[llvm] [SandboxVec][BottomUpVec] Clean up dead address instrs (PR #122536)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Jan 10 13:50:50 PST 2025
https://github.com/vporpo created https://github.com/llvm/llvm-project/pull/122536
When we vectorize loads or stores we only keep the address of the first lane. The rest may become dead. This patch adds the address operands of vectorized loads or stores to the dead candidates set, such that they get erased if dead.
>From d74dcff66d4a835db46d1df5a16ab1b5623c2fe5 Mon Sep 17 00:00:00 2001
From: Vasileios Porpodas <vporpodas at google.com>
Date: Thu, 14 Nov 2024 10:06:11 -0800
Subject: [PATCH] [SandboxVec][BottomUpVec] Clean up dead address instrs
When we vectorize loads or stores we only keep the address of the first lane.
The rest may become dead. This patch adds the address operands of vectorized
loads or stores to the dead candidates set, such that they get erased if dead.
---
.../SandboxVectorizer/Passes/BottomUpVec.h | 3 +-
.../SandboxVectorizer/Passes/BottomUpVec.cpp | 43 +++++++++++++++----
.../SandboxVectorizer/bottomup_basic.ll | 9 ----
.../SandboxVectorizer/bottomup_seed_slice.ll | 1 -
.../bottomup_seed_slice_pow2.ll | 3 --
5 files changed, 37 insertions(+), 22 deletions(-)
diff --git a/llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.h b/llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.h
index bd45634814b076..1a53ca6e06f5fd 100644
--- a/llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.h
+++ b/llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.h
@@ -25,7 +25,7 @@ namespace llvm::sandboxir {
class BottomUpVec final : public FunctionPass {
bool Change = false;
std::unique_ptr<LegalityAnalysis> Legality;
- SmallVector<Instruction *> DeadInstrCandidates;
+ DenseSet<Instruction *> DeadInstrCandidates;
/// Creates and returns a vector instruction that replaces the instructions in
/// \p Bndl. \p Operands are the already vectorized operands.
@@ -35,6 +35,7 @@ class BottomUpVec final : public FunctionPass {
void tryEraseDeadInstrs();
/// Packs all elements of \p ToPack into a vector and returns that vector.
Value *createPack(ArrayRef<Value *> ToPack);
+ void collectPotentiallyDeadInstrs(ArrayRef<Value *> Bndl);
/// Recursively try to vectorize \p Bndl and its operands.
Value *vectorizeRec(ArrayRef<Value *> Bndl, unsigned Depth);
/// Entry point for vectorization starting from \p Seeds.
diff --git a/llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp b/llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp
index 18e072c17d202b..6957a1743f72e8 100644
--- a/llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp
+++ b/llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp
@@ -157,9 +157,11 @@ Value *BottomUpVec::createVectorInstr(ArrayRef<Value *> Bndl,
void BottomUpVec::tryEraseDeadInstrs() {
// Visiting the dead instructions bottom-to-top.
- sort(DeadInstrCandidates,
+ SmallVector<Instruction *> SortedDeadInstrCandidates(
+ DeadInstrCandidates.begin(), DeadInstrCandidates.end());
+ sort(SortedDeadInstrCandidates,
[](Instruction *I1, Instruction *I2) { return I1->comesBefore(I2); });
- for (Instruction *I : reverse(DeadInstrCandidates)) {
+ for (Instruction *I : reverse(SortedDeadInstrCandidates)) {
if (I->hasNUses(0))
I->eraseFromParent();
}
@@ -218,6 +220,31 @@ Value *BottomUpVec::createPack(ArrayRef<Value *> ToPack) {
return LastInsert;
}
+void BottomUpVec::collectPotentiallyDeadInstrs(ArrayRef<Value *> Bndl) {
+ for (Value *V : Bndl)
+ DeadInstrCandidates.insert(cast<Instruction>(V));
+ // Also collect the GEPs of vectorized loads and stores.
