[llvm] [VPlan] Extend tryToFoldLiveIns to fold binary intrinsics (PR #161703)
Ramkumar Ramachandra via llvm-commits
llvm-commits at lists.llvm.org
Thu Oct 23 13:00:37 PDT 2025
https://github.com/artagnon updated https://github.com/llvm/llvm-project/pull/161703
>From eb9fccd5183a0f2ed25569e2b330467394a6fdd9 Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Thu, 2 Oct 2025 18:05:11 +0100
Subject: [PATCH 1/5] [LV] Pre-commit test for constantfolder-binaryintrinsic
---
.../LoopVectorize/constantfolder.ll | 77 +++++++++++++++++++
1 file changed, 77 insertions(+)
diff --git a/llvm/test/Transforms/LoopVectorize/constantfolder.ll b/llvm/test/Transforms/LoopVectorize/constantfolder.ll
index 66592b0ccf677..4bc26a7d19a9b 100644
--- a/llvm/test/Transforms/LoopVectorize/constantfolder.ll
+++ b/llvm/test/Transforms/LoopVectorize/constantfolder.ll
@@ -288,3 +288,80 @@ loop.latch:
exit:
ret void
}
+
+define void @const_fold_binaryintrinsic(ptr %dst, i64 %d) {
+; CHECK-LABEL: define void @const_fold_binaryintrinsic(
+; CHECK-SAME: ptr [[DST:%.*]], i64 [[D:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
+; CHECK: [[VECTOR_PH]]:
+; CHECK-NEXT: [[TMP0:%.*]] = call <4 x i64> @llvm.umax.v4i64(<4 x i64> zeroinitializer, <4 x i64> splat (i64 3))
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <4 x i64> [[TMP0]], i32 3
+; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
+; CHECK: [[VECTOR_BODY]]:
+; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: store i64 [[TMP1]], ptr [[DST]], align 2
+; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i64 [[INDEX_NEXT]], 100
+; CHECK-NEXT: br i1 [[TMP2]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]]
+; CHECK: [[MIDDLE_BLOCK]]:
+; CHECK-NEXT: br label %[[EXIT:.*]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %const.0 = xor i64 %d, %d
+ %trunc = call i64 @llvm.umax.i64(i64 %const.0, i64 3)
+ store i64 %trunc, ptr %dst, align 2
+ %iv.next = add i64 %iv, 1
+ %cmp = icmp ult i64 %iv.next, 100
+ br i1 %cmp, label %loop, label %exit
+
+exit:
+ ret void
+}
+
+define void @const_fold_widegep(ptr noalias %A, ptr noalias %B) {
+; CHECK-LABEL: define void @const_fold_widegep(
+; CHECK-SAME: ptr noalias [[A:%.*]], ptr noalias [[B:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
+; CHECK: [[VECTOR_PH]]:
+; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
+; CHECK: [[VECTOR_BODY]]:
+; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: store ptr [[A]], ptr [[B]], align 8
+; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT: [[TMP0:%.*]] = icmp eq i64 [[INDEX_NEXT]], 100
+; CHECK-NEXT: br i1 [[TMP0]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]
+; CHECK: [[MIDDLE_BLOCK]]:
+; CHECK-NEXT: br label %[[EXIT:.*]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ br label %loop.header
+
+loop.header:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop.latch ]
+ br i1 true, label %loop.latch, label %else
+
+else:
+ br label %loop.latch
+
+loop.latch:
+ %const.0 = phi i64 [ 0, %loop.header ], [ %iv, %else ]
+ %gep.A = getelementptr i64, ptr %A, i64 %const.0
+ %gep.B = getelementptr i64, ptr %B, i64 %const.0
+ store ptr %gep.A, ptr %gep.B
+ %iv.next = add nuw nsw i64 %iv, 1
+ %exit.cond = icmp ult i64 %iv.next, 100
+ br i1 %exit.cond, label %loop.header, label %exit
+
+exit:
+ ret void
+}
>From cc0f14127d13f41d1c19a064586a210b68533d72 Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Thu, 2 Oct 2025 18:07:40 +0100
Subject: [PATCH 2/5] [VPlan] Extend tryToFoldLiveIns to fold binary intrinsics
InstSimplifyFolder can fold binary intrinsics, so take the opportunity
to unify code with getOpcodeOrIntrinsicID, and handle the case. The
additional handling of WidenGEP is non-functional, as the GEP is
simplified before it is widened, as the included test shows.
