[llvm] [VPlan] Unroll VPReplicateRecipe by VF. (PR #142433)
Florian Hahn via llvm-commits
llvm-commits at lists.llvm.org
Tue Jun 10 02:49:56 PDT 2025
https://github.com/fhahn updated https://github.com/llvm/llvm-project/pull/142433
>From c354bdcb27ad73a1625c1b665d4bc7773f5ec8b2 Mon Sep 17 00:00:00 2001
From: Florian Hahn <flo at fhahn.com>
Date: Sun, 1 Jun 2025 21:58:24 +0100
Subject: [PATCH 1/2] [VPlan] Unroll VPRedplicateRecipes by VF.
---
.../Transforms/Vectorize/LoopVectorize.cpp | 1 +
llvm/lib/Transforms/Vectorize/VPlan.cpp | 8 ++
llvm/lib/Transforms/Vectorize/VPlan.h | 6 ++
.../Transforms/Vectorize/VPlanAnalysis.cpp | 2 +
.../lib/Transforms/Vectorize/VPlanRecipes.cpp | 61 +++++++++-----
.../Transforms/Vectorize/VPlanTransforms.cpp | 16 ++++
.../Transforms/Vectorize/VPlanTransforms.h | 4 +
llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp | 81 +++++++++++++++++++
.../X86/fixed-order-recurrence.ll | 6 --
...terleave-ptradd-with-replicated-operand.ll | 51 +++++-------
...6-sunk-instruction-used-outside-of-loop.ll | 2 +
.../invariant-store-vectorization-2.ll | 6 +-
.../LoopVectorize/iv_outside_user.ll | 5 --
.../LoopVectorize/load-deref-pred-align.ll | 2 +
.../Transforms/LoopVectorize/struct-return.ll | 8 +-
.../Transforms/LoopVectorize/uniform-blend.ll | 4 +
16 files changed, 194 insertions(+), 69 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index fc8ebebcf21b7..d008cfc464d35 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -7291,6 +7291,7 @@ DenseMap<const SCEV *, Value *> LoopVectorizationPlanner::executePlan(
// cost model is complete for better cost estimates.
VPlanTransforms::runPass(VPlanTransforms::unrollByUF, BestVPlan, BestUF,
OrigLoop->getHeader()->getContext());
+ VPlanTransforms::runPass(VPlanTransforms::unrollByVF, BestVPlan, BestVF);
VPlanTransforms::runPass(VPlanTransforms::materializeBroadcasts, BestVPlan);
VPlanTransforms::optimizeForVFAndUF(BestVPlan, BestVF, BestUF, PSE);
VPlanTransforms::simplifyRecipes(BestVPlan, *Legal->getWidestInductionType());
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.cpp b/llvm/lib/Transforms/Vectorize/VPlan.cpp
index 1838562f26b82..a9d0182800db8 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlan.cpp
@@ -261,6 +261,14 @@ Value *VPTransformState::get(const VPValue *Def, const VPLane &Lane) {
return Data.VPV2Scalars[Def][0];
}
+ // Look through BuildVector to avoid redundant extracts.
+ // TODO: Remove once replicate regions are unrolled explicitly.
+ auto *BV = dyn_cast<VPInstruction>(Def);
+ if (Lane.getKind() == VPLane::Kind::First && BV &&
+ BV->getOpcode() == VPInstruction::BuildVector) {
+ return get(BV->getOperand(Lane.getKnownLane()), true);
+ }
+
assert(hasVectorValue(Def));
auto *VecPart = Data.VPV2Vector[Def];
if (!VecPart->getType()->isVectorTy()) {
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 273df55188c16..3bc91eb9a1443 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -907,6 +907,12 @@ class VPInstruction : public VPRecipeWithIRFlags,
BranchOnCount,
BranchOnCond,
Broadcast,
+ /// Creates a vector containing all operands. The vector element count
+ /// matches the number of operands.
+ BuildVector,
+ /// Creates a struct of vectors containing all operands. The vector element
+ /// count matches the number of operands.
+ BuildStructVector,
ComputeAnyOfResult,
ComputeFindLastIVResult,
ComputeReductionResult,
diff --git a/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp b/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
index 81fc93bbf51fd..fac7d0ee13bbd 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
@@ -107,6 +107,8 @@ Type *VPTypeAnalysis::inferScalarTypeForRecipe(const VPInstruction *R) {
case VPInstruction::CalculateTripCountMinusVF:
case VPInstruction::CanonicalIVIncrementForPart:
case VPInstruction::AnyOf:
+ case VPInstruction::BuildVector:
+ case VPInstruction::BuildStructVector:
return SetResultTyFromOp();
case VPInstruction::FirstActiveLane:
return Type::getIntNTy(Ctx, 64);
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 90a04af60e3d8..a2b494e85611c 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -493,6 +493,9 @@ Value *VPInstruction::generate(VPTransformState &State) {
}
case Instruction::ExtractElement: {
assert(State.VF.isVector() && "Only extract elements from vectors");
+ return State.get(getOperand(0),
+ VPLane(cast<ConstantInt>(getOperand(1)->getLiveInIRValue())
+ ->getZExtValue()));
Value *Vec = State.get(getOperand(0));
Value *Idx = State.get(getOperand(1), /*IsScalar=*/true);
return Builder.CreateExtractElement(Vec, Idx, Name);
@@ -604,6 +607,33 @@ Value *VPInstruction::generate(VPTransformState &State) {
return Builder.CreateVectorSplat(
State.VF, State.get(getOperand(0), /*IsScalar*/ true), "broadcast");
}
+ case VPInstruction::BuildVector: {
+ auto *ScalarTy = State.TypeAnalysis.inferScalarType(getOperand(0));
+ Value *Res = PoisonValue::get(
+ toVectorizedTy(ScalarTy, ElementCount::getFixed(getNumOperands())));
+ for (const auto &[Idx, Op] : enumerate(operands()))
+ Res = State.Builder.CreateInsertElement(Res, State.get(Op, true),
+ State.Builder.getInt32(Idx));
+ return Res;
+ }
+ case VPInstruction::BuildStructVector: {
+ // For struct types, we need to build a new 'wide' struct type, where each
+ // element is widened.
+ auto *STy =
+ cast<StructType>(State.TypeAnalysis.inferScalarType(getOperand(0)));
+ Value *Res = PoisonValue::get(
+ toVectorizedTy(STy, ElementCount::getFixed(getNumOperands())));
+ for (const auto &[Idx, Op] : enumerate(operands())) {
+ for (unsigned I = 0, E = STy->getNumElements(); I != E; I++) {
+ Value *ScalarValue = Builder.CreateExtractValue(State.get(Op, true), I);
+ Value *VectorValue = Builder.CreateExtractValue(Res, I);
+ VectorValue =
+ Builder.CreateInsertElement(VectorValue, ScalarValue, Idx);
+ Res = Builder.CreateInsertValue(Res, VectorValue, I);
+ }
+ }
+ return Res;
+ }
case VPInstruction::ComputeAnyOfResult: {
// FIXME: The cross-recipe dependency on VPReductionPHIRecipe is temporary
// and will be removed by breaking up the recipe further.