+ auto Opcode = cast<Instruction>(Bndl[0])->getOpcode();
+ switch (Opcode) {
+ case Instruction::Opcode::Load: {
+ for (Value *V : drop_begin(Bndl))
+ if (auto *Ptr =
+ dyn_cast<Instruction>(cast<LoadInst>(V)->getPointerOperand()))
+ DeadInstrCandidates.insert(Ptr);
+ break;
+ }
+ case Instruction::Opcode::Store: {
+ for (Value *V : drop_begin(Bndl))
+ if (auto *Ptr =
+ dyn_cast<Instruction>(cast<StoreInst>(V)->getPointerOperand()))
+ DeadInstrCandidates.insert(Ptr);
+ break;
+ }
+ default:
+ break;
+ }
+}
+
Value *BottomUpVec::vectorizeRec(ArrayRef<Value *> Bndl, unsigned Depth) {
Value *NewVec = nullptr;
const auto &LegalityRes = Legality->canVectorize(Bndl);
@@ -225,7 +252,8 @@ Value *BottomUpVec::vectorizeRec(ArrayRef<Value *> Bndl, unsigned Depth) {
case LegalityResultID::Widen: {
auto *I = cast<Instruction>(Bndl[0]);
SmallVector<Value *, 2> VecOperands;
- switch (I->getOpcode()) {
+ auto Opcode = I->getOpcode();
+ switch (Opcode) {
case Instruction::Opcode::Load:
// Don't recurse towards the pointer operand.
VecOperands.push_back(cast<LoadInst>(I)->getPointerOperand());
@@ -247,11 +275,10 @@ Value *BottomUpVec::vectorizeRec(ArrayRef<Value *> Bndl, unsigned Depth) {
}
NewVec = createVectorInstr(Bndl, VecOperands);
- // Collect the original scalar instructions as they may be dead.
- if (NewVec != nullptr) {
- for (Value *V : Bndl)
- DeadInstrCandidates.push_back(cast<Instruction>(V));
- }
+ // Collect any potentially dead scalar instructions, including the original
+ // scalars and pointer operands of loads/stores.
+ if (NewVec != nullptr)
+ collectPotentiallyDeadInstrs(Bndl);
break;
}
case LegalityResultID::Pack: {
diff --git a/llvm/test/Transforms/SandboxVectorizer/bottomup_basic.ll b/llvm/test/Transforms/SandboxVectorizer/bottomup_basic.ll
index 785d1f4ef666fc..d34c8f88e4b3c6 100644
--- a/llvm/test/Transforms/SandboxVectorizer/bottomup_basic.ll
+++ b/llvm/test/Transforms/SandboxVectorizer/bottomup_basic.ll
@@ -5,7 +5,6 @@ define void @store_load(ptr %ptr) {
; CHECK-LABEL: define void @store_load(
; CHECK-SAME: ptr [[PTR:%.*]]) {
; CHECK-NEXT: [[PTR0:%.*]] = getelementptr float, ptr [[PTR]], i32 0
-; CHECK-NEXT: [[PTR1:%.*]] = getelementptr float, ptr [[PTR]], i32 1
; CHECK-NEXT: [[VECL:%.*]] = load <2 x float>, ptr [[PTR0]], align 4
; CHECK-NEXT: store <2 x float> [[VECL]], ptr [[PTR0]], align 4
; CHECK-NEXT: ret void
@@ -24,9 +23,7 @@ define void @store_fpext_load(ptr %ptr) {
; CHECK-LABEL: define void @store_fpext_load(
; CHECK-SAME: ptr [[PTR:%.*]]) {
; CHECK-NEXT: [[PTR0:%.*]] = getelementptr float, ptr [[PTR]], i32 0