---
.../llvm/Analysis/InstSimplifyFolder.h | 2 +-
.../Transforms/Vectorize/VPlanTransforms.cpp | 83 ++++++++++---------
.../LoopVectorize/constantfolder.ll | 4 +-
3 files changed, 46 insertions(+), 43 deletions(-)
diff --git a/llvm/include/llvm/Analysis/InstSimplifyFolder.h b/llvm/include/llvm/Analysis/InstSimplifyFolder.h
index 58793ed977f68..2832beb9e337c 100644
--- a/llvm/include/llvm/Analysis/InstSimplifyFolder.h
+++ b/llvm/include/llvm/Analysis/InstSimplifyFolder.h
@@ -120,7 +120,7 @@ class LLVM_ABI InstSimplifyFolder final : public IRBuilderFolder {
}
Value *FoldBinaryIntrinsic(Intrinsic::ID ID, Value *LHS, Value *RHS, Type *Ty,
- Instruction *FMFSource) const override {
+ Instruction *FMFSource = nullptr) const override {
return simplifyBinaryIntrinsic(ID, Ty, LHS, RHS, SQ,
dyn_cast_if_present<CallBase>(FMFSource));
}
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 3e85e6ff514ef..f6df5668bf451 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -943,12 +943,41 @@ static void recursivelyDeleteDeadRecipes(VPValue *V) {
}
}
+namespace {
+/// Get any instruction opcode or intrinsic ID data embedded in recipe \p R.
+/// Returns an optional pair, where the first element indicates whether it is
+/// an intrinsic ID.
+static std::optional<std::pair<bool, unsigned>>
+getOpcodeOrIntrinsicID(const VPSingleDefRecipe *R) {
+ return TypeSwitch<const VPSingleDefRecipe *,
+ std::optional<std::pair<bool, unsigned>>>(R)
+ .Case<VPInstruction, VPWidenRecipe, VPWidenCastRecipe,
+ VPWidenSelectRecipe, VPWidenGEPRecipe, VPReplicateRecipe>(
+ [](auto *I) { return std::make_pair(false, I->getOpcode()); })
+ .Case<VPWidenIntrinsicRecipe>([](auto *I) {
+ return std::make_pair(true, I->getVectorIntrinsicID());
+ })
+ .Case<VPVectorPointerRecipe, VPPredInstPHIRecipe>([](auto *I) {
+ // For recipes that do not directly map to LLVM IR instructions,
+ // assign opcodes after the last VPInstruction opcode (which is also
+ // after the last IR Instruction opcode), based on the VPDefID.
+ return std::make_pair(false,
+ VPInstruction::OpsEnd + 1 + I->getVPDefID());
+ })
+ .Default([](auto *) { return std::nullopt; });
+}
+} // namespace
+
/// Try to fold \p R using InstSimplifyFolder. Will succeed and return a
-/// non-nullptr Value for a handled \p Opcode if corresponding \p Operands are
-/// foldable live-ins.
-static Value *tryToFoldLiveIns(const VPRecipeBase &R, unsigned Opcode,
+/// non-nullptr Value for a handled opcode or intrinsic ID if corresponding \p
+/// Operands are foldable live-ins.
