AMDGPU: Teach isOperandLegal about SALU literal restrictions #127626
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
isOperandLegal mostly implemented the VALU operand rules, and largely ignored SALU restrictions. This theoretically avoids folding literals into SALU insts which already have a literal operand. This issue is currently avoided due to a bug in SIFoldOperands; this change will allow using raw operand legality rules. This breaks the formation of s_fmaak_f32 in SIFoldOperands, but it probably should not have been forming there in the first place. TwoAddressInsts or RA should generally handle that, and this only worked by accident.
|
@llvm/pr-subscribers-backend-amdgpu Author: Matt Arsenault (arsenm) ChangesisOperandLegal mostly implemented the VALU operand rules, and This breaks the formation of s_fmaak_f32 in SIFoldOperands, Full diff: https://github.com/llvm/llvm-project/pull/127626.diff 3 Files Affected:
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index ceab6c9dcca34..7dace11d208a0 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -5931,11 +5931,15 @@ bool SIInstrInfo::isOperandLegal(const MachineInstr &MI, unsigned OpIdx,
if (!MO)
MO = &MI.getOperand(OpIdx);
- const MachineOperand *UsedLiteral = nullptr;
+ const bool IsInlineConst = !MO->isReg() && isInlineConstant(*MO, OpInfo);
- int ConstantBusLimit = ST.getConstantBusLimit(MI.getOpcode());
- int LiteralLimit = !isVOP3(MI) || ST.hasVOP3Literal() ? 1 : 0;
- if (isVALU(MI) && usesConstantBus(MRI, *MO, OpInfo)) {
+ if (isVALU(MI) && !IsInlineConst && usesConstantBus(MRI, *MO, OpInfo)) {
+ const MachineOperand *UsedLiteral = nullptr;
+
+ int ConstantBusLimit = ST.getConstantBusLimit(MI.getOpcode());
+ int LiteralLimit = !isVOP3(MI) || ST.hasVOP3Literal() ? 1 : 0;
+
+ // TODO: Be more permissive with frame indexes.
if (!MO->isReg() && !isInlineConstant(*MO, OpInfo)) {
if (!LiteralLimit--)
return false;
@@ -5974,9 +5978,19 @@ bool SIInstrInfo::isOperandLegal(const MachineInstr &MI, unsigned OpIdx,
return false;
}
}
- } else if (ST.hasNoF16PseudoScalarTransInlineConstants() && !MO->isReg() &&
- isF16PseudoScalarTrans(MI.getOpcode()) &&
- isInlineConstant(*MO, OpInfo)) {
+ } else if (!IsInlineConst && !MO->isReg() && isSALU(MI)) {
+ // There can be at most one literal operand, but it can be repeated.
+ for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
+ if (i == OpIdx)
+ continue;
+ const MachineOperand &Op = MI.getOperand(i);
+ if (!Op.isReg() && !Op.isFI() &&
+ !isInlineConstant(Op, InstDesc.operands()[i]) &&
+ !Op.isIdenticalTo(*MO))
+ return false;
+ }
+ } else if (IsInlineConst && ST.hasNoF16PseudoScalarTransInlineConstants() &&
+ isF16PseudoScalarTrans(MI.getOpcode())) {
return false;
}
diff --git a/llvm/test/CodeGen/AMDGPU/fold-operands-scalar-fmac.mir b/llvm/test/CodeGen/AMDGPU/fold-operands-scalar-fmac.mir
index 08693ec9db1d4..2492eb2982aac 100644
--- a/llvm/test/CodeGen/AMDGPU/fold-operands-scalar-fmac.mir
+++ b/llvm/test/CodeGen/AMDGPU/fold-operands-scalar-fmac.mir
@@ -133,7 +133,8 @@ body: |
; CHECK: liveins: $sgpr0
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
- ; CHECK-NEXT: %fma:sreg_32 = nofpexcept S_FMAMK_F32 [[COPY]], 1234567890, 1056964608, implicit $mode
+ ; CHECK-NEXT: %noninlinable:sreg_32 = S_MOV_B32 1234567890
+ ; CHECK-NEXT: %fma:sreg_32 = nofpexcept S_FMAAK_F32 %noninlinable, [[COPY]], 1056964608, implicit $mode
; CHECK-NEXT: $sgpr0 = COPY %fma
%0:sreg_32 = COPY $sgpr0
%inlinable:sreg_32 = S_MOV_B32 1056964608
@@ -152,7 +153,8 @@ body: |
