[DAG] Avoid use of stale store.

Correct references to alignment of store which may be deleted in a
previous iteration of merge. Instead use first store that would be
merged.

Corrects pr33172's use-after-poison caught by ASan.

Reviewers: spatel, hfinkel, RKSimon

Reviewed By: RKSimon

Subscribers: thegameg, javed.absar, llvm-commits

Differential Revision: https://reviews.llvm.org/D33686

llvm-svn: 304299

2 files changed, 34 insertions, 2 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index ab36bc1417a..a54ca4d32cd 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -12777,10 +12777,10 @@ bool DAGCombiner::MergeConsecutiveStores(StoreSDNode *St) {
     }
 
     // If we have load/store pair instructions and we only have two values,
-    // don't bother.
+    // don't bother merging.
     unsigned RequiredAlignment;
     if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
-        St->getAlignment() >= RequiredAlignment) {
+        StoreNodes[0].MemNode->getAlignment() >= RequiredAlignment) {
       StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + 2);
       continue;
     }
diff --git a/llvm/test/CodeGen/AArch64/pr33172.ll b/llvm/test/CodeGen/AArch64/pr33172.ll
new file mode 100644
index 00000000000..1e1da78b28f
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/pr33172.ll
@@ -0,0 +1,32 @@
+; RUN: llc < %s | FileCheck %s
+
+; CHECK-LABEL: pr33172
+; CHECK: ldp
+; CHECK: stp
+
+target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
+target triple = "arm64-apple-ios10.3.0"
+
+@main.b = external global [200 x float], align 8
+@main.x = external global [200 x float], align 8
+
+; Function Attrs: nounwind ssp
+define void @pr33172() local_unnamed_addr  {
+entry:
+  %wide.load8281058.3 = load i64, i64* bitcast (float* getelementptr inbounds ([200 x float], [200 x float]* @main.b, i64 0, i64 12) to i64*), align 8
+  %wide.load8291059.3 = load i64, i64* bitcast (float* getelementptr inbounds ([200 x float], [200 x float]* @main.b, i64 0, i64 14) to i64*), align 8
+  store i64 %wide.load8281058.3, i64* bitcast (float* getelementptr inbounds ([200 x float], [200 x float]* @main.x, i64 0, i64 12) to i64*), align 8
+  store i64 %wide.load8291059.3, i64* bitcast (float* getelementptr inbounds ([200 x float], [200 x float]* @main.x, i64 0, i64 14) to i64*), align 8
+  %wide.load8281058.4 = load i64, i64* bitcast (float* getelementptr inbounds ([200 x float], [200 x float]* @main.b, i64 0, i64 16) to i64*), align 8
+  %wide.load8291059.4 = load i64, i64* bitcast (float* getelementptr inbounds ([200 x float], [200 x float]* @main.b, i64 0, i64 18) to i64*), align 8
+  store i64 %wide.load8281058.4, i64* bitcast (float* getelementptr inbounds ([200 x float], [200 x float]* @main.x, i64 0, i64 16) to i64*), align 8
+  store i64 %wide.load8291059.4, i64* bitcast (float* getelementptr inbounds ([200 x float], [200 x float]* @main.x, i64 0, i64 18) to i64*), align 8
+  tail call void @llvm.memset.p0i8.i64(i8* bitcast ([200 x float]* @main.b to i8*), i8 0, i64 undef, i32 8, i1 false) #2
+  unreachable
+}
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.memset.p0i8.i64(i8* nocapture writeonly, i8, i64, i32, i1) #1
+
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { nounwind }