+static Value *tryToFoldLiveIns(VPSingleDefRecipe &R,
ArrayRef<VPValue *> Operands,
const DataLayout &DL, VPTypeAnalysis &TypeInfo) {
+ auto OpcodeOrIID = getOpcodeOrIntrinsicID(&R);
+ if (!OpcodeOrIID)
+ return nullptr;
+
SmallVector<Value *, 4> Ops;
for (VPValue *Op : Operands) {
if (!Op->isLiveIn() || !Op->getLiveInIRValue())
@@ -957,6 +986,14 @@ static Value *tryToFoldLiveIns(const VPRecipeBase &R, unsigned Opcode,
}
InstSimplifyFolder Folder(DL);
+ if (OpcodeOrIID->first) {
+ if (R.getNumOperands() != 2)
+ return nullptr;
+ unsigned ID = OpcodeOrIID->second;
+ return Folder.FoldBinaryIntrinsic(ID, Ops[0], Ops[1],
+ TypeInfo.inferScalarType(&R));
+ }
+ unsigned Opcode = OpcodeOrIID->second;
if (Instruction::isBinaryOp(Opcode))
return Folder.FoldBinOp(static_cast<Instruction::BinaryOps>(Opcode), Ops[0],
Ops[1]);
@@ -1006,19 +1043,10 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
// Simplification of live-in IR values for SingleDef recipes using
// InstSimplifyFolder.
- if (TypeSwitch<VPRecipeBase *, bool>(&R)
- .Case<VPInstruction, VPWidenRecipe, VPWidenCastRecipe,
- VPReplicateRecipe, VPWidenSelectRecipe>([&](auto *I) {
- const DataLayout &DL =
- Plan->getScalarHeader()->getIRBasicBlock()->getDataLayout();
- Value *V = tryToFoldLiveIns(*I, I->getOpcode(), I->operands(), DL,
- TypeInfo);
- if (V)
- I->replaceAllUsesWith(Plan->getOrAddLiveIn(V));
- return V;
- })
- .Default([](auto *) { return false; }))
- return;
+ const DataLayout &DL =
+ Plan->getScalarHeader()->getIRBasicBlock()->getDataLayout();
+ if (Value *V = tryToFoldLiveIns(*Def, Def->operands(), DL, TypeInfo))
+ return Def->replaceAllUsesWith(Plan->getOrAddLiveIn(V));
// Fold PredPHI LiveIn -> LiveIn.
if (auto *PredPHI = dyn_cast<VPPredInstPHIRecipe>(&R)) {
@@ -1996,29 +2024,6 @@ struct VPCSEDenseMapInfo : public DenseMapInfo<VPSingleDefRecipe *> {
return Def == getEmptyKey() || Def == getTombstoneKey();
}
- /// Get any instruction opcode or intrinsic ID data embedded in recipe \p R.
- /// Returns an optional pair, where the first element indicates whether it is
- /// an intrinsic ID.
- static std::optional<std::pair<bool, unsigned>>
- getOpcodeOrIntrinsicID(const VPSingleDefRecipe *R) {
- return TypeSwitch<const VPSingleDefRecipe *,
- std::optional<std::pair<bool, unsigned>>>(R)
- .Case<VPInstruction, VPWidenRecipe, VPWidenCastRecipe,
- VPWidenSelectRecipe, VPWidenGEPRecipe, VPReplicateRecipe>(
- [](auto *I) { return std::make_pair(false, I->getOpcode()); })
- .Case<VPWidenIntrinsicRecipe>([](auto *I) {
- return std::make_pair(true, I->getVectorIntrinsicID());
- })
- .Case<VPVectorPointerRecipe, VPPredInstPHIRecipe>([](auto *I) {
- // For recipes that do not directly map to LLVM IR instructions,
- // assign opcodes after the last VPInstruction opcode (which is also
- // after the last IR Instruction opcode), based on the VPDefID.
- return std::make_pair(false,
- VPInstruction::OpsEnd + 1 + I->getVPDefID());
- })
- .Default([](auto *) { return std::nullopt; });
- }
-
/// If recipe \p R will lower to a GEP with a non-i8 source element type,
/// return that source element type.