@@ -872,10 +902,11 @@ void VPInstruction::execute(VPTransformState &State) {
if (!hasResult())
return;
assert(GeneratedValue && "generate must produce a value");
- assert(
- (GeneratedValue->getType()->isVectorTy() == !GeneratesPerFirstLaneOnly ||
- State.VF.isScalar()) &&
- "scalar value but not only first lane defined");
+ assert((((GeneratedValue->getType()->isVectorTy() ||
+ GeneratedValue->getType()->isStructTy()) ==
+ !GeneratesPerFirstLaneOnly) ||
+ State.VF.isScalar()) &&
+ "scalar value but not only first lane defined");
State.set(this, GeneratedValue,
/*IsScalar*/ GeneratesPerFirstLaneOnly);
}
@@ -889,6 +920,8 @@ bool VPInstruction::opcodeMayReadOrWriteFromMemory() const {
case Instruction::ICmp:
case Instruction::Select:
case VPInstruction::AnyOf:
+ case VPInstruction::BuildVector:
+ case VPInstruction::BuildStructVector:
case VPInstruction::CalculateTripCountMinusVF:
case VPInstruction::CanonicalIVIncrementForPart:
case VPInstruction::ExtractLastElement:
@@ -1008,6 +1041,12 @@ void VPInstruction::print(raw_ostream &O, const Twine &Indent,
case VPInstruction::Broadcast:
O << "broadcast";
break;
+ case VPInstruction::BuildVector:
+ O << "buildvector";
+ break;
+ case VPInstruction::BuildStructVector:
+ O << "buildstructvector";
+ break;
case VPInstruction::ExtractLastElement:
O << "extract-last-element";
break;
@@ -2763,20 +2802,6 @@ void VPReplicateRecipe::execute(VPTransformState &State) {
scalarizeInstruction(UI, this, VPLane(0), State);
return;
}
-
- // A store of a loop varying value to a uniform address only needs the last
- // copy of the store.
- if (isa<StoreInst>(UI) && vputils::isSingleScalar(getOperand(1))) {
- auto Lane = VPLane::getLastLaneForVF(State.VF);
- scalarizeInstruction(UI, this, VPLane(Lane), State);
- return;
- }
-
- // Generate scalar instances for all VF lanes.
- assert(!State.VF.isScalable() && "Can't scalarize a scalable vector");
- const unsigned EndLane = State.VF.getKnownMinValue();
- for (unsigned Lane = 0; Lane < EndLane; ++Lane)
- scalarizeInstruction(UI, this, VPLane(Lane), State);
}
bool VPReplicateRecipe::shouldPack() const {
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index ea617f042566b..4177b5b1b956d 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1140,6 +1140,22 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
return;
}
+ // Look through Extract(Last|Penultimate)Element (BuildVector ....).
+ if (match(&R,
+ m_VPInstruction<VPInstruction::ExtractLastElement>(m_VPValue(A))) ||
+ match(&R, m_VPInstruction<VPInstruction::ExtractPenultimateElement>(
+ m_VPValue(A)))) {
+ unsigned Offset = cast<VPInstruction>(&R)->getOpcode() ==
+ VPInstruction::ExtractLastElement
+ ? 1
+ : 2;
+ auto *BV = dyn_cast<VPInstruction>(A);
+ if (BV && BV->getOpcode() == VPInstruction::BuildVector) {
+ Def->replaceAllUsesWith(BV->getOperand(BV->getNumOperands() - Offset));
+ return;
+ }
+ }
+
// Some simplifications can only be applied after unrolling. Perform them
// below.
if (!Plan->isUnrolled())
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
index 34e2de4eb3b74..f45b7a7969d04 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
@@ -99,6 +99,10 @@ struct VPlanTransforms {
/// Explicitly unroll \p Plan by \p UF.
static void unrollByUF(VPlan &Plan, unsigned UF, LLVMContext &Ctx);
+ /// Explicitly unroll VPReplicateRecipes outside of replicate regions by \p
+ /// VF.
+ static void unrollByVF(VPlan &Plan, ElementCount VF);
+
/// Optimize \p Plan based on \p BestVF and \p BestUF. This may restrict the
/// resulting plan to \p BestVF and \p BestUF.
static void optimizeForVFAndUF(VPlan &Plan, ElementCount BestVF,
diff --git a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
index 335301a927ceb..e48611ab4b923 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
@@ -15,6 +15,7 @@
#include "VPlan.h"
#include "VPlanAnalysis.h"
#include "VPlanCFG.h"
+#include "VPlanHelpers.h"
#include "VPlanPatternMatch.h"
#include "VPlanTransforms.h"
#include "VPlanUtils.h"
@@ -430,3 +431,83 @@ void VPlanTransforms::unrollByUF(VPlan &Plan, unsigned UF, LLVMContext &Ctx) {
VPlanTransforms::removeDeadRecipes(Plan);
}
+
+/// Create a single-scalar clone of RepR for lane \p Lane.
+static VPReplicateRecipe *cloneForLane(VPlan &Plan, VPBuilder &Builder,
+ Type *IdxTy, VPReplicateRecipe *RepR,
+ VPLane Lane) {
+ // Collect the operands at Lane, creating extracts as needed.
+ SmallVector<VPValue *> NewOps;
+ for (VPValue *Op : RepR->operands()) {
+ if (vputils::isSingleScalar(Op)) {
+ NewOps.push_back(Op);
+ continue;
+ }
+ VPValue *Ext;
+ if (Lane.getKind() == VPLane::Kind::ScalableLast) {
+ Ext = Builder.createNaryOp(VPInstruction::ExtractLastElement, {Op});
+ } else {
+ // Look through buildvector to avoid unnecessary extracts.
+ auto *BV = dyn_cast<VPInstruction>(Op);
+ if (BV && BV->getOpcode() == VPInstruction::BuildVector) {
+ NewOps.push_back(BV->getOperand(Lane.getKnownLane()));
+ continue;
+ }
+ VPValue *Idx =
+ Plan.getOrAddLiveIn(ConstantInt::get(IdxTy, Lane.getKnownLane()));
+ Ext = Builder.createNaryOp(Instruction::ExtractElement, {Op, Idx});
+ }
+ NewOps.push_back(Ext);
+ }
+
+ auto *New =
+ new VPReplicateRecipe(RepR->getUnderlyingInstr(), NewOps,
+ /*IsSingleScalar=*/true, /*Mask=*/nullptr, *RepR);
+ New->insertBefore(RepR);
+ return New;
+}
+
+void VPlanTransforms::unrollByVF(VPlan &Plan, ElementCount VF) {
+ Type *IdxTy = IntegerType::get(
+ Plan.getScalarHeader()->getIRBasicBlock()->getContext(), 32);
+ for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
+ vp_depth_first_shallow(Plan.getVectorLoopRegion()->getEntry()))) {
+ for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
+ auto *RepR = dyn_cast<VPReplicateRecipe>(&R);
+ if (!RepR || RepR->isSingleScalar())
+ continue;
+
+ VPBuilder Builder(RepR);
+ SmallVector<VPValue *> LaneDefs;
+ // Stores to invariant addresses only need to store the last lane.
+ if (isa<StoreInst>(RepR->getUnderlyingInstr()) &&
+ vputils::isSingleScalar(RepR->getOperand(1))) {
+ cloneForLane(Plan, Builder, IdxTy, RepR, VPLane::getLastLaneForVF(VF));
+ RepR->eraseFromParent();
+ continue;
+ }
+
+ /// Create single-scalar version of RepR for all lanes.
+ for (unsigned I = 0; I != VF.getKnownMinValue(); ++I)
+ LaneDefs.push_back(cloneForLane(Plan, Builder, IdxTy, RepR, VPLane(I)));
+
+ /// Users that only demand the first lane can use the definition for lane
+ /// 0.
+ RepR->replaceUsesWithIf(LaneDefs[0], [RepR](VPUser &U, unsigned) {
+ return U.onlyFirstLaneUsed(RepR);
+ });
+
+ Type *ResTy = RepR->getUnderlyingInstr()->getType();
+ // If needed, create a Build(Struct)Vector recipe to insert the scalar
+ // lane values into a vector.
