diff --git a/clang/test/CodeGen/aarch64-ABI-align-packed.c b/clang/test/CodeGen/aarch64-ABI-align-packed.c
new file mode 100644
--- /dev/null
+++ b/clang/test/CodeGen/aarch64-ABI-align-packed.c
@@ -0,0 +1,240 @@
+// RUN: %clang_cc1 -fsyntax-only -triple aarch64-none-eabi \
+// RUN:   -target-feature +neon \
+// RUN:   -emit-llvm -o - %s | FileCheck %s
+#include <stdarg.h>
+#include <arm_neon.h>
+
+// natural alignment 16, adjusted alignment 16
+// expected alignment of copy on callee stack: 16
+struct non_packed_struct {
+  uint16x8_t M0; // member alignment 16
+};
+
+// natural alignment 1, adjusted alignment 1
+// expected alignment of copy on callee stack: 8
+struct __attribute((packed)) packed_struct {
+  uint16x8_t M0; // member alignment 1, because the field is packed when the struct is packed
+};
+
+// natural alignment 1, adjusted alignment 1
+// expected alignment of copy on callee stack: 8
+struct packed_member {
+  uint16x8_t M0 __attribute((packed)); // member alignment 1
+};
+
+// natural alignment 16, adjusted alignment 16 since __attribute((aligned (n))) sets the minimum alignment
+// expected alignment of copy on callee stack: 16
+struct __attribute((aligned (8))) aligned_struct_8 {
+  uint16x8_t M0; // member alignment 16
+};
+
+// natural alignment 16, adjusted alignment 16
+// expected alignment of copy on callee stack: 16
+struct aligned_member_8 {
+  uint16x8_t M0 __attribute((aligned (8))); // member alignment 16 since __attribute((aligned (n))) sets the minimum alignment
+};
+
+// natural alignment 8, adjusted alignment 8
+// expected alignment of copy on callee stack: 8
+#pragma pack(8)
+struct pragma_packed_struct_8 {
+  uint16x8_t M0; // member alignment 8 because the struct is subject to packed(8)
+};
+
+// natural alignment 4, adjusted alignment 4
+// expected alignment of copy on callee stack: 8
+#pragma pack(4)
+struct pragma_packed_struct_4 {
+  uint16x8_t M0; // member alignment 4 because the struct is subject to packed(4)
+};
+
+// Struct passed as a named argument
+// CHECK-LABEL: define dso_local void @named_arg_non_packed_struct
+// CHECK-SAME: ([1 x <8 x i16>] [[ARG_COERCE:%.*]]) #[[ATTR0:[0-9]+]] {
+void named_arg_non_packed_struct(struct non_packed_struct arg) {}
+
+// CHECK-LABEL: define dso_local void @named_arg_packed_struct
+// CHECK-SAME: ([1 x <8 x i16>] [[ARG_COERCE:%.*]]) #[[ATTR0]] {
+void named_arg_packed_struct(struct packed_struct arg) {}
+
+// CHECK-LABEL: define dso_local void @named_arg_packed_member
+// CHECK-SAME: ([1 x <8 x i16>] [[ARG_COERCE:%.*]]) #[[ATTR0]] {
+void named_arg_packed_member(struct packed_member arg) {}
+
+// CHECK-LABEL: define dso_local void @named_arg_aligned_struct_8
+// CHECK-SAME: ([1 x <8 x i16>] [[ARG_COERCE:%.*]]) #[[ATTR0]] {
+void named_arg_aligned_struct_8(struct aligned_struct_8 arg) {}
+
+// CHECK-LABEL: define dso_local void @named_arg_aligned_member_8
+// CHECK-SAME: ([1 x <8 x i16>] [[ARG_COERCE:%.*]]) #[[ATTR0]] {
+void named_arg_aligned_member_8(struct aligned_member_8 arg) {}
+
+// CHECK-LABEL: define dso_local void @named_arg_pragma_packed_struct_8
