[llvm] 79314c6 - [SLP]Fix PR90224: check that users of gep are all vectorized.
Alexey Bataev via llvm-commits
llvm-commits at lists.llvm.org
Fri Apr 26 11:51:27 PDT 2024
Author: Alexey Bataev
Date: 2024-04-26T11:49:12-07:00
New Revision: 79314c64d0f88fa4ccc1709dee373fdfb9a6ca49
URL: https://github.com/llvm/llvm-project/commit/79314c64d0f88fa4ccc1709dee373fdfb9a6ca49
DIFF: https://github.com/llvm/llvm-project/commit/79314c64d0f88fa4ccc1709dee373fdfb9a6ca49.diff
LOG: [SLP]Fix PR90224: check that users of gep are all vectorized.
Before deleting extractelement instruction for vectorized GEP with
external users, need to check that all users vectorized before deleting
this extractelement.
Added:
llvm/test/Transforms/SLPVectorizer/X86/gep-with-extractelement-many-users.ll
Modified:
llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index fbece8c0109c38..f62270fe62ebea 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -8499,6 +8499,12 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
const TreeEntry *VE = R.getTreeEntry(V);
if (!CheckedExtracts.insert(V).second ||
!R.areAllUsersVectorized(cast<Instruction>(V), &VectorizedVals) ||
+ any_of(EE->users(),
+ [&](User *U) {
+ return isa<GetElementPtrInst>(U) &&
+ !R.areAllUsersVectorized(cast<Instruction>(U),
+ &VectorizedVals);
+ }) ||
(VE && VE != E))
continue;
std::optional<unsigned> EEIdx = getExtractIndex(EE);
@@ -11397,6 +11403,8 @@ class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis {
any_of(EI->users(), [&](User *U) {
const TreeEntry *UTE = R.getTreeEntry(U);
return !UTE || R.MultiNodeScalars.contains(U) ||
+ (isa<GetElementPtrInst>(U) &&
+ !R.areAllUsersVectorized(cast<Instruction>(U))) ||
count_if(R.VectorizableTree,
[&](const std::unique_ptr<TreeEntry> &TE) {
return any_of(TE->UserTreeIndices,
@@ -13329,8 +13337,11 @@ Value *BoUpSLP::vectorizeTree(
// Leave the GEPs as is, they are free in most cases and better to
// keep them as GEPs.
auto *CloneGEP = GEP->clone();
- CloneGEP->insertBefore(*Builder.GetInsertBlock(),
- Builder.GetInsertPoint());
+ if (isa<Instruction>(Vec))
+ CloneGEP->insertBefore(*Builder.GetInsertBlock(),
+ Builder.GetInsertPoint());
+ else
+ CloneGEP->insertBefore(GEP);
if (GEP->hasName())
CloneGEP->takeName(GEP);
Ex = CloneGEP;
diff --git a/llvm/test/Transforms/SLPVectorizer/X86/gep-with-extractelement-many-users.ll b/llvm/test/Transforms/SLPVectorizer/X86/gep-with-extractelement-many-users.ll
new file mode 100644
index 00000000000000..37d166953c3336
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/X86/gep-with-extractelement-many-users.ll
@@ -0,0 +1,26 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt -S --passes=slp-vectorizer -mtriple=x86_64-unknown-linux-gnu -slp-threshold=-99999 < %s | FileCheck %s
+
+define void @test() {
+; CHECK-LABEL: define void @test() {
+; CHECK-NEXT: bb:
+; CHECK-NEXT: [[TMP0:%.*]] = extractelement <2 x ptr> zeroinitializer, i32 0
+; CHECK-NEXT: [[GETELEMENTPTR6:%.*]] = getelementptr i8, ptr [[TMP0]], i64 872
+; CHECK-NEXT: store double 0.000000e+00, ptr [[GETELEMENTPTR6]], align 8
+; CHECK-NEXT: br label [[BB9:%.*]]
+; CHECK: bb9:
+; CHECK-NEXT: [[TMP1:%.*]] = phi <2 x ptr> [ getelementptr (i8, <2 x ptr> zeroinitializer, <2 x i64> <i64 32, i64 872>), [[BB:%.*]] ]
+; CHECK-NEXT: ret void
+;
+bb:
+ %getelementptr = getelementptr i8, ptr null, i64 32
+ %0 = extractelement <2 x ptr> zeroinitializer, i32 0
+ %getelementptr6 = getelementptr i8, ptr %0, i64 872
+ store double 0.000000e+00, ptr %getelementptr6, align 8
+ br label %bb9
+
+bb9:
+ %phi10 = phi ptr [ %getelementptr, %bb ]
+ %phi11 = phi ptr [ %getelementptr6, %bb ]
+ ret void
+}
More information about the llvm-commits
mailing list