+ if (!ResTy->isVoidTy()) {
+ VPValue *VecRes = Builder.createNaryOp(
+ ResTy->isStructTy() ? VPInstruction::BuildStructVector
+ : VPInstruction::BuildVector,
+ LaneDefs);
+ RepR->replaceAllUsesWith(VecRes);
+ }
+ RepR->eraseFromParent();
+ }
+ }
+}
diff --git a/llvm/test/Transforms/LoopVectorize/X86/fixed-order-recurrence.ll b/llvm/test/Transforms/LoopVectorize/X86/fixed-order-recurrence.ll
index 83e9d6146755d..743aedee38012 100644
--- a/llvm/test/Transforms/LoopVectorize/X86/fixed-order-recurrence.ll
+++ b/llvm/test/Transforms/LoopVectorize/X86/fixed-order-recurrence.ll
@@ -398,12 +398,6 @@ define void @test_for_tried_to_force_scalar(ptr noalias %A, ptr noalias %B, ptr
; CHECK-NEXT: [[STRIDED_VEC:%.*]] = shufflevector <12 x float> [[WIDE_VEC]], <12 x float> poison, <4 x i32> <i32 0, i32 3, i32 6, i32 9>
; CHECK-NEXT: [[TMP30:%.*]] = extractelement <4 x float> [[STRIDED_VEC]], i32 3
; CHECK-NEXT: store float [[TMP30]], ptr [[C:%.*]], align 4
-; CHECK-NEXT: [[TMP31:%.*]] = extractelement <4 x ptr> [[TMP29]], i32 0
-; CHECK-NEXT: [[TMP38:%.*]] = load float, ptr [[TMP31]], align 4
-; CHECK-NEXT: [[TMP33:%.*]] = extractelement <4 x ptr> [[TMP29]], i32 1
-; CHECK-NEXT: [[TMP32:%.*]] = load float, ptr [[TMP33]], align 4
-; CHECK-NEXT: [[TMP35:%.*]] = extractelement <4 x ptr> [[TMP29]], i32 2
-; CHECK-NEXT: [[TMP34:%.*]] = load float, ptr [[TMP35]], align 4
; CHECK-NEXT: [[TMP37:%.*]] = extractelement <4 x ptr> [[TMP29]], i32 3
; CHECK-NEXT: [[TMP36:%.*]] = load float, ptr [[TMP37]], align 4
; CHECK-NEXT: store float [[TMP36]], ptr [[B:%.*]], align 4
diff --git a/llvm/test/Transforms/LoopVectorize/X86/interleave-ptradd-with-replicated-operand.ll b/llvm/test/Transforms/LoopVectorize/X86/interleave-ptradd-with-replicated-operand.ll
index cdc7839bfc0f0..95258e65bbe3d 100644
--- a/llvm/test/Transforms/LoopVectorize/X86/interleave-ptradd-with-replicated-operand.ll
+++ b/llvm/test/Transforms/LoopVectorize/X86/interleave-ptradd-with-replicated-operand.ll
@@ -32,42 +32,31 @@ define ptr @test_interleave_ptradd_with_replicated_op(ptr %m) #0 {
; CHECK-NEXT: [[TMP13:%.*]] = add i64 [[OFFSET_IDX]], 104
; CHECK-NEXT: [[TMP14:%.*]] = add i64 [[OFFSET_IDX]], 112
; CHECK-NEXT: [[TMP15:%.*]] = add i64 [[OFFSET_IDX]], 120
-; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP0]]
+; CHECK-NEXT: [[NEXT_GEP12:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP0]]
; CHECK-NEXT: [[NEXT_GEP2:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP1]]
; CHECK-NEXT: [[NEXT_GEP3:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP2]]
; CHECK-NEXT: [[NEXT_GEP4:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP3]]
-; CHECK-NEXT: [[NEXT_GEP5:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP4]]
+; CHECK-NEXT: [[NEXT_GEP13:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP4]]
; CHECK-NEXT: [[NEXT_GEP6:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP5]]
; CHECK-NEXT: [[NEXT_GEP7:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP6]]
; CHECK-NEXT: [[NEXT_GEP8:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP7]]
-; CHECK-NEXT: [[NEXT_GEP9:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP8]]
+; CHECK-NEXT: [[NEXT_GEP14:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP8]]
; CHECK-NEXT: [[NEXT_GEP10:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP9]]
; CHECK-NEXT: [[NEXT_GEP11:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP10]]
-; CHECK-NEXT: [[NEXT_GEP12:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP11]]
-; CHECK-NEXT: [[NEXT_GEP13:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP12]]
-; CHECK-NEXT: [[NEXT_GEP14:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP13]]
-; CHECK-NEXT: [[NEXT_GEP15:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP14]]
+; CHECK-NEXT: [[NEXT_GEP17:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP11]]
+; CHECK-NEXT: [[NEXT_GEP15:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP12]]
+; CHECK-NEXT: [[NEXT_GEP18:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP13]]
+; CHECK-NEXT: [[NEXT_GEP19:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP14]]
; CHECK-NEXT: [[NEXT_GEP16:%.*]] = getelementptr i8, ptr [[M]], i64 [[TMP15]]
-; CHECK-NEXT: [[TMP16:%.*]] = getelementptr i8, ptr [[NEXT_GEP]], i64 4
-; CHECK-NEXT: [[TMP17:%.*]] = getelementptr i8, ptr [[NEXT_GEP2]], i64 4
-; CHECK-NEXT: [[TMP18:%.*]] = getelementptr i8, ptr [[NEXT_GEP3]], i64 4
-; CHECK-NEXT: [[TMP19:%.*]] = getelementptr i8, ptr [[NEXT_GEP4]], i64 4
-; CHECK-NEXT: [[TMP20:%.*]] = getelementptr i8, ptr [[NEXT_GEP5]], i64 4
-; CHECK-NEXT: [[TMP21:%.*]] = getelementptr i8, ptr [[NEXT_GEP6]], i64 4
-; CHECK-NEXT: [[TMP22:%.*]] = getelementptr i8, ptr [[NEXT_GEP7]], i64 4
-; CHECK-NEXT: [[TMP23:%.*]] = getelementptr i8, ptr [[NEXT_GEP8]], i64 4
-; CHECK-NEXT: [[TMP24:%.*]] = getelementptr i8, ptr [[NEXT_GEP9]], i64 4
-; CHECK-NEXT: [[TMP25:%.*]] = getelementptr i8, ptr [[NEXT_GEP10]], i64 4
-; CHECK-NEXT: [[TMP26:%.*]] = getelementptr i8, ptr [[NEXT_GEP11]], i64 4
; CHECK-NEXT: [[TMP27:%.*]] = getelementptr i8, ptr [[NEXT_GEP12]], i64 4
; CHECK-NEXT: [[TMP28:%.*]] = getelementptr i8, ptr [[NEXT_GEP13]], i64 4
; CHECK-NEXT: [[TMP29:%.*]] = getelementptr i8, ptr [[NEXT_GEP14]], i64 4
; CHECK-NEXT: [[TMP30:%.*]] = getelementptr i8, ptr [[NEXT_GEP15]], i64 4
; CHECK-NEXT: [[TMP31:%.*]] = getelementptr i8, ptr [[NEXT_GEP16]], i64 4
-; CHECK-NEXT: [[TMP32:%.*]] = getelementptr i8, ptr [[TMP16]], i32 -4
-; CHECK-NEXT: [[TMP33:%.*]] = getelementptr i8, ptr [[TMP20]], i32 -4
-; CHECK-NEXT: [[TMP34:%.*]] = getelementptr i8, ptr [[TMP24]], i32 -4