+// CHECK-SAME: ([1 x <8 x i16>] [[ARG_COERCE:%.*]]) #[[ATTR0]] {
+void named_arg_pragma_packed_struct_8(struct pragma_packed_struct_8 arg) {}
+
+// CHECK-LABEL: define dso_local void @named_arg_pragma_packed_struct_4
+// CHECK-SAME: ([1 x <8 x i16>] [[ARG_COERCE:%.*]]) #[[ATTR0]] {
+void named_arg_pragma_packed_struct_4(struct pragma_packed_struct_4 arg) {}
+
+// Struct passed as a variadic argument
+// CHECK-LABEL: define dso_local void @variadic_non_packed_struct
+// CHECK:       vaarg.end:
+// CHECK:       call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[ON_CALLEE_STACK:.*]], ptr align 16 [[VAARGS_ADDR:.*]], i64 16, i1 false)
+// CHECK-NEXT:    ret void
+void variadic_non_packed_struct(int named_arg, ...) {
+  va_list vl;
+  va_start(vl, named_arg);
+  struct non_packed_struct on_callee_stack;
+  on_callee_stack = va_arg(vl, struct non_packed_struct);
+}
+
+// CHECK-LABEL: define dso_local void @variadic_packed_struct
+// CHECK:       vaarg.end:
+// CHECK:       call void @llvm.memcpy.p0.p0.i64(ptr align 1 [[ON_CALLEE_STACK:.*]], ptr align 8 [[VAARGS_ADDR:.*]], i64 16, i1 false)
+// CHECK-NEXT:    ret void
+void variadic_packed_struct(int named_arg, ...) {
+  va_list vl;
+  va_start(vl, named_arg);
+  struct packed_struct on_callee_stack;
+  on_callee_stack = va_arg(vl, struct packed_struct);
+}
+
+// CHECK-LABEL: define dso_local void @variadic_packed_member
+// CHECK:       vaarg.end:
+// CHECK:       call void @llvm.memcpy.p0.p0.i64(ptr align 1 [[ON_CALLEE_STACK:.*]], ptr align 8 [[VAARGS_ADDR:.*]], i64 16, i1 false)
+// CHECK-NEXT:    ret void
+void variadic_packed_member(int named_arg, ...) {
+  va_list vl;
+  va_start(vl, named_arg);
+  struct packed_member on_callee_stack;
+  on_callee_stack = va_arg(vl, struct packed_member);
+}
+
+// CHECK-LABEL: define dso_local void @variadic_aligned_struct_8
+// CHECK:       vaarg.end:
+// CHECK:       call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[ON_CALLEE_STACK:.*]], ptr align 16 [[VAARGS_ADDR:.*]], i64 16, i1 false)
+// CHECK-NEXT:    ret void
+void variadic_aligned_struct_8(int named_arg, ...) {
+  va_list vl;
+  va_start(vl, named_arg);
+  struct aligned_struct_8 on_callee_stack;
+  on_callee_stack = va_arg(vl, struct aligned_struct_8);
+}
+
+// CHECK-LABEL: define dso_local void @variadic_aligned_member_8
+// CHECK:       vaarg.end:
+// CHECK:       call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[ON_CALLEE_STACK:.*]], ptr align 16 [[VAARGS_ADDR:.*]], i64 16, i1 false)
+// CHECK-NEXT:    ret void
+void variadic_aligned_member_8(int named_arg, ...) {
+  va_list vl;
+  va_start(vl, named_arg);
+  struct aligned_member_8 on_callee_stack;
+  on_callee_stack = va_arg(vl, struct aligned_member_8);
+}
+
+// CHECK-LABEL: define dso_local void @variadic_pragma_packed_struct_8
+// CHECK:       vaarg.end:
+// CHECK:       call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[ON_CALLEE_STACK:.*]], ptr align 8 [[VAARGS_ADDR:.*]], i64 16, i1 false)
+// CHECK-NEXT:    ret void
+void variadic_pragma_packed_struct_8(int named_arg, ...) {
