Index: lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -12777,10 +12777,10 @@
     }
 
     // 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;
     }
Index: test/CodeGen/AArch64/pr33172.ll
===================================================================
--- /dev/null
+++ test/CodeGen/AArch64/pr33172.ll
@@ -0,0 +1,37 @@
+; RUN: llc < %s | FileCheck %s
+
+; CHECK-LABEL: pr33172
+; CHECK: ldp
+; CHECK: sdp
+
+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, !tbaa !1
+  %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, !tbaa !1
+  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, !tbaa !1
+  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, !tbaa !1
+  %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, !tbaa !1
+  %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, !tbaa !1
+  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, !tbaa !1
+  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, !tbaa !1
+  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 }
+
+!1 = !{!2, !2, i64 0}
+!2 = !{!"float", !3, i64 0}
+!3 = !{!"omnipotent char", !4, i64 0}
+!4 = !{!"Simple C/C++ TBAA"}