-; CHECK-NEXT: [[TMP35:%.*]] = getelementptr i8, ptr [[TMP28]], i32 -4
+; CHECK-NEXT: [[TMP32:%.*]] = getelementptr i8, ptr [[TMP27]], i32 -4
+; CHECK-NEXT: [[TMP33:%.*]] = getelementptr i8, ptr [[TMP28]], i32 -4
+; CHECK-NEXT: [[TMP34:%.*]] = getelementptr i8, ptr [[TMP29]], i32 -4
+; CHECK-NEXT: [[TMP35:%.*]] = getelementptr i8, ptr [[TMP30]], i32 -4
; CHECK-NEXT: [[WIDE_VEC:%.*]] = load <8 x i32>, ptr [[TMP32]], align 4
; CHECK-NEXT: [[STRIDED_VEC:%.*]] = shufflevector <8 x i32> [[WIDE_VEC]], <8 x i32> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
; CHECK-NEXT: [[STRIDED_VEC17:%.*]] = shufflevector <8 x i32> [[WIDE_VEC]], <8 x i32> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
@@ -85,7 +74,7 @@ define ptr @test_interleave_ptradd_with_replicated_op(ptr %m) #0 {
; CHECK-NEXT: [[TMP38:%.*]] = add <4 x i32> [[STRIDED_VEC23]], [[STRIDED_VEC22]]
; CHECK-NEXT: [[TMP39:%.*]] = add <4 x i32> [[STRIDED_VEC26]], [[STRIDED_VEC25]]
; CHECK-NEXT: [[TMP40:%.*]] = extractelement <4 x i32> [[TMP36]], i32 0
-; CHECK-NEXT: store i32 [[TMP40]], ptr [[NEXT_GEP]], align 4
+; CHECK-NEXT: store i32 [[TMP40]], ptr [[NEXT_GEP12]], align 4
; CHECK-NEXT: [[TMP41:%.*]] = extractelement <4 x i32> [[TMP36]], i32 1
; CHECK-NEXT: store i32 [[TMP41]], ptr [[NEXT_GEP2]], align 4
; CHECK-NEXT: [[TMP42:%.*]] = extractelement <4 x i32> [[TMP36]], i32 2
@@ -93,7 +82,7 @@ define ptr @test_interleave_ptradd_with_replicated_op(ptr %m) #0 {
; CHECK-NEXT: [[TMP43:%.*]] = extractelement <4 x i32> [[TMP36]], i32 3
; CHECK-NEXT: store i32 [[TMP43]], ptr [[NEXT_GEP4]], align 4
; CHECK-NEXT: [[TMP44:%.*]] = extractelement <4 x i32> [[TMP37]], i32 0
-; CHECK-NEXT: store i32 [[TMP44]], ptr [[NEXT_GEP5]], align 4
+; CHECK-NEXT: store i32 [[TMP44]], ptr [[NEXT_GEP13]], align 4
; CHECK-NEXT: [[TMP45:%.*]] = extractelement <4 x i32> [[TMP37]], i32 1
; CHECK-NEXT: store i32 [[TMP45]], ptr [[NEXT_GEP6]], align 4
; CHECK-NEXT: [[TMP46:%.*]] = extractelement <4 x i32> [[TMP37]], i32 2
@@ -101,19 +90,19 @@ define ptr @test_interleave_ptradd_with_replicated_op(ptr %m) #0 {
; CHECK-NEXT: [[TMP47:%.*]] = extractelement <4 x i32> [[TMP37]], i32 3
; CHECK-NEXT: store i32 [[TMP47]], ptr [[NEXT_GEP8]], align 4
; CHECK-NEXT: [[TMP48:%.*]] = extractelement <4 x i32> [[TMP38]], i32 0
-; CHECK-NEXT: store i32 [[TMP48]], ptr [[NEXT_GEP9]], align 4
+; CHECK-NEXT: store i32 [[TMP48]], ptr [[NEXT_GEP14]], align 4
; CHECK-NEXT: [[TMP49:%.*]] = extractelement <4 x i32> [[TMP38]], i32 1
; CHECK-NEXT: store i32 [[TMP49]], ptr [[NEXT_GEP10]], align 4
; CHECK-NEXT: [[TMP50:%.*]] = extractelement <4 x i32> [[TMP38]], i32 2
; CHECK-NEXT: store i32 [[TMP50]], ptr [[NEXT_GEP11]], align 4
; CHECK-NEXT: [[TMP51:%.*]] = extractelement <4 x i32> [[TMP38]], i32 3
-; CHECK-NEXT: store i32 [[TMP51]], ptr [[NEXT_GEP12]], align 4
+; CHECK-NEXT: store i32 [[TMP51]], ptr [[NEXT_GEP17]], align 4
; CHECK-NEXT: [[TMP52:%.*]] = extractelement <4 x i32> [[TMP39]], i32 0
-; CHECK-NEXT: store i32 [[TMP52]], ptr [[NEXT_GEP13]], align 4
+; CHECK-NEXT: store i32 [[TMP52]], ptr [[NEXT_GEP15]], align 4
; CHECK-NEXT: [[TMP53:%.*]] = extractelement <4 x i32> [[TMP39]], i32 1
-; CHECK-NEXT: store i32 [[TMP53]], ptr [[NEXT_GEP14]], align 4
+; CHECK-NEXT: store i32 [[TMP53]], ptr [[NEXT_GEP18]], align 4
; CHECK-NEXT: [[TMP54:%.*]] = extractelement <4 x i32> [[TMP39]], i32 2
-; CHECK-NEXT: store i32 [[TMP54]], ptr [[NEXT_GEP15]], align 4
+; CHECK-NEXT: store i32 [[TMP54]], ptr [[NEXT_GEP19]], align 4
; CHECK-NEXT: [[TMP55:%.*]] = extractelement <4 x i32> [[TMP39]], i32 3
; CHECK-NEXT: store i32 [[TMP55]], ptr [[NEXT_GEP16]], align 4
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 16
@@ -123,11 +112,11 @@ define ptr @test_interleave_ptradd_with_replicated_op(ptr %m) #0 {
; CHECK-NEXT: br i1 false, label %[[EXIT:.*]], label %[[SCALAR_PH]]
; CHECK: [[SCALAR_PH]]:
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi ptr [ [[IND_END]], %[[MIDDLE_BLOCK]] ], [ [[M]], %[[ENTRY]] ]
-; CHECK-NEXT: [[BC_RESUME_VAL1:%.*]] = phi i32 [ 97, %[[MIDDLE_BLOCK]] ], [ 1, %[[ENTRY]] ]
+; CHECK-NEXT: [[BC_RESUME_VAL26:%.*]] = phi i32 [ 97, %[[MIDDLE_BLOCK]] ], [ 1, %[[ENTRY]] ]
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
; CHECK-NEXT: [[PTR_IV:%.*]] = phi ptr [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[PTR_IV_NEXT:%.*]], %[[LOOP]] ]
-; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL1]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL26]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
; CHECK-NEXT: [[PTR_IV_NEXT]] = getelementptr i8, ptr [[PTR_IV]], i64 8
; CHECK-NEXT: [[P_4:%.*]] = getelementptr i8, ptr [[PTR_IV]], i64 4
; CHECK-NEXT: [[L_1:%.*]] = load i32, ptr [[P_4]], align 4
diff --git a/llvm/test/Transforms/LoopVectorize/X86/pr51366-sunk-instruction-used-outside-of-loop.ll b/llvm/test/Transforms/LoopVectorize/X86/pr51366-sunk-instruction-used-outside-of-loop.ll
index 644f10b617eb7..db9be20f7d190 100644
--- a/llvm/test/Transforms/LoopVectorize/X86/pr51366-sunk-instruction-used-outside-of-loop.ll
+++ b/llvm/test/Transforms/LoopVectorize/X86/pr51366-sunk-instruction-used-outside-of-loop.ll
@@ -10,6 +10,8 @@ define ptr @test(ptr noalias %src, ptr noalias %dst) {
; CHECK-NEXT: [[TMP1:%.*]] = add i64 [[INDEX]], 1
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[SRC:%.*]], i64 [[TMP0]]
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i32, ptr [[SRC]], i64 [[TMP1]]
+; CHECK-NEXT: [[TMP16:%.*]] = insertelement <2 x ptr> poison, ptr [[TMP6]], i32 0
+; CHECK-NEXT: [[TMP18:%.*]] = insertelement <2 x ptr> [[TMP16]], ptr [[TMP2]], i32 1