+  va_list vl;
+  va_start(vl, named_arg);
+  struct pragma_packed_struct_8 on_callee_stack;
+  on_callee_stack = va_arg(vl, struct pragma_packed_struct_8);
+}
+
+// CHECK-LABEL: define dso_local void @variadic_pragma_packed_struct_4
+// CHECK:       vaarg.end:
+// CHECK:       call void @llvm.memcpy.p0.p0.i64(ptr align 4 [[ON_CALLEE_STACK:.*]], ptr align 8 [[VAARGS_ADDR:.*]], i64 16, i1 false)
+// CHECK-NEXT:    ret void
+void variadic_pragma_packed_struct_4(int named_arg, ...) {
+  va_list vl;
+  va_start(vl, named_arg);
+  struct pragma_packed_struct_4 on_callee_stack;
+  on_callee_stack = va_arg(vl, struct pragma_packed_struct_4);
+}
+
+// CHECK-LABEL: define dso_local void @test_non_packed_struct
+// CHECK:       call void @named_arg_non_packed_struct([1 x <8 x i16>] [[TMP3:.*]])
+// CHECK-NEXT:    [[COERCE_DIVE7:%.*]] = getelementptr inbounds [[STRUCT_NON_PACKED_STRUCT:.*]], ptr [[P_NON_PACKED_STRUCT:.*]], i32 0, i32 0
+// CHECK-NEXT:    [[TMP4:%.*]] = load [1 x <8 x i16>], ptr [[COERCE_DIVE7:.*]], align 16
+// CHECK-NEXT:    call void (i32, ...) @variadic_non_packed_struct(i32 noundef 0, [1 x <8 x i16>] [[TMP4:.*]])
+// CHECK-NEXT:    ret void
+void test_non_packed_struct() {
+  struct non_packed_struct P_non_packed_struct = {vld1q_u16(((const uint16_t[8]){1234, 0, 0, 0, 0, 0, 0, 0}))};
+  named_arg_non_packed_struct(P_non_packed_struct);
+  variadic_non_packed_struct(0, P_non_packed_struct);
+}
+
+// CHECK-LABEL: define dso_local void @test_packed_struct
+// CHECK:         call void @named_arg_packed_struct([1 x <8 x i16>] [[TMP3:.*]])
+// CHECK-NEXT:    [[COERCE_DIVE7:%.*]] = getelementptr inbounds [[STRUCT_PACKED_STRUCT:.*]], ptr [[P_PACKED_STRUCT:.*]], i32 0, i32 0
+// CHECK-NEXT:    [[TMP4:%.*]] = load [1 x <8 x i16>], ptr [[COERCE_DIVE7:.*]], align 1
+// CHECK-NEXT:    call void (i32, ...) @variadic_packed_struct(i32 noundef 0, [1 x <8 x i16>] [[TMP4:.*]])
+// CHECK-NEXT:    ret void
+void test_packed_struct() {
+  struct packed_struct P_packed_struct = {vld1q_u16(((const uint16_t[8]){1234, 0, 0, 0, 0, 0, 0, 0}))};
+  named_arg_packed_struct(P_packed_struct);
+  variadic_packed_struct(0, P_packed_struct);
+}
+
+// CHECK-LABEL: define dso_local void @test_packed_member
+// CHECK:         call void @named_arg_packed_member([1 x <8 x i16>] [[TMP3:.*]])
+// CHECK-NEXT:    [[COERCE_DIVE7:%.*]] = getelementptr inbounds [[STRUCT_PACKED_MEMBER:.*]], ptr [[P_PACKED_MEMBER:.*]], i32 0, i32 0
+// CHECK-NEXT:    [[TMP4:%.*]] = load [1 x <8 x i16>], ptr [[COERCE_DIVE7:.*]], align 1
+// CHECK-NEXT:    call void (i32, ...) @variadic_packed_member(i32 noundef 0, [1 x <8 x i16>] [[TMP4:.*]])
+// CHECK-NEXT:    ret void
+void test_packed_member() {
+  struct packed_member P_packed_member = {vld1q_u16(((const uint16_t[8]){1234, 0, 0, 0, 0, 0, 0, 0}))};
+  named_arg_packed_member(P_packed_member);
+  variadic_packed_member(0, P_packed_member);
+}
+
+// CHECK-LABEL: define dso_local void @test_aligned_struct_8
+// CHECK:         call void @named_arg_aligned_struct_8([1 x <8 x i16>] [[TMP3:.*]])