-; CHECK-NEXT: [[PTR1:%.*]] = getelementptr float, ptr [[PTR]], i32 1
; CHECK-NEXT: [[PTRD0:%.*]] = getelementptr double, ptr [[PTR]], i32 0
-; CHECK-NEXT: [[PTRD1:%.*]] = getelementptr double, ptr [[PTR]], i32 1
; CHECK-NEXT: [[VECL:%.*]] = load <2 x float>, ptr [[PTR0]], align 4
; CHECK-NEXT: [[VCAST:%.*]] = fpext <2 x float> [[VECL]] to <2 x double>
; CHECK-NEXT: store <2 x double> [[VCAST]], ptr [[PTRD0]], align 8
@@ -49,9 +46,7 @@ define void @store_fcmp_zext_load(ptr %ptr) {
; CHECK-LABEL: define void @store_fcmp_zext_load(
; CHECK-SAME: ptr [[PTR:%.*]]) {
; CHECK-NEXT: [[PTR0:%.*]] = getelementptr float, ptr [[PTR]], i32 0
-; CHECK-NEXT: [[PTR1:%.*]] = getelementptr float, ptr [[PTR]], i32 1
; CHECK-NEXT: [[PTRB0:%.*]] = getelementptr i32, ptr [[PTR]], i32 0
-; CHECK-NEXT: [[PTRB1:%.*]] = getelementptr i32, ptr [[PTR]], i32 1
; CHECK-NEXT: [[VECL1:%.*]] = load <2 x float>, ptr [[PTR0]], align 4
; CHECK-NEXT: [[VECL:%.*]] = load <2 x float>, ptr [[PTR0]], align 4
; CHECK-NEXT: [[VCMP:%.*]] = fcmp ogt <2 x float> [[VECL]], [[VECL1]]
@@ -80,7 +75,6 @@ define void @store_fadd_load(ptr %ptr) {
; CHECK-LABEL: define void @store_fadd_load(
; CHECK-SAME: ptr [[PTR:%.*]]) {
; CHECK-NEXT: [[PTR0:%.*]] = getelementptr float, ptr [[PTR]], i32 0
-; CHECK-NEXT: [[PTR1:%.*]] = getelementptr float, ptr [[PTR]], i32 1
; CHECK-NEXT: [[VECL:%.*]] = load <2 x float>, ptr [[PTR0]], align 4
; CHECK-NEXT: [[VECL1:%.*]] = load <2 x float>, ptr [[PTR0]], align 4
; CHECK-NEXT: [[VEC:%.*]] = fadd <2 x float> [[VECL]], [[VECL1]]
@@ -104,7 +98,6 @@ define void @store_fneg_load(ptr %ptr) {
; CHECK-LABEL: define void @store_fneg_load(
; CHECK-SAME: ptr [[PTR:%.*]]) {
; CHECK-NEXT: [[PTR0:%.*]] = getelementptr float, ptr [[PTR]], i32 0
-; CHECK-NEXT: [[PTR1:%.*]] = getelementptr float, ptr [[PTR]], i32 1
; CHECK-NEXT: [[VECL:%.*]] = load <2 x float>, ptr [[PTR0]], align 4
; CHECK-NEXT: [[VEC:%.*]] = fneg <2 x float> [[VECL]]
; CHECK-NEXT: store <2 x float> [[VEC]], ptr [[PTR0]], align 4
@@ -147,7 +140,6 @@ define void @pack_scalars(ptr %ptr, ptr %ptr2) {
; CHECK-LABEL: define void @pack_scalars(
; CHECK-SAME: ptr [[PTR:%.*]], ptr [[PTR2:%.*]]) {
; CHECK-NEXT: [[PTR0:%.*]] = getelementptr float, ptr [[PTR]], i32 0
-; CHECK-NEXT: [[PTR1:%.*]] = getelementptr float, ptr [[PTR]], i32 1
; CHECK-NEXT: [[LD0:%.*]] = load float, ptr [[PTR0]], align 4
; CHECK-NEXT: [[LD1:%.*]] = load float, ptr [[PTR2]], align 4
; CHECK-NEXT: [[PACK:%.*]] = insertelement <2 x float> poison, float [[LD0]], i32 0
@@ -191,7 +183,6 @@ define void @pack_vectors(ptr %ptr, ptr %ptr2) {
; CHECK-LABEL: define void @pack_vectors(