; CHECK-NEXT: [[TMP3:%.*]] = icmp eq <2 x i64> [[VEC_IND]], zeroinitializer
; CHECK-NEXT: [[TMP4:%.*]] = xor <2 x i1> [[TMP3]], splat (i1 true)
; CHECK-NEXT: [[TMP5:%.*]] = extractelement <2 x i1> [[TMP4]], i32 0
diff --git a/llvm/test/Transforms/LoopVectorize/invariant-store-vectorization-2.ll b/llvm/test/Transforms/LoopVectorize/invariant-store-vectorization-2.ll
index 0306567c5ce98..a02fddc4cf72d 100644
--- a/llvm/test/Transforms/LoopVectorize/invariant-store-vectorization-2.ll
+++ b/llvm/test/Transforms/LoopVectorize/invariant-store-vectorization-2.ll
@@ -134,18 +134,14 @@ define void @inv_val_store_to_inv_address_conditional_inv(ptr %a, i64 %n, ptr %b
; CHECK-NEXT: br i1 [[FOUND_CONFLICT]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]]
; CHECK: vector.ph:
; CHECK-NEXT: [[N_VEC:%.*]] = and i64 [[SMAX2]], 9223372036854775804
-; CHECK-NEXT: [[BROADCAST_SPLAT6:%.*]] = insertelement <4 x i32> poison, i32 [[K]], i64 0
-; CHECK-NEXT: [[TMP1:%.*]] = insertelement <4 x i1> poison, i1 [[CMP]], i64 0
; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[NTRUNC]], i64 0
; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
-; CHECK-NEXT: [[PREDPHI:%.*]] = select <4 x i1> [[TMP1]], <4 x i32> [[BROADCAST_SPLAT]], <4 x i32> [[BROADCAST_SPLAT6]]
-; CHECK-NEXT: [[TMP2:%.*]] = extractelement <4 x i32> [[PREDPHI]], i64 0
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
; CHECK: vector.body:
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDEX]]
; CHECK-NEXT: store <4 x i32> [[BROADCAST_SPLAT]], ptr [[TMP3]], align 4, !alias.scope [[META9:![0-9]+]], !noalias [[META12:![0-9]+]]
-; CHECK-NEXT: store i32 [[TMP2]], ptr [[A]], align 4, !alias.scope [[META12]]
+; CHECK-NEXT: store i32 [[K]], ptr [[A]], align 4, !alias.scope [[META12]]
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP14:![0-9]+]]
diff --git a/llvm/test/Transforms/LoopVectorize/iv_outside_user.ll b/llvm/test/Transforms/LoopVectorize/iv_outside_user.ll
index f801443b85d3f..7b7735434e67d 100644
--- a/llvm/test/Transforms/LoopVectorize/iv_outside_user.ll
+++ b/llvm/test/Transforms/LoopVectorize/iv_outside_user.ll
@@ -1101,7 +1101,6 @@ define i32 @test_iv_uniform_with_outside_use_scev_simplification(ptr %dst) {
; VEC-NEXT: [[TMP1:%.*]] = getelementptr inbounds i16, ptr [[DST]], i32 [[TMP0]]
; VEC-NEXT: [[TMP2:%.*]] = getelementptr inbounds i16, ptr [[TMP1]], i32 0
; VEC-NEXT: store <2 x i16> zeroinitializer, ptr [[TMP2]], align 2
-; VEC-NEXT: [[TMP4:%.*]] = add i32 [[STEP_2]], [[TMP0]]
; VEC-NEXT: [[TMP5:%.*]] = add i32 [[STEP_2]], [[TMP6]]
; VEC-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
; VEC-NEXT: [[TMP3:%.*]] = icmp eq i32 [[INDEX_NEXT]], 8
@@ -1293,8 +1292,6 @@ define i32 @iv_ext_used_outside( ptr %dst) {
; VEC-NEXT: [[TMP2:%.*]] = getelementptr inbounds nuw i32, ptr [[TMP1]], i32 0
; VEC-NEXT: store <2 x i32> zeroinitializer, ptr [[TMP2]], align 4
; VEC-NEXT: [[TMP5:%.*]] = add nuw nsw <2 x i16> [[VEC_IND]], splat (i16 1)
-; VEC-NEXT: [[TMP3:%.*]] = extractelement <2 x i16> [[TMP5]], i32 0
-; VEC-NEXT: [[TMP4:%.*]] = zext nneg i16 [[TMP3]] to i32
; VEC-NEXT: [[TMP8:%.*]] = extractelement <2 x i16> [[TMP5]], i32 1
; VEC-NEXT: [[TMP7:%.*]] = zext nneg i16 [[TMP8]] to i32
; VEC-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
@@ -1389,9 +1386,7 @@ define i64 @test_iv_increment_incremented(ptr %dst) {
; VEC-NEXT: [[TMP1:%.*]] = getelementptr i16, ptr [[TMP0]], i32 0
; VEC-NEXT: [[TMP2:%.*]] = getelementptr i16, ptr [[TMP1]], i32 -1
; VEC-NEXT: store <2 x i16> splat (i16 1), ptr [[TMP2]], align 2
-; VEC-NEXT: [[TMP3:%.*]] = add i64 2, -1
; VEC-NEXT: [[TMP5:%.*]] = add i64 1, -1
-; VEC-NEXT: [[TMP4:%.*]] = add i64 [[TMP3]], 1
; VEC-NEXT: [[TMP6:%.*]] = add i64 [[TMP5]], 1
; VEC-NEXT: br label %[[MIDDLE_BLOCK:.*]]
; VEC: [[MIDDLE_BLOCK]]:
diff --git a/llvm/test/Transforms/LoopVectorize/load-deref-pred-align.ll b/llvm/test/Transforms/LoopVectorize/load-deref-pred-align.ll
index 7a33fd092e293..8a326c9d0c083 100644
--- a/llvm/test/Transforms/LoopVectorize/load-deref-pred-align.ll
+++ b/llvm/test/Transforms/LoopVectorize/load-deref-pred-align.ll
@@ -669,6 +669,8 @@ define void @test_rev_loops_strided_deref_loads(ptr nocapture noundef writeonly
; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1024 x i32], ptr [[LOCAL_SRC]], i64 0, i64 [[TMP9]]
; CHECK-NEXT: [[TMP11:%.*]] = load i32, ptr [[TMP8]], align 4
; CHECK-NEXT: [[TMP12:%.*]] = load i32, ptr [[TMP10]], align 4
+; CHECK-NEXT: [[TMP23:%.*]] = insertelement <2 x i32> poison, i32 [[TMP11]], i32 0
+; CHECK-NEXT: [[TMP24:%.*]] = insertelement <2 x i32> [[TMP23]], i32 [[TMP12]], i32 1
; CHECK-NEXT: [[TMP13:%.*]] = extractelement <2 x i1> [[TMP5]], i32 0
; CHECK-NEXT: br i1 [[TMP13]], label [[PRED_STORE_IF:%.*]], label [[PRED_STORE_CONTINUE:%.*]]
; CHECK: pred.store.if:
diff --git a/llvm/test/Transforms/LoopVectorize/struct-return.ll b/llvm/test/Transforms/LoopVectorize/struct-return.ll
index 50b9ba12af82d..0c5cf09430bb9 100644
--- a/llvm/test/Transforms/LoopVectorize/struct-return.ll
+++ b/llvm/test/Transforms/LoopVectorize/struct-return.ll
@@ -79,20 +79,20 @@ define void @struct_return_f32_replicate(ptr noalias %in, ptr noalias writeonly
; CHECK: [[CALL_LANE_1:%.*]] = tail call { float, float } @foo(float {{%.*}})
; // Lane 0
; CHECK: [[A_0:%.*]] = extractvalue { float, float } [[CALL_LANE_0]], 0
-; CHECK: [[VEC_A_0:%.*]] = insertelement <2 x float> poison, float [[A_0]], i32 0
+; CHECK: [[VEC_A_0:%.*]] = insertelement <2 x float> poison, float [[A_0]], i64 0
; CHECK: [[WIDE_A_0:%.*]] = insertvalue { <2 x float>, <2 x float> } poison, <2 x float> [[VEC_A_0]], 0
