Simplify atomic load/store IRGen.

Also fixes a couple minor bugs along the way; see testcases.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@186049 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/CodeGen/CGExpr.cpp b/lib/CodeGen/CGExpr.cpp
index 419075f..b283174 100644
--- a/lib/CodeGen/CGExpr.cpp
+++ b/lib/CodeGen/CGExpr.cpp
@@ -2717,11 +2717,10 @@
   case CK_BuiltinFnToFnPtr:
     llvm_unreachable("builtin functions are handled elsewhere");
 
-  // These two casts are currently treated as no-ops, although they could
-  // potentially be real operations depending on the target's ABI.
+  // These are never l-values; just use the aggregate emission code.
   case CK_NonAtomicToAtomic:
   case CK_AtomicToNonAtomic:
-    return EmitLValue(E->getSubExpr());
+    return EmitAggExprToLValue(E);
 
   case CK_Dynamic: {
     LValue LV = EmitLValue(E->getSubExpr());