SLPVectorizer: fix wrong scheduling of atomic load/stores.

This fixes PR22306.

llvm-svn: 227077

2 files changed, 43 insertions, 1 deletions

diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index fb1f64b..4dee2d9 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -309,6 +309,17 @@ static AliasAnalysis::Location getLocation(Instruction *I, AliasAnalysis *AA) {
   return AliasAnalysis::Location();
 }
 
+/// \returns True if the instruction is not a volatile or atomic load/store.
+static bool isSimple(Instruction *I) {
+  if (LoadInst *LI = dyn_cast<LoadInst>(I))
+    return LI->isSimple();
+  if (StoreInst *SI = dyn_cast<StoreInst>(I))
+    return SI->isSimple();
+  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
+    return !MI->isVolatile();
+  return true;
+}
+
 /// Bottom Up SLP Vectorizer.
 class BoUpSLP {
 public:
@@ -501,7 +512,7 @@ private:
     }
     AliasAnalysis::Location Loc2 = getLocation(Inst2, AA);
     bool aliased = true;
-    if (Loc1.Ptr && Loc2.Ptr) {
+    if (Loc1.Ptr && Loc2.Ptr && isSimple(Inst1) && isSimple(Inst2)) {
       // Do the alias check.
       aliased = AA->alias(Loc1, Loc2);
     }
diff --git a/llvm/test/Transforms/SLPVectorizer/X86/atomics.ll b/llvm/test/Transforms/SLPVectorizer/X86/atomics.ll
new file mode 100644
index 0000000..6cb322e
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/X86/atomics.ll
@@ -0,0 +1,31 @@
+; RUN: opt < %s -basicaa -slp-vectorizer -S |FileCheck %s
+target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
+
+@x = global [4 x i32] zeroinitializer, align 16
+@a = global [4 x i32] zeroinitializer, align 16
+
+; The SLPVectorizer should not vectorize atomic stores and it should not
+; schedule regular stores around atomic stores.
+
+; CHECK-LABEL: test
+; CHECK: store i32
+; CHECK: store atomic i32
+; CHECK: store i32
+; CHECK: store atomic i32
+; CHECK: store i32
+; CHECK: store atomic i32
+; CHECK: store i32
+; CHECK: store atomic i32
+define void @test() {
+entry:
+  store i32 0, i32* getelementptr inbounds ([4 x i32]* @a, i64 0, i64 0), align 16
+  store atomic i32 0, i32* getelementptr inbounds ([4 x i32]* @x, i64 0, i64 0) release, align 16
+  store i32 0, i32* getelementptr inbounds ([4 x i32]* @a, i64 0, i64 1), align 4
+  store atomic i32 1, i32* getelementptr inbounds ([4 x i32]* @x, i64 0, i64 1) release, align 4
+  store i32 0, i32* getelementptr inbounds ([4 x i32]* @a, i64 0, i64 2), align 8
+  store atomic i32 2, i32* getelementptr inbounds ([4 x i32]* @x, i64 0, i64 2) release, align 8
+  store i32 0, i32* getelementptr inbounds ([4 x i32]* @a, i64 0, i64 3), align 4
+  store atomic i32 3, i32* getelementptr inbounds ([4 x i32]* @x, i64 0, i64 3) release, align 4
+  ret void
+}
+