static Type *getGEPSourceElementType(const VPSingleDefRecipe *R) {
diff --git a/llvm/test/Transforms/LoopVectorize/constantfolder.ll b/llvm/test/Transforms/LoopVectorize/constantfolder.ll
index 4bc26a7d19a9b..0d6578532e815 100644
--- a/llvm/test/Transforms/LoopVectorize/constantfolder.ll
+++ b/llvm/test/Transforms/LoopVectorize/constantfolder.ll
@@ -295,12 +295,10 @@ define void @const_fold_binaryintrinsic(ptr %dst, i64 %d) {
; CHECK-NEXT: [[ENTRY:.*:]]
; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
; CHECK: [[VECTOR_PH]]:
-; CHECK-NEXT: [[TMP0:%.*]] = call <4 x i64> @llvm.umax.v4i64(<4 x i64> zeroinitializer, <4 x i64> splat (i64 3))
-; CHECK-NEXT: [[TMP1:%.*]] = extractelement <4 x i64> [[TMP0]], i32 3
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
; CHECK: [[VECTOR_BODY]]:
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
-; CHECK-NEXT: store i64 [[TMP1]], ptr [[DST]], align 2
+; CHECK-NEXT: store i64 3, ptr [[DST]], align 2
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i64 [[INDEX_NEXT]], 100
; CHECK-NEXT: br i1 [[TMP2]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]]
>From 94145cd2fccf51e3b0fa1ba44a7905189be74817 Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Tue, 21 Oct 2025 10:49:24 +0100
Subject: [PATCH 3/5] [VPlan] Strip anon namespace
---
llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp | 2 --
1 file changed, 2 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index f6df5668bf451..86a52cb3af1e9 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -943,7 +943,6 @@ static void recursivelyDeleteDeadRecipes(VPValue *V) {
}
}
-namespace {
/// Get any instruction opcode or intrinsic ID data embedded in recipe \p R.
/// Returns an optional pair, where the first element indicates whether it is
/// an intrinsic ID.
@@ -966,7 +965,6 @@ getOpcodeOrIntrinsicID(const VPSingleDefRecipe *R) {
})
.Default([](auto *) { return std::nullopt; });
}
-} // namespace
/// Try to fold \p R using InstSimplifyFolder. Will succeed and return a
/// non-nullptr Value for a handled opcode or intrinsic ID if corresponding \p
>From ad7dd7dbf214be7f6871d72799069feecfdd0470 Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Tue, 21 Oct 2025 17:50:09 +0100
Subject: [PATCH 4/5] [VPlan] Return VPValue from tryToFoldLiveIns
---
.../Transforms/Vectorize/VPlanTransforms.cpp | 111 ++++++++++--------
1 file changed, 59 insertions(+), 52 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 86a52cb3af1e9..6eb7f27dac9c8 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -969,12 +969,62 @@ getOpcodeOrIntrinsicID(const VPSingleDefRecipe *R) {
/// Try to fold \p R using InstSimplifyFolder. Will succeed and return a
/// non-nullptr Value for a handled opcode or intrinsic ID if corresponding \p
/// Operands are foldable live-ins.
-static Value *tryToFoldLiveIns(VPSingleDefRecipe &R,
- ArrayRef<VPValue *> Operands,
- const DataLayout &DL, VPTypeAnalysis &TypeInfo) {
- auto OpcodeOrIID = getOpcodeOrIntrinsicID(&R);
- if (!OpcodeOrIID)
+static VPValue *tryToFoldLiveIns(VPSingleDefRecipe &R,
+ ArrayRef<VPValue *> Operands,
+ const DataLayout &DL,
+ VPTypeAnalysis &TypeInfo) {
+ auto FoldIROperands = [&R, &DL, &TypeInfo](ArrayRef<Value *> Ops) -> Value * {
+ auto OpcodeOrIID = getOpcodeOrIntrinsicID(&R);
+ if (!OpcodeOrIID)
+ return nullptr;
+
+ InstSimplifyFolder Folder(DL);
+ if (OpcodeOrIID->first) {
+ if (R.getNumOperands() != 2)
+ return nullptr;
+ unsigned ID = OpcodeOrIID->second;
