[llvm] 5d5e706 - [VPlan] Restrict hoisting of broadcast operations using VPDominatorTree (#117138)
Mel Chen via llvm-commits
llvm-commits at lists.llvm.org
Thu Mar 13 07:16:58 PDT 2025
Author: Mel Chen
Date: 2025-03-13T07:16:04-07:00
New Revision: 5d5e706691c6d811820598a73694d3430c1d11b0
URL: https://github.com/llvm/llvm-project/commit/5d5e706691c6d811820598a73694d3430c1d11b0
DIFF: https://github.com/llvm/llvm-project/commit/5d5e706691c6d811820598a73694d3430c1d11b0.diff
LOG: [VPlan] Restrict hoisting of broadcast operations using VPDominatorTree (#117138)
This patch restricts broadcast operations from being hoisted to the vector
preheader unless the basic block that defines the broadcasted value properly
dominates the vector preheader.
This prevents potential use-before-definition issues when the broadcasted
value is defined within the plan. VPDominatorTree is used to confirm this
restriction while still allowing safe hoisting for broadcasted values defined
outside the plan.
Issue https://github.com/llvm/llvm-project/issues/117139
Added:
Modified:
llvm/lib/Transforms/Vectorize/VPlan.cpp
llvm/lib/Transforms/Vectorize/VPlanDominatorTree.h
llvm/lib/Transforms/Vectorize/VPlanHelpers.h
llvm/test/Transforms/LoopVectorize/AArch64/extractvalue-no-scalarization-required.ll
llvm/test/Transforms/LoopVectorize/AArch64/sve-widen-extractvalue.ll
llvm/test/Transforms/LoopVectorize/RISCV/truncate-to-minimal-bitwidth-cost.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.cpp b/llvm/lib/Transforms/Vectorize/VPlan.cpp
index 944a11b96325d..e595347d62bf5 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlan.cpp
@@ -19,6 +19,7 @@
#include "VPlan.h"
#include "LoopVectorizationPlanner.h"
#include "VPlanCFG.h"
+#include "VPlanDominatorTree.h"
#include "VPlanHelpers.h"
#include "VPlanPatternMatch.h"
#include "VPlanTransforms.h"
@@ -221,7 +222,7 @@ VPTransformState::VPTransformState(const TargetTransformInfo *TTI,
Loop *CurrentParentLoop, Type *CanonicalIVTy)
: TTI(TTI), VF(VF), CFG(DT), LI(LI), Builder(Builder), ILV(ILV), Plan(Plan),
CurrentParentLoop(CurrentParentLoop), LVer(nullptr),
- TypeAnalysis(CanonicalIVTy) {}
+ TypeAnalysis(CanonicalIVTy), VPDT(*Plan) {}
Value *VPTransformState::get(const VPValue *Def, const VPLane &Lane) {
if (Def->isLiveIn())
@@ -264,7 +265,11 @@ Value *VPTransformState::get(const VPValue *Def, bool NeedsScalar) {
return Data.VPV2Vector[Def];
auto GetBroadcastInstrs = [this, Def](Value *V) {
- bool SafeToHoist = Def->isDefinedOutsideLoopRegions();
+ bool SafeToHoist =
+ !Def->hasDefiningRecipe() ||
+ VPDT.properlyDominates(Def->getDefiningRecipe()->getParent(),
+ Plan->getVectorPreheader());
+
if (VF.isScalar())
return V;
// Place the code for broadcasting invariant variables in the new preheader.
@@ -929,6 +934,10 @@ void VPlan::execute(VPTransformState *State) {
State->CFG.PrevVPBB = nullptr;
State->CFG.ExitBB = State->CFG.PrevBB->getSingleSuccessor();
+ // Update VPDominatorTree since VPBasicBlock may be removed after State was
+ // constructed.
+ State->VPDT.recalculate(*this);
+
// Disconnect VectorPreHeader from ExitBB in both the CFG and DT.