; CHECK: liveins: $sgpr0
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
- ; CHECK-NEXT: %fma:sreg_32 = nofpexcept S_FMAMK_F32 [[COPY]], 1234567890, 1056964608, implicit $mode
+ ; CHECK-NEXT: %noninlinable:sreg_32 = S_MOV_B32 1234567890
+ ; CHECK-NEXT: %fma:sreg_32 = nofpexcept S_FMAAK_F32 [[COPY]], %noninlinable, 1056964608, implicit $mode
; CHECK-NEXT: $sgpr0 = COPY %fma
%0:sreg_32 = COPY $sgpr0
%inlinable:sreg_32 = S_MOV_B32 1056964608
@@ -210,7 +212,8 @@ body: |
; CHECK: liveins: $sgpr0
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
- ; CHECK-NEXT: %fma:sreg_32 = nofpexcept S_FMAAK_F32 [[COPY]], 1056964608, 1234567890, implicit $mode
+ ; CHECK-NEXT: %noninlinable:sreg_32 = S_MOV_B32 1234567890
+ ; CHECK-NEXT: %fma:sreg_32 = nofpexcept S_FMAMK_F32 [[COPY]], 1056964608, %noninlinable, implicit $mode
; CHECK-NEXT: $sgpr0 = COPY %fma
%0:sreg_32 = COPY $sgpr0
%inlinable:sreg_32 = S_MOV_B32 1056964608
diff --git a/llvm/test/CodeGen/AMDGPU/fold-sgpr-multi-imm.mir b/llvm/test/CodeGen/AMDGPU/fold-sgpr-multi-imm.mir
index 5f985605c082d..c8afb89aa272a 100644
--- a/llvm/test/CodeGen/AMDGPU/fold-sgpr-multi-imm.mir
+++ b/llvm/test/CodeGen/AMDGPU/fold-sgpr-multi-imm.mir
@@ -69,3 +69,202 @@ body: |
%0:sreg_32 = S_MOV_B32 63
%1:sreg_32 = S_ADD_I32 %stack.0, %0, implicit-def $scc
...
+
+# GCN-LABEL: name: test_no_fold_literal_already_inline_lhs{{$}}
+# GCN: %0:sreg_32 = S_MOV_B32 80
+# GCN-NEXT: %1:sreg_32 = S_ADD_I32 70, %0
+---
+name: test_no_fold_literal_already_inline_lhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 80
+ %1:sreg_32 = S_ADD_I32 70, %0, implicit-def $scc
+...
+
+# GCN-LABEL: name: test_no_fold_literal_already_inline_rhs{{$}}
+# GCN: %0:sreg_32 = S_MOV_B32 80
+# GCN-NEXT: %1:sreg_32 = S_ADD_I32 %0, 70
+---
+name: test_no_fold_literal_already_inline_rhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 80
+ %1:sreg_32 = S_ADD_I32 %0, 70, implicit-def $scc
+...
+
+# GCN-LABEL: name: test_fold_literal_inlineimm_lhs{{$}}
+# GCN: %1:sreg_32 = S_ADD_I32 64, 80
+---
+name: test_fold_literal_inlineimm_lhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 80
+ %1:sreg_32 = S_ADD_I32 64, %0, implicit-def $scc
+...
+
+# GCN-LABEL: name: test_fold_literal_inlineimm_rhs{{$}}
+# GCN: %1:sreg_32 = S_ADD_I32 80, 64
+---
+name: test_fold_literal_inlineimm_rhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 80
+ %1:sreg_32 = S_ADD_I32 %0, 64, implicit-def $scc
+...
+
+# GCN-LABEL: name: test_fold_same_literal_2x{{$}}
+# GCN: %2:sreg_32 = S_ADD_I32 70, %1
+---
+name: test_fold_same_literal_2x
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 70
+ %1:sreg_32 = S_MOV_B32 70
+ %2:sreg_32 = S_ADD_I32 %0, %1, implicit-def $scc
+...
+
+# GCN-LABEL: name: test_fold_same_literal_lhs{{$}}
+# GCN: %1:sreg_32 = S_ADD_I32 70, %0
+---
+name: test_fold_same_literal_lhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 70
+ %1:sreg_32 = S_ADD_I32 70, %0, implicit-def $scc
+...
+
+# GCN-LABEL: name: test_fold_same_literal_rhs{{$}}
+# GCN: %1:sreg_32 = S_ADD_I32 %0, 70
+---
+name: test_fold_same_literal_rhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 70
+ %1:sreg_32 = S_ADD_I32 %0, 70, implicit-def $scc
+...
+
+
+# GCN-LABEL: name: test_s_cselect_b32_2x_literal_fold{{$}}
+# GCN: %2:sreg_32 = S_CSELECT_B32 70, %1, implicit $scc
+---
+name: test_s_cselect_b32_2x_literal_fold
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 70
+ %1:sreg_32 = S_MOV_B32 80
+ $scc = IMPLICIT_DEF
+ %2:sreg_32 = S_CSELECT_B32 %0, %1, implicit $scc
+...