; CHECK-SAME: ptr [[PTR:%.*]], ptr [[PTR2:%.*]]) {
; CHECK-NEXT: [[PTR0:%.*]] = getelementptr <2 x float>, ptr [[PTR]], i32 0
-; CHECK-NEXT: [[PTR1:%.*]] = getelementptr float, ptr [[PTR]], i32 2
; CHECK-NEXT: [[LD0:%.*]] = load <2 x float>, ptr [[PTR0]], align 8
; CHECK-NEXT: [[LD1:%.*]] = load float, ptr [[PTR2]], align 4
; CHECK-NEXT: [[VPACK:%.*]] = extractelement <2 x float> [[LD0]], i32 0
diff --git a/llvm/test/Transforms/SandboxVectorizer/bottomup_seed_slice.ll b/llvm/test/Transforms/SandboxVectorizer/bottomup_seed_slice.ll
index 46cda3c80aaa35..8459c3addaa83f 100644
--- a/llvm/test/Transforms/SandboxVectorizer/bottomup_seed_slice.ll
+++ b/llvm/test/Transforms/SandboxVectorizer/bottomup_seed_slice.ll
@@ -7,7 +7,6 @@ define void @slice_seeds(ptr %ptr, float %val) {
; CHECK-LABEL: define void @slice_seeds(
; CHECK-SAME: ptr [[PTR:%.*]], float [[VAL:%.*]]) {
; CHECK-NEXT: [[PTR0:%.*]] = getelementptr float, ptr [[PTR]], i32 0
-; CHECK-NEXT: [[PTR1:%.*]] = getelementptr float, ptr [[PTR]], i32 1
; CHECK-NEXT: [[PTR2:%.*]] = getelementptr float, ptr [[PTR]], i32 2
; CHECK-NEXT: [[LD2:%.*]] = load float, ptr [[PTR2]], align 4
; CHECK-NEXT: store float [[LD2]], ptr [[PTR2]], align 4
diff --git a/llvm/test/Transforms/SandboxVectorizer/bottomup_seed_slice_pow2.ll b/llvm/test/Transforms/SandboxVectorizer/bottomup_seed_slice_pow2.ll
index 22119c4491b929..e186d5fa86e4a7 100644
--- a/llvm/test/Transforms/SandboxVectorizer/bottomup_seed_slice_pow2.ll
+++ b/llvm/test/Transforms/SandboxVectorizer/bottomup_seed_slice_pow2.ll
@@ -6,7 +6,6 @@ define void @pow2(ptr %ptr, float %val) {
; POW2-LABEL: define void @pow2(
; POW2-SAME: ptr [[PTR:%.*]], float [[VAL:%.*]]) {
; POW2-NEXT: [[PTR0:%.*]] = getelementptr float, ptr [[PTR]], i32 0
-; POW2-NEXT: [[PTR1:%.*]] = getelementptr float, ptr [[PTR]], i32 1
; POW2-NEXT: [[PTR2:%.*]] = getelementptr float, ptr [[PTR]], i32 2
; POW2-NEXT: [[VECL:%.*]] = load <2 x float>, ptr [[PTR0]], align 4
; POW2-NEXT: [[LD2:%.*]] = load float, ptr [[PTR2]], align 4
@@ -17,8 +16,6 @@ define void @pow2(ptr %ptr, float %val) {
; NON-POW2-LABEL: define void @pow2(
; NON-POW2-SAME: ptr [[PTR:%.*]], float [[VAL:%.*]]) {
; NON-POW2-NEXT: [[PTR0:%.*]] = getelementptr float, ptr [[PTR]], i32 0
-; NON-POW2-NEXT: [[PTR1:%.*]] = getelementptr float, ptr [[PTR]], i32 1
-; NON-POW2-NEXT: [[PTR2:%.*]] = getelementptr float, ptr [[PTR]], i32 2
; NON-POW2-NEXT: [[PACK2:%.*]] = load <3 x float>, ptr [[PTR0]], align 4
; NON-POW2-NEXT: store <3 x float> [[PACK2]], ptr [[PTR0]], align 4
; NON-POW2-NEXT: ret void
More information about the llvm-commits
mailing list