; CHECK: [[B_0:%.*]] = extractvalue { float, float } [[CALL_LANE_0]], 1
; CHECK: [[UNDEF_B_0:%.*]] = extractvalue { <2 x float>, <2 x float> } [[WIDE_A_0]], 1
-; CHECK: [[VEC_B_0:%.*]] = insertelement <2 x float> [[UNDEF_B_0]], float [[B_0]], i32 0
+; CHECK: [[VEC_B_0:%.*]] = insertelement <2 x float> [[UNDEF_B_0]], float [[B_0]], i64 0
; CHECK: [[WIDE_0:%.*]] = insertvalue { <2 x float>, <2 x float> } [[WIDE_A_0]], <2 x float> [[VEC_B_0]], 1
; // Lane 1
; CHECK: [[A_1:%.*]] = extractvalue { float, float } [[CALL_LANE_1]], 0
; CHECK: [[VEC_A_0_EXT:%.*]] = extractvalue { <2 x float>, <2 x float> } [[WIDE_0]], 0
-; CHECK: [[VEC_A:%.*]] = insertelement <2 x float> [[VEC_A_0_EXT]], float [[A_1]], i32 1
+; CHECK: [[VEC_A:%.*]] = insertelement <2 x float> [[VEC_A_0_EXT]], float [[A_1]], i64 1
; CHECK: [[WIDE_A:%.*]] = insertvalue { <2 x float>, <2 x float> } [[WIDE_0]], <2 x float> [[VEC_A]], 0
; CHECK: [[B_1:%.*]] = extractvalue { float, float } [[CALL_LANE_1]], 1
; CHECK: [[VEC_B_0_EXT:%.*]] = extractvalue { <2 x float>, <2 x float> } [[WIDE_A]], 1
-; CHECK: [[VEC_B:%.*]] = insertelement <2 x float> [[VEC_B_0_EXT]], float [[B_1]], i32 1
+; CHECK: [[VEC_B:%.*]] = insertelement <2 x float> [[VEC_B_0_EXT]], float [[B_1]], i64 1
; CHECK: [[WIDE:%.*]] = insertvalue { <2 x float>, <2 x float> } [[WIDE_A]], <2 x float> [[VEC_B]], 1
; // Store wide values:
; CHECK: [[VEC_A_EXT:%.*]] = extractvalue { <2 x float>, <2 x float> } [[WIDE]], 0
diff --git a/llvm/test/Transforms/LoopVectorize/uniform-blend.ll b/llvm/test/Transforms/LoopVectorize/uniform-blend.ll
index 130db548ca8cb..bdd7fef37e66d 100644
--- a/llvm/test/Transforms/LoopVectorize/uniform-blend.ll
+++ b/llvm/test/Transforms/LoopVectorize/uniform-blend.ll
@@ -229,6 +229,10 @@ define void @redundant_branch_and_blends_without_mask(ptr %A) {
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP1]]
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP2]]
; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP3]]
+; CHECK-NEXT: [[TMP35:%.*]] = insertelement <4 x ptr> poison, ptr [[TMP5]], i32 0
+; CHECK-NEXT: [[TMP36:%.*]] = insertelement <4 x ptr> [[TMP35]], ptr [[TMP6]], i32 1
+; CHECK-NEXT: [[TMP37:%.*]] = insertelement <4 x ptr> [[TMP36]], ptr [[TMP7]], i32 2
+; CHECK-NEXT: [[TMP38:%.*]] = insertelement <4 x ptr> [[TMP37]], ptr [[TMP8]], i32 3
; CHECK-NEXT: [[TMP9:%.*]] = extractelement <4 x i1> [[TMP4]], i32 0
; CHECK-NEXT: br i1 [[TMP9]], label %[[PRED_LOAD_IF:.*]], label %[[PRED_LOAD_CONTINUE:.*]]
; CHECK: [[PRED_LOAD_IF]]:
>From 0800450605848882025ddee06fd65a49c1f813ad Mon Sep 17 00:00:00 2001
From: Florian Hahn <flo at fhahn.com>
Date: Tue, 10 Jun 2025 10:49:21 +0100
Subject: [PATCH 2/2] !fixup address latest comments, thanks
---
.../Transforms/Vectorize/LoopVectorize.cpp | 2 +-
llvm/lib/Transforms/Vectorize/VPlan.cpp | 7 +-
llvm/lib/Transforms/Vectorize/VPlan.h | 9 +--
.../Transforms/Vectorize/VPlanAnalysis.cpp | 2 +-
.../Transforms/Vectorize/VPlanPatternMatch.h | 16 ++++
.../lib/Transforms/Vectorize/VPlanRecipes.cpp | 81 +++++++++----------
.../Transforms/Vectorize/VPlanTransforms.cpp | 27 +++----
.../Transforms/Vectorize/VPlanTransforms.h | 7 +-
llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp | 28 +++----
9 files changed, 95 insertions(+), 84 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 0ce6ee8a80556..87a5fb2fe741b 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -7294,7 +7294,7 @@ DenseMap<const SCEV *, Value *> LoopVectorizationPlanner::executePlan(
// cost model is complete for better cost estimates.
VPlanTransforms::runPass(VPlanTransforms::unrollByUF, BestVPlan, BestUF,
OrigLoop->getHeader()->getContext());
- VPlanTransforms::runPass(VPlanTransforms::unrollByVF, BestVPlan, BestVF);
+ VPlanTransforms::runPass(VPlanTransforms::replicateByVF, BestVPlan, BestVF);
VPlanTransforms::runPass(VPlanTransforms::materializeBroadcasts, BestVPlan);
VPlanTransforms::optimizeForVFAndUF(BestVPlan, BestVF, BestUF, PSE);
VPlanTransforms::simplifyRecipes(BestVPlan, *Legal->getWidestInductionType());
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.cpp b/llvm/lib/Transforms/Vectorize/VPlan.cpp
index a9d0182800db8..ec641e569731e 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlan.cpp
@@ -263,10 +263,9 @@ Value *VPTransformState::get(const VPValue *Def, const VPLane &Lane) {
// Look through BuildVector to avoid redundant extracts.
// TODO: Remove once replicate regions are unrolled explicitly.
- auto *BV = dyn_cast<VPInstruction>(Def);
- if (Lane.getKind() == VPLane::Kind::First && BV &&
- BV->getOpcode() == VPInstruction::BuildVector) {
- return get(BV->getOperand(Lane.getKnownLane()), true);
+ if (Lane.getKind() == VPLane::Kind::First && match(Def, m_BuildVector())) {
+ auto *BuildVector = cast<VPInstruction>(Def);
+ return get(BuildVector->getOperand(Lane.getKnownLane()), true);
}
assert(hasVectorValue(Def));
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index a4436e6b2532e..454accba7b4c1 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -907,12 +907,11 @@ class VPInstruction : public VPRecipeWithIRFlags,
BranchOnCount,
BranchOnCond,
Broadcast,
- /// Creates a vector containing all operands. The vector element count
- /// matches the number of operands.
- BuildVector,
- /// Creates a struct of vectors containing all operands. The vector element
- /// count matches the number of operands.
+ /// Creates a struct of fixed-width vectors containing all operands. The number of operands
+/// matches the number of fields in the struct.
