %subreg.sub1 + ; GFX1200-NEXT: [[BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN [[COPY8]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1200-NEXT: $exec_lo = S_XOR_B32_term $exec_lo, [[S_AND_SAVEEXEC_B32_]], implicit-def $scc + ; GFX1200-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.4: + ; GFX1200-NEXT: successors: %bb.5(0x80000000) + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: $exec_lo = S_MOV_B32_term [[S_MOV_B32_]] + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.5: + ; GFX1200-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN]] + ; GFX1200-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; + ; GFX1250-LABEL: name: struct_buffer_atomic_sub_i32__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; GFX1250: bb.1 (%ir-block.0): + ; GFX1250-NEXT: successors: %bb.2(0x80000000) + ; GFX1250-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1250-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1250-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1250-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1250-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX1250-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128_align2 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1250-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; GFX1250-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; GFX1250-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; GFX1250-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[COPY6]] + ; GFX1250-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_MOV_B32 $exec_lo + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.2: + ; GFX1250-NEXT: successors: %bb.3(0x80000000) + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; GFX1250-NEXT: [[REG_SEQUENCE1:%[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 + ; GFX1250-NEXT: [[COPY11:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; GFX1250-NEXT: [[COPY12:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; GFX1250-NEXT: [[COPY13:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; GFX1250-NEXT: [[COPY14:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; GFX1250-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY13]], [[COPY11]], implicit $exec + ; GFX1250-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY14]], [[COPY12]], implicit $exec + ; GFX1250-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]], implicit $exec + ; GFX1250-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY7]], implicit $exec + ; GFX1250-NEXT: [[S_AND_B32_1:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[S_AND_B32_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; GFX1250-NEXT: [[S_AND_SAVEEXEC_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_SAVEEXEC_B32 killed [[S_AND_B32_1]], implicit-def $exec, implicit-def $scc, implicit $exec + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.3: + ; GFX1250-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY10]], %subreg.sub1 + ; GFX1250-NEXT: [[BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN [[COPY8]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1250-NEXT: $exec_lo = S_XOR_B32_term $exec_lo, [[S_AND_SAVEEXEC_B32_]], implicit-def $scc + ; GFX1250-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.4: + ; GFX1250-NEXT: successors: %bb.5(0x80000000) + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: $exec_lo = S_MOV_B32_term [[S_MOV_B32_]] + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.5: + ; GFX1250-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN]] + ; GFX1250-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %ret = call i32 @llvm.amdgcn.struct.buffer.atomic.sub.i32(i32 %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i32 %ret to float + ret float %cast +} + +; All register operands need legalization +define amdgpu_ps void @struct_buffer_atomic_sub_i32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset(i32 inreg %val, <4 x i32> %rsrc, i32 inreg %vindex, i32 inreg %voffset, i32 %soffset) { + ; GFX8-LABEL: name: struct_buffer_atomic_sub_i32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: successors: %bb.2(0x80000000) + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; GFX8-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; GFX8-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; GFX8-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[COPY6]] + ; GFX8-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.2: + ; GFX8-NEXT: successors: %bb.3(0x80000000) + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; GFX8-NEXT: [[REG_SEQUENCE1:%[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 + ; GFX8-NEXT: [[COPY11:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; GFX8-NEXT: [[COPY12:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; GFX8-NEXT: [[COPY13:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; GFX8-NEXT: [[COPY14:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; GFX8-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY13]], [[COPY11]], implicit $exec + ; GFX8-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY14]], [[COPY12]], implicit $exec + ; GFX8-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; GFX8-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]], implicit $exec + ; GFX8-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY7]], implicit $exec + ; GFX8-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; GFX8-NEXT: [[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 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.3: + ; GFX8-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY10]], %subreg.sub1 + ; GFX8-NEXT: BUFFER_ATOMIC_SUB_BOTHEN [[COPY8]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc + ; GFX8-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.4: + ; GFX8-NEXT: successors: %bb.5(0x80000000) + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: $exec = S_MOV_B64_term [[S_MOV_B64_]] + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: bb.5: + ; GFX8-NEXT: S_ENDPGM 0 + ; + ; GFX1200-LABEL: name: struct_buffer_atomic_sub_i32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; GFX1200: bb.1 (%ir-block.0): + ; GFX1200-NEXT: successors: %bb.2(0x80000000) + ; GFX1200-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1200-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1200-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1200-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1200-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX1200-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1200-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1200-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; GFX1200-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; GFX1200-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; GFX1200-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[COPY6]] + ; GFX1200-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_MOV_B32 $exec_lo + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.2: + ; GFX1200-NEXT: successors: %bb.3(0x80000000) + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; GFX1200-NEXT: [[REG_SEQUENCE1:%[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 + ; GFX1200-NEXT: [[COPY11:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; GFX1200-NEXT: [[COPY12:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; GFX1200-NEXT: [[COPY13:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; GFX1200-NEXT: [[COPY14:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; GFX1200-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY13]], [[COPY11]], implicit $exec + ; GFX1200-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY14]], [[COPY12]], implicit $exec + ; GFX1200-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; GFX1200-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]], implicit $exec + ; GFX1200-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY7]], implicit $exec + ; GFX1200-NEXT: [[S_AND_B32_1:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[S_AND_B32_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; GFX1200-NEXT: [[S_AND_SAVEEXEC_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_SAVEEXEC_B32 killed [[S_AND_B32_1]], implicit-def $exec, implicit-def $scc, implicit $exec + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.3: + ; GFX1200-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY10]], %subreg.sub1 + ; GFX1200-NEXT: BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN [[COPY8]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1200-NEXT: $exec_lo = S_XOR_B32_term $exec_lo, [[S_AND_SAVEEXEC_B32_]], implicit-def $scc + ; GFX1200-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.4: + ; GFX1200-NEXT: successors: %bb.5(0x80000000) + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: $exec_lo = S_MOV_B32_term [[S_MOV_B32_]] + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: bb.5: + ; GFX1200-NEXT: S_ENDPGM 0 + ; + ; GFX1250-LABEL: name: struct_buffer_atomic_sub_i32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; GFX1250: bb.1 (%ir-block.0): + ; GFX1250-NEXT: successors: %bb.2(0x80000000) + ; GFX1250-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1250-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1250-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1250-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1250-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; GFX1250-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128_align2 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1250-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; GFX1250-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; GFX1250-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; GFX1250-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[COPY6]] + ; GFX1250-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_MOV_B32 $exec_lo + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.2: + ; GFX1250-NEXT: successors: %bb.3(0x80000000) + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; GFX1250-NEXT: [[REG_SEQUENCE1:%[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 + ; GFX1250-NEXT: [[COPY11:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; GFX1250-NEXT: [[COPY12:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; GFX1250-NEXT: [[COPY13:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; GFX1250-NEXT: [[COPY14:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; GFX1250-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY13]], [[COPY11]], implicit $exec + ; GFX1250-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U64_e64 [[COPY14]], [[COPY12]], implicit $exec + ; GFX1250-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; GFX1250-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]], implicit $exec + ; GFX1250-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY7]], implicit $exec + ; GFX1250-NEXT: [[S_AND_B32_1:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_B32 [[S_AND_B32_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; GFX1250-NEXT: [[S_AND_SAVEEXEC_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_AND_SAVEEXEC_B32 killed [[S_AND_B32_1]], implicit-def $exec, implicit-def $scc, implicit $exec + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.3: + ; GFX1250-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY10]], %subreg.sub1 + ; GFX1250-NEXT: BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN [[COPY8]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1250-NEXT: $exec_lo = S_XOR_B32_term $exec_lo, [[S_AND_SAVEEXEC_B32_]], implicit-def $scc + ; GFX1250-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.4: + ; GFX1250-NEXT: successors: %bb.5(0x80000000) + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: $exec_lo = S_MOV_B32_term [[S_MOV_B32_]] + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: bb.5: + ; GFX1250-NEXT: S_ENDPGM 0 + %ret = call i32 @llvm.amdgcn.struct.buffer.atomic.sub.i32(i32 %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) + ret void +} + +; Natural mapping + slc +define amdgpu_ps float @struct_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc(i32 %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { + ; GFX8-LABEL: name: struct_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc + ; GFX8: bb.1 (%ir-block.0): + ; GFX8-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX8-NEXT: {{ $}} + ; GFX8-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX8-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX8-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX8-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX8-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX8-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX8-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX8-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX8-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX8-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 + ; GFX8-NEXT: [[BUFFER_ATOMIC_SUB_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 3, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX8-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_BOTHEN_RTN]] + ; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; + ; GFX1200-LABEL: name: struct_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc + ; GFX1200: bb.1 (%ir-block.0): + ; GFX1200-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX1200-NEXT: {{ $}} + ; GFX1200-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1200-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1200-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1200-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1200-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1200-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1200-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1200-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1200-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1200-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 + ; GFX1200-NEXT: [[BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 3, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1200-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN]] + ; GFX1200-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; + ; GFX1250-LABEL: name: struct_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc + ; GFX1250: bb.1 (%ir-block.0): + ; GFX1250-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; GFX1250-NEXT: {{ $}} + ; GFX1250-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GFX1250-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; GFX1250-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; GFX1250-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; GFX1250-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; GFX1250-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; GFX1250-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; GFX1250-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; GFX1250-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; GFX1250-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 + ; GFX1250-NEXT: [[BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 3, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8) + ; GFX1250-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_VBUFFER_BOTHEN_RTN]] + ; GFX1250-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %ret = call i32 @llvm.amdgcn.struct.buffer.atomic.sub.i32(i32 %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 2) + %cast = bitcast i32 %ret to float + ret float %cast +} + +declare i32 @llvm.amdgcn.struct.buffer.atomic.sub.i32(i32, <4 x i32>, i32, i32, i32, i32 immarg) #0 +declare i64 @llvm.amdgcn.struct.buffer.atomic.sub.i64(i64, <4 x i32>, i32, i32, i32, i32 immarg) #0 + +attributes #0 = { nounwind } diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.ptr.buffer.atomic.add.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.ptr.buffer.atomic.add.ll index 968e2ba5887c..aa2c78aaa4b9 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.ptr.buffer.atomic.add.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.ptr.buffer.atomic.add.ll @@ -1,5 +1,5 @@ ; 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 -o - %s | FileCheck %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -stop-after=instruction-select -o - %s | FileCheck %s ; Natural mapping define amdgpu_ps float @struct_ptr_buffer_atomic_add_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { @@ -25,7 +25,7 @@ define amdgpu_ps float @struct_ptr_buffer_atomic_add_i32__vgpr_val__sgpr_rsrc__v ret float %cast } -define amdgpu_ps float @struct_ptr_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { +define amdgpu_ps void @struct_ptr_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { ; CHECK-LABEL: name: struct_ptr_buffer_atomic_add_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset ; CHECK: bb.1 (%ir-block.0): ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 @@ -40,12 +40,10 @@ define amdgpu_ps float @struct_ptr_buffer_atomic_add_i32_noret__vgpr_val__sgpr_r ; CHECK-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 ; CHECK-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 - ; CHECK-NEXT: [[BUFFER_ATOMIC_ADD_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) - ; CHECK-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_BOTHEN_RTN]] - ; CHECK-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + ; CHECK-NEXT: BUFFER_ATOMIC_ADD_BOTHEN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: S_ENDPGM 0 %ret = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.add.