[SLP]Fix PR49898: Infinite loop in SLP vectorizer.
We should not re-try attempt of finding of the consecutive store chain
if it was tried before.
Differential Revision: https://reviews.llvm.org/D100131
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index 76bfdcc..431d0d2 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -6214,6 +6214,9 @@
break;
}
+ // Tracks if we tried to vectorize stores starting from the given tail
+ // already.
+ SmallBitVector TriedTails(E, false);
// For stores that start but don't end a link in the chain:
for (int Cnt = E; Cnt > 0; --Cnt) {
int I = Cnt - 1;
@@ -6230,8 +6233,9 @@
// Mark the new end in the chain and go back, if required. It might be
// required if the original stores come in reversed order, for example.
if (ConsecutiveChain[I].first != E &&
- Tails.test(ConsecutiveChain[I].first) &&
+ Tails.test(ConsecutiveChain[I].first) && !TriedTails.test(I) &&
!VectorizedStores.count(Stores[ConsecutiveChain[I].first])) {
+ TriedTails.set(I);
Tails.reset(ConsecutiveChain[I].first);
if (Cnt < ConsecutiveChain[I].first + 2)
Cnt = ConsecutiveChain[I].first + 2;
diff --git a/llvm/test/Transforms/SLPVectorizer/X86/several_store_chains.ll b/llvm/test/Transforms/SLPVectorizer/X86/several_store_chains.ll
new file mode 100644
index 0000000..4dfe694
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/X86/several_store_chains.ll
@@ -0,0 +1,40 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -S -mtriple=x86_64-unknown -slp-vectorizer | FileCheck %s
+
+; Test for PR49898.
+define void @fusion_1506(i8* %temp_buf1) local_unnamed_addr {
+; CHECK-LABEL: @fusion_1506(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds i8, i8* [[TEMP_BUF1:%.*]], i64 5621415936
+; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds i8, i8* [[TEMP_BUF1]], i64 7278166016
+; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i8, i8* [[TEMP_BUF1]], i64 5097127936
+; CHECK-NEXT: [[TMP3:%.*]] = bitcast i8* [[TMP2]] to float*
+; CHECK-NEXT: [[TMP4:%.*]] = bitcast i8* [[TMP1]] to float*
+; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds float, float* [[TMP4]], i64 undef
+; CHECK-NEXT: store float undef, float* [[TMP5]], align 16
+; CHECK-NEXT: [[TMP6:%.*]] = bitcast i8* [[TMP0]] to float*
+; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds float, float* [[TMP6]], i64 undef
+; CHECK-NEXT: store float undef, float* [[TMP7]], align 16
+; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds float, float* [[TMP6]], i64 undef
+; CHECK-NEXT: store float undef, float* [[TMP8]], align 4
+; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds float, float* [[TMP3]], i64 undef
+; CHECK-NEXT: store float undef, float* [[TMP9]], align 4
+; CHECK-NEXT: ret void
+;
+entry:
+ %0 = getelementptr inbounds i8, i8* %temp_buf1, i64 5621415936
+ %1 = getelementptr inbounds i8, i8* %temp_buf1, i64 7278166016
+ %2 = getelementptr inbounds i8, i8* %temp_buf1, i64 5097127936
+ %3 = bitcast i8* %2 to float*
+ %4 = bitcast i8* %1 to float*
+ %5 = getelementptr inbounds float, float* %4, i64 undef
+ store float undef, float* %5, align 16
+ %6 = bitcast i8* %0 to float*
+ %7 = getelementptr inbounds float, float* %6, i64 undef
+ store float undef, float* %7, align 16
+ %8 = getelementptr inbounds float, float* %6, i64 undef
+ store float undef, float* %8, align 4
+ %9 = getelementptr inbounds float, float* %3, i64 undef
+ store float undef, float* %9, align 4
+ ret void
+}