BuildStructVector,
+ /// Creates a fixed-width vector containing all operands. The number of operands matches the vector element count.
+ BuildVector,
ComputeAnyOfResult,
ComputeFindLastIVResult,
ComputeReductionResult,
diff --git a/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp b/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
index b79badf986f7b..da4a52203db3f 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
@@ -108,8 +108,8 @@ Type *VPTypeAnalysis::inferScalarTypeForRecipe(const VPInstruction *R) {
case VPInstruction::CalculateTripCountMinusVF:
case VPInstruction::CanonicalIVIncrementForPart:
case VPInstruction::AnyOf:
- case VPInstruction::BuildVector:
case VPInstruction::BuildStructVector:
+ case VPInstruction::BuildVector:
return SetResultTyFromOp();
case VPInstruction::FirstActiveLane:
return Type::getIntNTy(Ctx, 64);
diff --git a/llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h b/llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h
index dfd9fc3d4d719..7c427ac18e92b 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h
@@ -221,6 +221,9 @@ struct Recipe_match {
if ((!matchRecipeAndOpcode<RecipeTys>(R) && ...))
return false;
+ auto *VPI = dyn_cast<VPInstruction>(R);
+ if (VPI && VPI->getOpcode() == VPInstruction::BuildVector)
+ return true;
assert(R->getNumOperands() == std::tuple_size<Ops_t>::value &&
"recipe with matched opcode does not have the expected number of "
"operands");
@@ -260,6 +263,10 @@ struct Recipe_match {
}
};
+template <unsigned Opcode, typename... RecipeTys>
+using ZeroOpRecipe_match =
+ Recipe_match<std::tuple<>, Opcode, false, RecipeTys...>;
+
template <typename Op0_t, unsigned Opcode, typename... RecipeTys>
using UnaryRecipe_match =
Recipe_match<std::tuple<Op0_t>, Opcode, false, RecipeTys...>;
@@ -268,6 +275,10 @@ template <typename Op0_t, unsigned Opcode>
using UnaryVPInstruction_match =
UnaryRecipe_match<Op0_t, Opcode, VPInstruction>;
+template <unsigned Opcode>
+using ZeroOpVPInstruction_match =
+ ZeroOpRecipe_match<Opcode, VPInstruction>;
+
template <typename Op0_t, unsigned Opcode>
using AllUnaryRecipe_match =
UnaryRecipe_match<Op0_t, Opcode, VPWidenRecipe, VPReplicateRecipe,
@@ -299,6 +310,11 @@ using AllBinaryRecipe_match =
BinaryRecipe_match<Op0_t, Op1_t, Opcode, Commutative, VPWidenRecipe,
VPReplicateRecipe, VPWidenCastRecipe, VPInstruction>;
+inline ZeroOpVPInstruction_match<VPInstruction::BuildVector>
+m_BuildVector() {
+ return ZeroOpVPInstruction_match<VPInstruction::BuildVector>();
+}
+
template <unsigned Opcode, typename Op0_t>
inline UnaryVPInstruction_match<Op0_t, Opcode>
m_VPInstruction(const Op0_t &Op0) {
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 7e0cdafd2f0a2..2eb03ab640b8c 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -493,12 +493,10 @@ Value *VPInstruction::generate(VPTransformState &State) {
}
case Instruction::ExtractElement: {
assert(State.VF.isVector() && "Only extract elements from vectors");
+ unsigned IdxToExtract = cast<ConstantInt>(getOperand(1)->getLiveInIRValue())
+ ->getZExtValue();
return State.get(getOperand(0),
- VPLane(cast<ConstantInt>(getOperand(1)->getLiveInIRValue())
- ->getZExtValue()));
- Value *Vec = State.get(getOperand(0));
- Value *Idx = State.get(getOperand(1), /*IsScalar=*/true);
- return Builder.CreateExtractElement(Vec, Idx, Name);
+ VPLane(IdxToExtract));
}
case Instruction::Freeze: {
Value *Op = State.get(getOperand(0), vputils::onlyFirstLaneUsed(this));
@@ -607,24 +605,17 @@ Value *VPInstruction::generate(VPTransformState &State) {
return Builder.CreateVectorSplat(
State.VF, State.get(getOperand(0), /*IsScalar*/ true), "broadcast");
}
- case VPInstruction::BuildVector: {
- auto *ScalarTy = State.TypeAnalysis.inferScalarType(getOperand(0));
- Value *Res = PoisonValue::get(
- toVectorizedTy(ScalarTy, ElementCount::getFixed(getNumOperands())));
- for (const auto &[Idx, Op] : enumerate(operands()))
- Res = State.Builder.CreateInsertElement(Res, State.get(Op, true),
- State.Builder.getInt32(Idx));
- return Res;
- }
case VPInstruction::BuildStructVector: {
// For struct types, we need to build a new 'wide' struct type, where each
// element is widened.
- auto *STy =
+ auto *StructTy =
cast<StructType>(State.TypeAnalysis.inferScalarType(getOperand(0)));
+ auto NumOfElements = ElementCount::getFixed(getNumOperands());
Value *Res = PoisonValue::get(
- toVectorizedTy(STy, ElementCount::getFixed(getNumOperands())));
+ toVectorizedTy(StructTy, NumOfElements));
+ assert(NumOfElements.getKnownMinValue() == StructTy->getNumElements() && "number of operands must match number of elements in StructTy");
for (const auto &[Idx, Op] : enumerate(operands())) {
- for (unsigned I = 0, E = STy->getNumElements(); I != E; I++) {
+ for (unsigned I = 0 ; I != NumOfElements .getKnownMinValue(); I++) {
Value *ScalarValue = Builder.CreateExtractValue(State.get(Op, true), I);
Value *VectorValue = Builder.CreateExtractValue(Res, I);
VectorValue =
@@ -634,6 +625,16 @@ Value *VPInstruction::generate(VPTransformState &State) {
}
return Res;
}
+ case VPInstruction::BuildVector: {
+ auto *ScalarTy = State.TypeAnalysis.inferScalarType(getOperand(0));
+ auto NumOfElements = ElementCount::getFixed(getNumOperands());
+ Value *Res = PoisonValue::get(
+ toVectorizedTy(ScalarTy, NumOfElements));
+ for (const auto &[Idx, Op] : enumerate(operands()))
+ Res = State.Builder.CreateInsertElement(Res, State.get(Op, true),
+ State.Builder.getInt32(Idx));
+ return Res;
+ }
case VPInstruction::ReductionStartVector: {
if (State.VF.isScalar())
return State.get(getOperand(0), true);
@@ -933,8 +934,8 @@ bool VPInstruction::opcodeMayReadOrWriteFromMemory() const {
case Instruction::ICmp:
case Instruction::Select:
case VPInstruction::AnyOf:
- case VPInstruction::BuildVector:
case VPInstruction::BuildStructVector:
+ case VPInstruction::BuildVector:
case VPInstruction::CalculateTripCountMinusVF:
case VPInstruction::CanonicalIVIncrementForPart:
case VPInstruction::ExtractLastElement:
@@ -1056,12 +1057,12 @@ void VPInstruction::print(raw_ostream &O, const Twine &Indent,
case VPInstruction::Broadcast:
O << "broadcast";
break;
- case VPInstruction::BuildVector:
- O << "buildvector";
- break;
case VPInstruction::BuildStructVector:
O << "buildstructvector";
break;
+ case VPInstruction::BuildVector:
+ O << "buildvector";
+ break;
case VPInstruction::ExtractLastElement:
O << "extract-last-element";
break;
@@ -2797,30 +2798,28 @@ static void scalarizeInstruction(const Instruction *Instr,
void VPReplicateRecipe::execute(VPTransformState &State) {
Instruction *UI = getUnderlyingInstr();
- if (State.Lane) { // Generate a single instance.