BasicBlock *VectorPreHeader = State->CFG.PrevBB;
cast<BranchInst>(VectorPreHeader->getTerminator())->setSuccessor(0, nullptr);
diff --git a/llvm/lib/Transforms/Vectorize/VPlanDominatorTree.h b/llvm/lib/Transforms/Vectorize/VPlanDominatorTree.h
index 787cd940c4e26..995c6b8b2c2fb 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanDominatorTree.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanDominatorTree.h
@@ -39,8 +39,10 @@ class VPDominatorTree : public DominatorTreeBase<VPBlockBase, false> {
public:
VPDominatorTree() = default;
+ explicit VPDominatorTree(VPlan &Plan) { recalculate(Plan); }
/// Returns true if \p A properly dominates \p B.
+ using Base::properlyDominates;
bool properlyDominates(const VPRecipeBase *A, const VPRecipeBase *B);
};
diff --git a/llvm/lib/Transforms/Vectorize/VPlanHelpers.h b/llvm/lib/Transforms/Vectorize/VPlanHelpers.h
index 8bdbf556efbb3..d53b72bb2258a 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanHelpers.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanHelpers.h
@@ -16,6 +16,7 @@
#define LLVM_TRANSFORMS_VECTORIZE_VPLANHELPERS_H
#include "VPlanAnalysis.h"
+#include "VPlanDominatorTree.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
@@ -360,6 +361,9 @@ struct VPTransformState {
/// VPlan-based type analysis.
VPTypeAnalysis TypeAnalysis;
+
+ /// VPlan-based dominator tree.
+ VPDominatorTree VPDT;
};
/// Struct to hold various analysis needed for cost computations.
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/extractvalue-no-scalarization-required.ll b/llvm/test/Transforms/LoopVectorize/AArch64/extractvalue-no-scalarization-required.ll
index f969be80c3c00..ab21794baa924 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/extractvalue-no-scalarization-required.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/extractvalue-no-scalarization-required.ll
@@ -23,9 +23,9 @@
; Check that the extractvalue operands are actually free in vector code.
; FORCED: [[E1:%.+]] = extractvalue { i64, i64 } %sv, 0
-; FORCED-NEXT: [[E2:%.+]] = extractvalue { i64, i64 } %sv, 1
; FORCED-NEXT: %broadcast.splatinsert = insertelement <2 x i64> poison, i64 [[E1]], i64 0
; FORCED-NEXT: %broadcast.splat = shufflevector <2 x i64> %broadcast.splatinsert, <2 x i64> poison, <2 x i32> zeroinitializer
+; FORCED-NEXT: [[E2:%.+]] = extractvalue { i64, i64 } %sv, 1
; FORCED-NEXT: %broadcast.splatinsert1 = insertelement <2 x i64> poison, i64 [[E2]], i64 0
; FORCED-NEXT: %broadcast.splat2 = shufflevector <2 x i64> %broadcast.splatinsert1, <2 x i64> poison, <2 x i32> zeroinitializer
; FORCED-NEXT: [[ADD:%.+]] = add <2 x i64> %broadcast.splat, %broadcast.splat2
@@ -75,9 +75,9 @@ declare float @powf(float, float) readnone nounwind
; FORCED-LABEL: define void @test_getVectorCallCost
; FORCED: [[E1:%.+]] = extractvalue { float, float } %sv, 0
-; FORCED-NEXT: [[E2:%.+]] = extractvalue { float, float } %sv, 1
; FORCED-NEXT: %broadcast.splatinsert = insertelement <2 x float> poison, float [[E1]], i64 0
; FORCED-NEXT: %broadcast.splat = shufflevector <2 x float> %broadcast.splatinsert, <2 x float> poison, <2 x i32> zeroinitializer
+; FORCED-NEXT: [[E2:%.+]] = extractvalue { float, float } %sv, 1
