[llvm] [VPlan] Extend tryToFoldLiveIns to fold binary intrinsics (PR #161703)
Ramkumar Ramachandra via llvm-commits
llvm-commits at lists.llvm.org
Mon Oct 20 09:14:07 PDT 2025
https://github.com/artagnon updated https://github.com/llvm/llvm-project/pull/161703
>From b82b37068dccb62073187ba195e4706140b4c99b 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 1/2] [LV] Pre-commit test for constantfolder-binaryintrinsic
---
.../LoopVectorize/constantfolder.ll | 83 +++++++++++++++++++
1 file changed, 83 insertions(+)
diff --git a/llvm/test/Transforms/LoopVectorize/constantfolder.ll b/llvm/test/Transforms/LoopVectorize/constantfolder.ll
index 66592b0ccf677..0890347f1b6d3 100644
--- a/llvm/test/Transforms/LoopVectorize/constantfolder.ll
+++ b/llvm/test/Transforms/LoopVectorize/constantfolder.ll
@@ -288,3 +288,86 @@ 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.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 [ %d, %else ], [ 0, %loop.header ]
+ %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.header, 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 edd1c045994eb2ec9aa73134b503d88f1b9b337b 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 2/2] [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 e060e7081042a..a30a494fdcb0c 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.getVPSingleValue()));
+ }
+ 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)) {
@@ -1999,29 +2027,6 @@ 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 0890347f1b6d3..bbd208e8a0586 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]+]]
More information about the llvm-commits
mailing list