- assert((State.VF.isScalar() || !isSingleScalar()) &&
- "uniform recipe shouldn't be predicated");
- assert(!State.VF.isScalable() && "Can't scalarize a scalable vector");
- scalarizeInstruction(UI, this, *State.Lane, State);
- // Insert scalar instance packing it into a vector.
- if (State.VF.isVector() && shouldPack()) {
- // If we're constructing lane 0, initialize to start from poison.
- if (State.Lane->isFirstLane()) {
- assert(!State.VF.isScalable() && "VF is assumed to be non scalable.");
- Value *Poison =
- PoisonValue::get(VectorType::get(UI->getType(), State.VF));
- State.set(this, Poison);
- }
- State.packScalarIntoVectorizedValue(this, *State.Lane);
- }
- return;
- }
- if (IsSingleScalar) {
- // Uniform within VL means we need to generate lane 0.
+ if (!State.Lane) {
+ assert(IsSingleScalar && "VPReplicateRecipes outside replicate regions must be unrolled");
scalarizeInstruction(UI, this, VPLane(0), State);
return;
}
+
+ assert((State.VF.isScalar() || !isSingleScalar()) &&
+ "uniform recipe shouldn't be predicated");
+ assert(!State.VF.isScalable() && "Can't scalarize a scalable vector");
+ scalarizeInstruction(UI, this, *State.Lane, State);
+ // Insert scalar instance packing it into a vector.
+ if (State.VF.isVector() && shouldPack()) {
+ // If we're constructing lane 0, initialize to start from poison.
+ if (State.Lane->isFirstLane()) {
+ assert(!State.VF.isScalable() && "VF is assumed to be non scalable.");
+ Value *Poison =
+ PoisonValue::get(VectorType::get(UI->getType(), State.VF));
+ State.set(this, Poison);
+ }
+ State.packScalarIntoVectorizedValue(this, *State.Lane);
+ }
}
bool VPReplicateRecipe::shouldPack() const {
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index b9d029dd2eea5..df65874f605e5 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1140,21 +1140,20 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
return;
}
- // Look through Extract(Last|Penultimate)Element (BuildVector ....).
- if (match(&R,
- m_VPInstruction<VPInstruction::ExtractLastElement>(m_VPValue(A))) ||
- match(&R, m_VPInstruction<VPInstruction::ExtractPenultimateElement>(
- m_VPValue(A)))) {
- unsigned Offset = cast<VPInstruction>(&R)->getOpcode() ==
- VPInstruction::ExtractLastElement
- ? 1
- : 2;
- auto *BV = dyn_cast<VPInstruction>(A);
- if (BV && BV->getOpcode() == VPInstruction::BuildVector) {
- Def->replaceAllUsesWith(BV->getOperand(BV->getNumOperands() - Offset));
- return;
- }
+ // Look through ExtractLastElement (BuildVector ....).
+ if (match(&R, m_VPInstruction<VPInstruction::ExtractLastElement>(
+ m_BuildVector()))) {
+ auto *BuildVector = cast<VPInstruction>(R.getOperand(0));
+ Def->replaceAllUsesWith(BuildVector->getOperand(BuildVector->getNumOperands() - 1));
+ return;
}
+ // Look through ExtractPenultimateElement (BuildVector ....).
+ if (match(&R, m_VPInstruction<VPInstruction::ExtractPenultimateElement>(
+ m_BuildVector()))) {
+ auto *BuildVector = cast<VPInstruction>(R.getOperand(0));
+ Def->replaceAllUsesWith(BuildVector->getOperand(BuildVector->getNumOperands() - 2));
+ return;
+}
// Some simplifications can only be applied after unrolling. Perform them
// below.
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
index f45b7a7969d04..c848c2057c6d1 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
@@ -99,9 +99,10 @@ struct VPlanTransforms {
/// Explicitly unroll \p Plan by \p UF.
static void unrollByUF(VPlan &Plan, unsigned UF, LLVMContext &Ctx);
- /// Explicitly unroll VPReplicateRecipes outside of replicate regions by \p
- /// VF.
- static void unrollByVF(VPlan &Plan, ElementCount VF);
+ /// Replace replicating VPReplicateRecipes outside replicate regions in \p
+ /// Plan with \p VF single-scalar recipes.
+ /// TODO: Also unroll VPReplicateRegions by VF.
+ static void replicateByVF(VPlan &Plan, ElementCount VF);
/// Optimize \p Plan based on \p BestVF and \p BestUF. This may restrict the
/// resulting plan to \p BestVF and \p BestUF.
diff --git a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
index 1a86926937f4e..9b04761149f00 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
@@ -447,7 +447,7 @@ void VPlanTransforms::unrollByUF(VPlan &Plan, unsigned UF, LLVMContext &Ctx) {
VPlanTransforms::removeDeadRecipes(Plan);
}
-/// Create a single-scalar clone of RepR for lane \p Lane.
+/// Create a single-scalar clone of \p RepR for lane \p Lane.
static VPReplicateRecipe *cloneForLane(VPlan &Plan, VPBuilder &Builder,
Type *IdxTy, VPReplicateRecipe *RepR,
VPLane Lane) {
@@ -458,20 +458,18 @@ static VPReplicateRecipe *cloneForLane(VPlan &Plan, VPBuilder &Builder,
NewOps.push_back(Op);
continue;
}
- VPValue *Ext;
if (Lane.getKind() == VPLane::Kind::ScalableLast) {
- Ext = Builder.createNaryOp(VPInstruction::ExtractLastElement, {Op});
- } else {
- // Look through buildvector to avoid unnecessary extracts.
- auto *BV = dyn_cast<VPInstruction>(Op);
- if (BV && BV->getOpcode() == VPInstruction::BuildVector) {
- NewOps.push_back(BV->getOperand(Lane.getKnownLane()));
- continue;
- }
- VPValue *Idx =
- Plan.getOrAddLiveIn(ConstantInt::get(IdxTy, Lane.getKnownLane()));
- Ext = Builder.createNaryOp(Instruction::ExtractElement, {Op, Idx});
+ NewOps.push_back(Builder.createNaryOp(VPInstruction::ExtractLastElement, {Op}));
+ continue;
+ }
+ // Look through buildvector to avoid unnecessary extracts.
+ if (match(Op, m_BuildVector())) {
+ NewOps.push_back(cast<VPInstruction>(Op)->getOperand(Lane.getKnownLane()));
+ continue;
}
+ VPValue *Idx =
+ Plan.getOrAddLiveIn(ConstantInt::get(IdxTy, Lane.getKnownLane()));
+ VPValue *Ext = Builder.createNaryOp(Instruction::ExtractElement, {Op, Idx});
NewOps.push_back(Ext);
}
@@ -482,7 +480,7 @@ static VPReplicateRecipe *cloneForLane(VPlan &Plan, VPBuilder &Builder,
return New;
}
-void VPlanTransforms::unrollByVF(VPlan &Plan, ElementCount VF) {
+void VPlanTransforms::replicateByVF(VPlan &Plan, ElementCount VF) {
Type *IdxTy = IntegerType::get(
Plan.getScalarHeader()->getIRBasicBlock()->getContext(), 32);
for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
@@ -494,7 +492,7 @@ void VPlanTransforms::unrollByVF(VPlan &Plan, ElementCount VF) {
VPBuilder Builder(RepR);
SmallVector<VPValue *> LaneDefs;
- // Stores to invariant addresses only need to store the last lane.
+ // Stores to invariant addresses need to store the last lane only.
if (isa<StoreInst>(RepR->getUnderlyingInstr()) &&
vputils::isSingleScalar(RepR->getOperand(1))) {
cloneForLane(Plan, Builder, IdxTy, RepR, VPLane::getLastLaneForVF(VF));
More information about the llvm-commits
mailing list