diff --git a/clang/lib/CodeGen/TargetInfo.cpp b/clang/lib/CodeGen/TargetInfo.cpp
--- a/clang/lib/CodeGen/TargetInfo.cpp
+++ b/clang/lib/CodeGen/TargetInfo.cpp
@@ -7050,10 +7050,10 @@
 
   // Empty records are ignored for parameter passing purposes.
   if (isEmptyRecord(getContext(), Ty, true)) {
-    Address Addr = Address(CGF.Builder.CreateLoad(VAListAddr),
-                           getVAListElementType(CGF), SlotSize);
-    Addr = CGF.Builder.CreateElementBitCast(Addr, CGF.ConvertTypeForMem(Ty));
-    return Addr;
+    VAListAddr = CGF.Builder.CreateElementBitCast(VAListAddr, CGF.Int8PtrTy);
+    auto *Load = CGF.Builder.CreateLoad(VAListAddr);
+    Address Addr = Address(Load, CGF.Int8Ty, SlotSize);
+    return CGF.Builder.CreateElementBitCast(Addr, CGF.ConvertTypeForMem(Ty));
   }
 
   CharUnits TySize = getContext().getTypeSizeInChars(Ty);
diff --git a/clang/test/CodeGen/arm-vaarg.c b/clang/test/CodeGen/arm-vaarg.c
new file mode 100644
--- /dev/null
+++ b/clang/test/CodeGen/arm-vaarg.c
@@ -0,0 +1,23 @@
+// RUN: %clang -Xclang -no-opaque-pointers -mfloat-abi=soft -target arm-linux-gnu -emit-llvm -S -o - %s | FileCheck %s
+
+struct Empty {};
+
+struct Empty emptyvar;
+
+void take_args(int a, ...) {
+// CHECK: [[ALLOCA_VA_LIST:%[a-zA-Z0-9._]+]] = alloca %struct.__va_list, align 4
+// CHECK: call void @llvm.va_start
+// CHECK-NEXT: [[AP_ADDR:%[a-zA-Z0-9._]+]] = bitcast %struct.__va_list* [[ALLOCA_VA_LIST]] to i8**
+// CHECK-NEXT: [[LOAD_AP:%[a-zA-Z0-9._]+]] = load i8*, i8** [[AP_ADDR]], align 4
+// CHECK-NEXT: [[EMPTY_PTR:%[a-zA-Z0-9._]+]] = bitcast i8* [[LOAD_AP]] to %struct.Empty*
+
+  // It's conceivable that EMPTY_PTR may not actually be a valid pointer
+  // (e.g. it's at the very bottom of the stack and the next page is
+  // invalid). This doesn't matter provided it's never loaded (there's no
+  // well-defined way to tell), but it becomes a problem if we do try to use it.
+// CHECK-NOT: load %struct.Empty, %struct.Empty* [[EMPTY_PTR]]
+  __builtin_va_list l;
+  __builtin_va_start(l, a);
+  emptyvar = __builtin_va_arg(l, struct Empty);
+  __builtin_va_end(l);
+}