+ return Folder.FoldBinaryIntrinsic(ID, Ops[0], Ops[1],
+ TypeInfo.inferScalarType(&R));
+ }
+ unsigned Opcode = OpcodeOrIID->second;
+ if (Instruction::isBinaryOp(Opcode))
+ return Folder.FoldBinOp(static_cast<Instruction::BinaryOps>(Opcode),
+ Ops[0], Ops[1]);
+ if (Instruction::isCast(Opcode))
+ return Folder.FoldCast(static_cast<Instruction::CastOps>(Opcode), Ops[0],
+ TypeInfo.inferScalarType(R.getVPSingleValue()));
+ switch (Opcode) {
+ case VPInstruction::LogicalAnd:
+ return Folder.FoldSelect(Ops[0], Ops[1],
+ ConstantInt::getNullValue(Ops[1]->getType()));
+ case VPInstruction::Not:
+ return Folder.FoldBinOp(Instruction::BinaryOps::Xor, Ops[0],
+ Constant::getAllOnesValue(Ops[0]->getType()));
+ case Instruction::Select:
+ return Folder.FoldSelect(Ops[0], Ops[1], Ops[2]);
+ case Instruction::ICmp:
+ case Instruction::FCmp:
+ return Folder.FoldCmp(cast<VPRecipeWithIRFlags>(R).getPredicate(), Ops[0],
+ Ops[1]);
+ case Instruction::GetElementPtr: {
+ auto &RFlags = cast<VPRecipeWithIRFlags>(R);
+ auto *GEP = cast<GetElementPtrInst>(RFlags.getUnderlyingInstr());
+ return Folder.FoldGEP(GEP->getSourceElementType(), Ops[0],
+ Ops.drop_front(), RFlags.getGEPNoWrapFlags());
+ }
+ case VPInstruction::PtrAdd:
+ case VPInstruction::WidePtrAdd:
+ return Folder.FoldGEP(IntegerType::getInt8Ty(TypeInfo.getContext()),
+ Ops[0], Ops[1],
+ cast<VPRecipeWithIRFlags>(R).getGEPNoWrapFlags());
+ // An extract of a live-in is an extract of a broadcast, so return the
+ // broadcasted element.
+ case Instruction::ExtractElement:
+ assert(!Ops[0]->getType()->isVectorTy() && "Live-ins should be scalar");
+ return Ops[0];
+ }
return nullptr;
+ };
SmallVector<Value *, 4> Ops;
for (VPValue *Op : Operands) {
@@ -983,51 +1033,8 @@ static Value *tryToFoldLiveIns(VPSingleDefRecipe &R,
Ops.push_back(Op->getLiveInIRValue());
}
- InstSimplifyFolder Folder(DL);
- if (OpcodeOrIID->first) {
- if (R.getNumOperands() != 2)
- return nullptr;
- unsigned ID = OpcodeOrIID->second;
- return Folder.FoldBinaryIntrinsic(ID, Ops[0], Ops[1],
- TypeInfo.inferScalarType(&R));
- }
- unsigned Opcode = OpcodeOrIID->second;
- if (Instruction::isBinaryOp(Opcode))
- return Folder.FoldBinOp(static_cast<Instruction::BinaryOps>(Opcode), Ops[0],
- Ops[1]);
- if (Instruction::isCast(Opcode))
- return Folder.FoldCast(static_cast<Instruction::CastOps>(Opcode), Ops[0],
- TypeInfo.inferScalarType(R.getVPSingleValue()));
- switch (Opcode) {
- case VPInstruction::LogicalAnd:
- return Folder.FoldSelect(Ops[0], Ops[1],
- ConstantInt::getNullValue(Ops[1]->getType()));
- case VPInstruction::Not:
- return Folder.FoldBinOp(Instruction::BinaryOps::Xor, Ops[0],
- Constant::getAllOnesValue(Ops[0]->getType()));
- case Instruction::Select:
- return Folder.FoldSelect(Ops[0], Ops[1], Ops[2]);
- case Instruction::ICmp:
- case Instruction::FCmp:
- return Folder.FoldCmp(cast<VPRecipeWithIRFlags>(R).getPredicate(), Ops[0],
- Ops[1]);
- case Instruction::GetElementPtr: {
- auto &RFlags = cast<VPRecipeWithIRFlags>(R);
- auto *GEP = cast<GetElementPtrInst>(RFlags.getUnderlyingInstr());
- return Folder.FoldGEP(GEP->getSourceElementType(), Ops[0], drop_begin(Ops),
- RFlags.getGEPNoWrapFlags());
- }
- case VPInstruction::PtrAdd:
- case VPInstruction::WidePtrAdd:
- return Folder.FoldGEP(IntegerType::getInt8Ty(TypeInfo.getContext()), Ops[0],
- Ops[1],
- cast<VPRecipeWithIRFlags>(R).getGEPNoWrapFlags());
- // An extract of a live-in is an extract of a broadcast, so return the
- // broadcasted element.