+// CHECK-NEXT:    [[COERCE_DIVE7:%.*]] = getelementptr inbounds [[STRUCT_ALIGNED_STRUCT_8:.*]], ptr [[P_ALIGNED_STRUCT_8:.*]], i32 0, i32 0
+// CHECK-NEXT:    [[TMP4:%.*]] = load [1 x <8 x i16>], ptr [[COERCE_DIVE7:.*]], align 16
+// CHECK-NEXT:    call void (i32, ...) @variadic_aligned_struct_8(i32 noundef 0, [1 x <8 x i16>] [[TMP4:.*]])
+// CHECK-NEXT:    ret void
+void test_aligned_struct_8() {
+  struct aligned_struct_8 P_aligned_struct_8 = {vld1q_u16(((const uint16_t[8]){1234, 0, 0, 0, 0, 0, 0, 0}))};
+  named_arg_aligned_struct_8(P_aligned_struct_8);
+  variadic_aligned_struct_8(0, P_aligned_struct_8);
+}
+
+// CHECK-LABEL: define dso_local void @test_aligned_member_8
+// CHECK:         call void @named_arg_aligned_member_8([1 x <8 x i16>] [[TMP3:.*]])
+// CHECK-NEXT:    [[COERCE_DIVE7:%.*]] = getelementptr inbounds [[STRUCT_ALIGNED_MEMBER_8:.*]], ptr [[P_ALIGNED_MEMBER_8:.*]], i32 0, i32 0
+// CHECK-NEXT:    [[TMP4:%.*]] = load [1 x <8 x i16>], ptr [[COERCE_DIVE7]], align 16
+// CHECK-NEXT:    call void (i32, ...) @variadic_aligned_member_8(i32 noundef 0, [1 x <8 x i16>] [[TMP4:.*]])
+// CHECK-NEXT:    ret void
+void test_aligned_member_8() {
+  struct aligned_member_8 P_aligned_member_8 = {vld1q_u16(((const uint16_t[8]){1234, 0, 0, 0, 0, 0, 0, 0}))};
+  named_arg_aligned_member_8(P_aligned_member_8);
+  variadic_aligned_member_8(0, P_aligned_member_8);
+}
+
+// CHECK-LABEL: define dso_local void @test_pragma_packed_struct_8
+// CHECK:         call void @named_arg_pragma_packed_struct_8([1 x <8 x i16>] [[TMP3:.*]])
+// CHECK-NEXT:    [[COERCE_DIVE7:%.*]] = getelementptr inbounds [[STRUCT_PRAGMA_PACKED_STRUCT_8:.*]], ptr [[P_PRAGMA_PACKED_STRUCT_8:.*]], i32 0, i32 0
+// CHECK-NEXT:    [[TMP4:%.*]] = load [1 x <8 x i16>], ptr [[COERCE_DIVE7:.*]], align 8
+// CHECK-NEXT:    call void (i32, ...) @variadic_pragma_packed_struct_8(i32 noundef 0, [1 x <8 x i16>] [[TMP4:.*]])
+// CHECK-NEXT:    ret void
+void test_pragma_packed_struct_8() {
+  struct pragma_packed_struct_8 P_pragma_packed_struct_8 = {vld1q_u16(((const uint16_t[8]){1234, 0, 0, 0, 0, 0, 0, 0}))};
+  named_arg_pragma_packed_struct_8(P_pragma_packed_struct_8);
+  variadic_pragma_packed_struct_8(0, P_pragma_packed_struct_8);
+}
+
+// CHECK-LABEL: define dso_local void @test_pragma_packed_struct_4
+// CHECK:         call void @named_arg_pragma_packed_struct_4([1 x <8 x i16>] [[TMP3:.*]])
+// CHECK-NEXT:    [[COERCE_DIVE7:%.*]] = getelementptr inbounds [[STRUCT_PRAGMA_PACKED_STRUCT_4:.*]], ptr [[P_PRAGMA_PACKED_STRUCT_4:.*]], i32 0, i32 0
+// CHECK-NEXT:    [[TMP4:%.*]] = load [1 x <8 x i16>], ptr [[COERCE_DIVE7:.*]], align 4
+// CHECK-NEXT:    call void (i32, ...) @variadic_pragma_packed_struct_4(i32 noundef 0, [1 x <8 x i16>] [[TMP4:.*]])
+// CHECK-NEXT:    ret void
+void test_pragma_packed_struct_4() {
+  struct pragma_packed_struct_4 P_pragma_packed_struct_4 = {vld1q_u16(((const uint16_t[8]){1234, 0, 0, 0, 0, 0, 0, 0}))};
+  named_arg_pragma_packed_struct_4(P_pragma_packed_struct_4);
+  variadic_pragma_packed_struct_4(0, P_pragma_packed_struct_4);
+}