+
+# GCN-LABEL: name: test_s_cselect_b32_fold_literal_literal_lhs{{$}}
+# GCN: %1:sreg_32 = S_CSELECT_B32 70, %0, implicit $scc
+---
+name: test_s_cselect_b32_fold_literal_literal_lhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 80
+ $scc = IMPLICIT_DEF
+ %1:sreg_32 = S_CSELECT_B32 70, %0, implicit $scc
+...
+
+# GCN-LABEL: name: test_s_cselect_b32_fold_literal_literal_rhs{{$}}
+# GCN: %1:sreg_32 = S_CSELECT_B32 %0, 70, implicit $scc
+---
+name: test_s_cselect_b32_fold_literal_literal_rhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 80
+ $scc = IMPLICIT_DEF
+ %1:sreg_32 = S_CSELECT_B32 %0, 70, implicit $scc
+...
+
+# GCN-LABEL: name: test_s_cselect_b32_fold_literal_inlineimm_lhs{{$}}
+# GCN: %1:sreg_32 = S_CSELECT_B32 64, 80, implicit $scc
+---
+name: test_s_cselect_b32_fold_literal_inlineimm_lhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 80
+ $scc = IMPLICIT_DEF
+ %1:sreg_32 = S_CSELECT_B32 64, %0, implicit $scc
+...
+
+# GCN-LABEL: name: test_s_cselect_b32_fold_literal_inlineimm_rhs{{$}}
+# GCN: %1:sreg_32 = S_CSELECT_B32 80, 64, implicit $scc
+---
+name: test_s_cselect_b32_fold_literal_inlineimm_rhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 80
+ $scc = IMPLICIT_DEF
+ %1:sreg_32 = S_CSELECT_B32 %0, 64, implicit $scc
+...
+
+# GCN-LABEL: name: test_s_cmp_b32_2x_literal_fold{{$}}
+# GCN: S_CMP_EQ_U32 70, %1, implicit-def $scc
+---
+name: test_s_cmp_b32_2x_literal_fold
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 70
+ %1:sreg_32 = S_MOV_B32 80
+ $scc = IMPLICIT_DEF
+ S_CMP_EQ_U32 %0, %1, implicit-def $scc
+...
+
+# GCN-LABEL: name: test_s_cmp_b32_literal_literal_lhs{{$}}
+# GCN: S_CMP_EQ_U32 70, %0, implicit-def $scc
+---
+name: test_s_cmp_b32_literal_literal_lhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 80
+ S_CMP_EQ_U32 70, %0, implicit-def $scc
+...
+
+# GCN-LABEL: name: test_s_cmp_b32_literal_literal_rhs{{$}}
+# GCN: S_CMP_EQ_U32 %0, 70, implicit-def $scc
+---
+name: test_s_cmp_b32_literal_literal_rhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 80
+ S_CMP_EQ_U32 %0, 70, implicit-def $scc
+...
+
+# GCN-LABEL: name: test_s_cmp_b32_literal_inlineimm_lhs{{$}}
+# GCN: S_CMP_EQ_U32 64, 80, implicit-def $scc
+---
+name: test_s_cmp_b32_literal_inlineimm_lhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 80
+ S_CMP_EQ_U32 64, %0, implicit-def $scc
+...
+
+# GCN-LABEL: name: test_s_cmp_b32_literal_inlineimm_rhs{{$}}
+# GCN: S_CMP_EQ_U32 80, 64, implicit-def $scc
+---
+name: test_s_cmp_b32_literal_inlineimm_rhs
+tracksRegLiveness: true
+body: |
+ bb.0:
+ %0:sreg_32 = S_MOV_B32 80
+ S_CMP_EQ_U32 %0, 64, implicit-def $scc
+...
|
arsenm
requested review from
jayfoad,
jrbyrnes,
kerbowa,
mbrkusanin,
perlfu,
Pierre-vh,
rampitec and
shiltian
rampitec
approved these changes
Feb 18, 2025
arsenm
deleted the
users/arsenm/amdgpu/avoid-double-literal-salu-isOperandLegal
branch
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
isOperandLegal mostly implemented the VALU operand rules, and
largely ignored SALU restrictions. This theoretically avoids
folding literals into SALU insts which already have a literal
operand. This issue is currently avoided due to a bug in
SIFoldOperands; this change will allow using raw operand
legality rules.
This breaks the formation of s_fmaak_f32 in SIFoldOperands,
but it probably should not have been forming there in the first
place. TwoAddressInsts or RA should generally handle that,
and this only worked by accident.