- case Instruction::ExtractElement:
- assert(!Ops[0]->getType()->isVectorTy() && "Live-ins should be scalar");
- return Ops[0];
- }
+ if (Value *V = FoldIROperands(Ops))
+ return R.getParent()->getPlan()->getOrAddLiveIn(V);
return nullptr;
}
@@ -1043,8 +1050,8 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
// InstSimplifyFolder.
const DataLayout &DL =
Plan->getScalarHeader()->getIRBasicBlock()->getDataLayout();
- if (Value *V = tryToFoldLiveIns(*Def, Def->operands(), DL, TypeInfo))
- return Def->replaceAllUsesWith(Plan->getOrAddLiveIn(V));
+ if (VPValue *V = tryToFoldLiveIns(*Def, Def->operands(), DL, TypeInfo))
+ return Def->replaceAllUsesWith(V);
// Fold PredPHI LiveIn -> LiveIn.
if (auto *PredPHI = dyn_cast<VPPredInstPHIRecipe>(&R)) {
>From 4ed655367100a67c3369dc86c0adf43d84bdfe7b Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Tue, 21 Oct 2025 21:49:12 +0100
Subject: [PATCH 5/5] [VPlan] Refactor to FoldToIRValue
---
.../Transforms/Vectorize/VPlanTransforms.cpp | 20 +++++++++----------
1 file changed, 10 insertions(+), 10 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 6eb7f27dac9c8..5c46a22e348e7 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -973,11 +973,18 @@ static VPValue *tryToFoldLiveIns(VPSingleDefRecipe &R,
ArrayRef<VPValue *> Operands,
const DataLayout &DL,
VPTypeAnalysis &TypeInfo) {
- auto FoldIROperands = [&R, &DL, &TypeInfo](ArrayRef<Value *> Ops) -> Value * {
+ auto FoldToIRValue = [&]() -> Value * {
auto OpcodeOrIID = getOpcodeOrIntrinsicID(&R);
if (!OpcodeOrIID)
return nullptr;
+ SmallVector<Value *, 4> Ops;
+ for (VPValue *Op : Operands) {
+ if (!Op->isLiveIn() || !Op->getLiveInIRValue())
+ return nullptr;
+ Ops.push_back(Op->getLiveInIRValue());
+ }
+
InstSimplifyFolder Folder(DL);
if (OpcodeOrIID->first) {
if (R.getNumOperands() != 2)
@@ -1010,7 +1017,7 @@ static VPValue *tryToFoldLiveIns(VPSingleDefRecipe &R,
auto &RFlags = cast<VPRecipeWithIRFlags>(R);
auto *GEP = cast<GetElementPtrInst>(RFlags.getUnderlyingInstr());
return Folder.FoldGEP(GEP->getSourceElementType(), Ops[0],
- Ops.drop_front(), RFlags.getGEPNoWrapFlags());
+ drop_begin(Ops), RFlags.getGEPNoWrapFlags());
}
case VPInstruction::PtrAdd:
case VPInstruction::WidePtrAdd:
@@ -1026,14 +1033,7 @@ static VPValue *tryToFoldLiveIns(VPSingleDefRecipe &R,
return nullptr;
};
- SmallVector<Value *, 4> Ops;
- for (VPValue *Op : Operands) {
- if (!Op->isLiveIn() || !Op->getLiveInIRValue())
- return nullptr;
- Ops.push_back(Op->getLiveInIRValue());
- }
-
- if (Value *V = FoldIROperands(Ops))
+ if (Value *V = FoldToIRValue())
return R.getParent()->getPlan()->getOrAddLiveIn(V);
return nullptr;
}
More information about the llvm-commits
mailing list