Index: llvm/trunk/include/llvm/IR/Metadata.h
===================================================================
--- llvm/trunk/include/llvm/IR/Metadata.h
+++ llvm/trunk/include/llvm/IR/Metadata.h
@@ -879,6 +879,7 @@
   static MDNode *getMostGenericTBAA(MDNode *A, MDNode *B);
   static MDNode *getMostGenericFPMath(MDNode *A, MDNode *B);
   static MDNode *getMostGenericRange(MDNode *A, MDNode *B);
+  static MDNode *getMostGenericAliasScope(MDNode *A, MDNode *B);
 };
 
 /// \brief Tuple of metadata.
Index: llvm/trunk/lib/Analysis/TypeBasedAliasAnalysis.cpp
===================================================================
--- llvm/trunk/lib/Analysis/TypeBasedAliasAnalysis.cpp
+++ llvm/trunk/lib/Analysis/TypeBasedAliasAnalysis.cpp
@@ -623,8 +623,8 @@
     N.TBAA = getMetadata(LLVMContext::MD_tbaa);
 
   if (Merge)
-    N.Scope =
-        MDNode::intersect(N.Scope, getMetadata(LLVMContext::MD_alias_scope));
+    N.Scope = MDNode::getMostGenericAliasScope(
+        N.Scope, getMetadata(LLVMContext::MD_alias_scope));
   else
     N.Scope = getMetadata(LLVMContext::MD_alias_scope);
 
Index: llvm/trunk/lib/IR/Metadata.cpp
===================================================================
--- llvm/trunk/lib/IR/Metadata.cpp
+++ llvm/trunk/lib/IR/Metadata.cpp
@@ -782,6 +782,28 @@
   return getOrSelfReference(A->getContext(), MDs);
 }
 
+MDNode *MDNode::getMostGenericAliasScope(MDNode *A, MDNode *B) {
+  if (!A || !B)
+    return nullptr;
+
+  SmallVector<Metadata *, 4> MDs(B->op_begin(), B->op_end());
+  for (unsigned i = 0, ie = A->getNumOperands(); i != ie; ++i) {
+    Metadata *MD = A->getOperand(i);
+    bool insert = true;
+    for (unsigned j = 0, je = B->getNumOperands(); j != je; ++j)
+      if (MD == B->getOperand(j)) {
+        insert = false;
+        break;
+      }
+    if (insert)
+        MDs.push_back(MD);
+  }
+
+  // FIXME: This preserves long-standing behaviour, but is it really the right
+  // behaviour?  Or was that an unintended side-effect of node uniquing?
+  return getOrSelfReference(A->getContext(), MDs);
+}
+
 MDNode *MDNode::getMostGenericFPMath(MDNode *A, MDNode *B) {
   if (!A || !B)
     return nullptr;
Index: llvm/trunk/lib/Transforms/Utils/Local.cpp
===================================================================
--- llvm/trunk/lib/Transforms/Utils/Local.cpp
+++ llvm/trunk/lib/Transforms/Utils/Local.cpp
@@ -1334,6 +1334,8 @@
         K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
         break;
       case LLVMContext::MD_alias_scope:
+        K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD));
+        break;
       case LLVMContext::MD_noalias:
         K->setMetadata(Kind, MDNode::intersect(JMD, KMD));
         break;
Index: llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp
===================================================================
--- llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -217,6 +217,8 @@
         MD = MDNode::getMostGenericTBAA(MD, IMD);
         break;
       case LLVMContext::MD_alias_scope:
+        MD = MDNode::getMostGenericAliasScope(MD, IMD);
+        break;
       case LLVMContext::MD_noalias:
         MD = MDNode::intersect(MD, IMD);
         break;
Index: llvm/trunk/test/Transforms/Util/combine-alias-scope-metadata.ll
===================================================================
--- llvm/trunk/test/Transforms/Util/combine-alias-scope-metadata.ll
+++ llvm/trunk/test/Transforms/Util/combine-alias-scope-metadata.ll
@@ -0,0 +1,24 @@
+; RUN: opt < %s -S -basicaa -memcpyopt | FileCheck %s
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+define void @test(i8* noalias dereferenceable(1) %in, i8* noalias dereferenceable(1) %out) {
+  %tmp = alloca i8
+  %tmp2 = alloca i8
+; CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* %out, i8* %in, i64 1, i32 8, i1 false)
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %tmp, i8* %in, i64 1, i32 8, i1 false), !alias.scope !4
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %tmp2, i8* %tmp, i64 1, i32 8, i1 false), !alias.scope !5
+
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %out, i8* %tmp2, i64 1, i32 8, i1 false), !noalias !6
+
+  ret void
+}
+
+declare void @llvm.memcpy.p0i8.p0i8.i64(i8*, i8*, i64, i32, i1)
+
+!0 = !{!0}
+!1 = distinct !{!1, !0, !"in"}
+!2 = distinct !{!2, !0, !"tmp"}
+!3 = distinct !{!3, !0, !"tmp2"}
+!4 = distinct !{!1, !2}
+!5 = distinct !{!2, !3}
+!6 = distinct !{!1, !2}