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) - %cast = bitcast i32 %ret to float - ret float %cast + ret void } define amdgpu_ps <2 x float> @struct_ptr_buffer_atomic_add_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i64 %val, ptr addrspace(8) inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.ptr.buffer.atomic.sub.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.ptr.buffer.atomic.sub.ll new file mode 100644 index 000000000000..db8e444d909a --- /dev/null +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.struct.ptr.buffer.atomic.sub.ll @@ -0,0 +1,249 @@ +; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -stop-after=instruction-select -o - %s | FileCheck %s + +; Natural mapping +define amdgpu_ps float @struct_ptr_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { + ; CHECK-LABEL: name: struct_ptr_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; CHECK-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; CHECK-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 + ; CHECK-NEXT: [[BUFFER_ATOMIC_SUB_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_BOTHEN_RTN]] + ; CHECK-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %ret = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.sub.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i32 %ret to float + ret float %cast +} + +define amdgpu_ps void @struct_ptr_buffer_atomic_sub_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { + ; CHECK-LABEL: name: struct_ptr_buffer_atomic_sub_i32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; CHECK-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; CHECK-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 + ; CHECK-NEXT: BUFFER_ATOMIC_SUB_BOTHEN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: S_ENDPGM 0 + %ret = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.sub.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) + ret void +} + +define amdgpu_ps <2 x float> @struct_ptr_buffer_atomic_sub_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i64 %val, ptr addrspace(8) inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { + ; CHECK-LABEL: name: struct_ptr_buffer_atomic_sub_i64__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2, $vgpr3 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; CHECK-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; CHECK-NEXT: [[COPY8:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; CHECK-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; CHECK-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY7]], %subreg.sub1 + ; CHECK-NEXT: [[BUFFER_ATOMIC_SUB_X2_BOTHEN_RTN:%[0-9]+]]:vreg_64 = BUFFER_ATOMIC_SUB_X2_BOTHEN_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[COPY8]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_BOTHEN_RTN]].sub0 + ; CHECK-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_SUB_X2_BOTHEN_RTN]].sub1 + ; CHECK-NEXT: $vgpr0 = COPY [[COPY9]] + ; CHECK-NEXT: $vgpr1 = COPY [[COPY10]] + ; CHECK-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0, implicit $vgpr1 + %ret = call i64 @llvm.amdgcn.struct.ptr.buffer.atomic.sub.i64(i64 %val, ptr addrspace(8) %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i64 %ret to <2 x float> + ret <2 x float> %cast +} + +define amdgpu_ps void @struct_ptr_buffer_atomic_sub_i64_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(i64 %val, ptr addrspace(8) inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { + ; CHECK-LABEL: name: struct_ptr_buffer_atomic_sub_i64_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2, $vgpr3 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; CHECK-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; CHECK-NEXT: [[COPY8:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; CHECK-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3 + ; CHECK-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY7]], %subreg.sub1 + ; CHECK-NEXT: BUFFER_ATOMIC_SUB_X2_BOTHEN [[REG_SEQUENCE]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[COPY8]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: S_ENDPGM 0 + %ret = call i64 @llvm.amdgcn.struct.ptr.buffer.atomic.sub.i64(i64 %val, ptr addrspace(8) %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) + ret void +} + +; All register operands need legalization +define amdgpu_ps float @struct_ptr_buffer_atomic_sub_i32__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset(i32 inreg %val, ptr addrspace(8) %rsrc, i32 inreg %vindex, i32 inreg %voffset, i32 %soffset) { + ; CHECK-LABEL: name: struct_ptr_buffer_atomic_sub_i32__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: successors: %bb.2(0x80000000) + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; CHECK-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; CHECK-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; CHECK-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[COPY6]] + ; CHECK-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.2: + ; CHECK-NEXT: successors: %bb.3(0x80000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; CHECK-NEXT: [[REG_SEQUENCE1:%[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-NEXT: [[COPY11:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; CHECK-NEXT: [[COPY12:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; CHECK-NEXT: [[COPY13:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; CHECK-NEXT: [[COPY14:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; CHECK-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY13]], [[COPY11]], implicit $exec + ; CHECK-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY14]], [[COPY12]], implicit $exec + ; CHECK-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]], implicit $exec + ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY7]], implicit $exec + ; CHECK-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; CHECK-NEXT: [[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-NEXT: {{ $}} + ; CHECK-NEXT: bb.3: + ; CHECK-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY10]], %subreg.sub1 + ; CHECK-NEXT: [[BUFFER_ATOMIC_SUB_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_BOTHEN_RTN [[COPY8]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc + ; CHECK-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.4: + ; CHECK-NEXT: successors: %bb.5(0x80000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: $exec = S_MOV_B64_term [[S_MOV_B64_]] + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.5: + ; CHECK-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_BOTHEN_RTN]] + ; CHECK-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %ret = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.sub.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) + %cast = bitcast i32 %ret to float + ret float %cast +} + +; All register operands need legalization +define amdgpu_ps void @struct_ptr_buffer_atomic_sub_i32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset(i32 inreg %val, ptr addrspace(8) %rsrc, i32 inreg %vindex, i32 inreg %voffset, i32 %soffset) { + ; CHECK-LABEL: name: struct_ptr_buffer_atomic_sub_i32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: successors: %bb.2(0x80000000) + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr4 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; CHECK-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY]] + ; CHECK-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[COPY5]] + ; CHECK-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[COPY6]] + ; CHECK-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.2: + ; CHECK-NEXT: successors: %bb.3(0x80000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec + ; CHECK-NEXT: [[REG_SEQUENCE1:%[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-NEXT: [[COPY11:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1 + ; CHECK-NEXT: [[COPY12:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3 + ; CHECK-NEXT: [[COPY13:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1 + ; CHECK-NEXT: [[COPY14:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3 + ; CHECK-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY13]], [[COPY11]], implicit $exec + ; CHECK-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY14]], [[COPY12]], implicit $exec + ; CHECK-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc + ; CHECK-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY7]], implicit $exec + ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY7]], implicit $exec + ; CHECK-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc + ; CHECK-NEXT: [[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-NEXT: {{ $}} + ; CHECK-NEXT: bb.3: + ; CHECK-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY10]], %subreg.sub1 + ; CHECK-NEXT: BUFFER_ATOMIC_SUB_BOTHEN [[COPY8]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc + ; CHECK-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.4: + ; CHECK-NEXT: successors: %bb.5(0x80000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: $exec = S_MOV_B64_term [[S_MOV_B64_]] + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.5: + ; CHECK-NEXT: S_ENDPGM 0 + %ret = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.sub.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0) + ret void +} + +; Natural mapping + slc +define amdgpu_ps float @struct_ptr_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc(i32 %val, ptr addrspace(8) inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) { + ; CHECK-LABEL: name: struct_ptr_buffer_atomic_sub_i32__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc + ; CHECK: bb.1 (%ir-block.0): + ; CHECK-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3 + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4 + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5 + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1 + ; CHECK-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2 + ; CHECK-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6 + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3 + ; CHECK-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1 + ; CHECK-NEXT: [[BUFFER_ATOMIC_SUB_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_SUB_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 3, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8) + ; CHECK-NEXT: $vgpr0 = COPY [[BUFFER_ATOMIC_SUB_BOTHEN_RTN]] + ; CHECK-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0 + %ret = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.sub.i32(i32 %val, ptr addrspace(8) %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 2) + %cast = bitcast i32 %ret to float + ret float %cast +} + +declare i32 @llvm.amdgcn.struct.ptr.buffer.atomic.sub.i32(i32, ptr addrspace(8), i32, i32, i32, i32 immarg) #0 +declare i64 @llvm.amdgcn.struct.ptr.buffer.atomic.sub.i64(i64, ptr addrspace(8), i32, i32, i32, i32 immarg) #0 + +attributes #0 = { nounwind }