; FORCED-NEXT: %broadcast.splatinsert1 = insertelement <2 x float> poison, float [[E2]], i64 0
; FORCED-NEXT: %broadcast.splat2 = shufflevector <2 x float> %broadcast.splatinsert1, <2 x float> poison, <2 x i32> zeroinitializer
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/sve-widen-extractvalue.ll b/llvm/test/Transforms/LoopVectorize/AArch64/sve-widen-extractvalue.ll
index 91dd6e475ec47..473fabfc9fecc 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/sve-widen-extractvalue.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/sve-widen-extractvalue.ll
@@ -5,9 +5,9 @@ target triple = "aarch64-unknown-linux-gnu"
define void @widen_extractvalue(ptr %dst, {i64, i64} %sv) #0 {
; CHECK-LABEL: @widen_extractvalue(
; CHECK: [[EXTRACT0:%.*]] = extractvalue { i64, i64 } [[SV:%.*]], 0
-; CHECK-NEXT: [[EXTRACT1:%.*]] = extractvalue { i64, i64 } [[SV]], 1
; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[EXTRACT0]], i64 0
; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 2 x i64> [[DOTSPLATINSERT]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
+; CHECK-NEXT: [[EXTRACT1:%.*]] = extractvalue { i64, i64 } [[SV]], 1
; CHECK-NEXT: [[DOTSPLATINSERT1:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[EXTRACT1]], i64 0
; CHECK-NEXT: [[DOTSPLAT2:%.*]] = shufflevector <vscale x 2 x i64> [[DOTSPLATINSERT1]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
; CHECK: [[ADD:%.*]] = add <vscale x 2 x i64> [[DOTSPLAT]], [[DOTSPLAT2]]
diff --git a/llvm/test/Transforms/LoopVectorize/RISCV/truncate-to-minimal-bitwidth-cost.ll b/llvm/test/Transforms/LoopVectorize/RISCV/truncate-to-minimal-bitwidth-cost.ll
index a5728daf63d4c..94c0036e4c8b6 100644
--- a/llvm/test/Transforms/LoopVectorize/RISCV/truncate-to-minimal-bitwidth-cost.ll
+++ b/llvm/test/Transforms/LoopVectorize/RISCV/truncate-to-minimal-bitwidth-cost.ll
@@ -276,10 +276,10 @@ define void @icmp_only_first_op_truncated(ptr noalias %dst, i32 %x, i64 %N, i64
; CHECK-NEXT: [[TMP8:%.*]] = icmp eq <vscale x 2 x i32> [[TMP7]], [[BROADCAST_SPLAT2]]
; CHECK-NEXT: [[TMP9:%.*]] = zext i32 [[X]] to i64
; CHECK-NEXT: [[TMP10:%.*]] = getelementptr double, ptr [[SRC]], i64 [[TMP9]]
-; CHECK-NEXT: [[BROADCAST_SPLATINSERT5:%.*]] = insertelement <vscale x 2 x ptr> poison, ptr [[DST]], i64 0
-; CHECK-NEXT: [[BROADCAST_SPLAT6:%.*]] = shufflevector <vscale x 2 x ptr> [[BROADCAST_SPLATINSERT5]], <vscale x 2 x ptr> poison, <vscale x 2 x i32> zeroinitializer
; CHECK-NEXT: [[BROADCAST_SPLATINSERT6:%.*]] = insertelement <vscale x 2 x ptr> poison, ptr [[TMP10]], i64 0
; CHECK-NEXT: [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 2 x ptr> [[BROADCAST_SPLATINSERT6]], <vscale x 2 x ptr> poison, <vscale x 2 x i32> zeroinitializer
+; CHECK-NEXT: [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 2 x ptr> poison, ptr [[DST]], i64 0
+; CHECK-NEXT: [[BROADCAST_SPLAT6:%.*]] = shufflevector <vscale x 2 x ptr> [[BROADCAST_SPLATINSERT3]], <vscale x 2 x ptr> poison, <vscale x 2 x i32> zeroinitializer
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
; CHECK: [[VECTOR_BODY]]:
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
More information about the llvm-commits
mailing list