[X86] Add explicit alignment to __m128/__m128i/__m128d/etc. to allow matching of MSVC behavior with #pragma pack.

[craig.topper]

Summary

With MSVC, #pragma pack is ignored when there is explicit alignment. This differs from gcc. Clang emulates this difference when compiling for Windows.

It appears that MSVC and headers consider the m128/m128i/__m128d/etc. types to be explicitly aligned and ignores #pragma pack for them. Since we don't have explicit alignment on them in our headers, we don't match the MSVC behavior here.

This patch adds explicit alignment to match this behavior. I'm hoping this won't cause any problems when we're not emulating MSVC. But if someone knows of something that would be different we can swith to conditionally adding the alignment based on _MSC_VER.

I had to add explicitly unaligned types as well so we could use them in the loadu/storeu intrinsics which use attribute(packed). So the explicitly aligned types wouldn't produce align 1 accesses when targeting Windows.

[erichkeane]

The justification is correct based on my research into this problem, and the code changes themselves look correct. I cannot think of any reason why re-stating the alignment will matter in GCC mode, so I think this is OK.

I'll accept, but I'd really like to give @rnk a chance to take a look to make sure there isn't something I'm missing.

test/CodeGen/x86-vec-struct-packing.c

14

We default to C++14, do we not use static-asserts for this?

[craig.topper]

Forgot to do MMX's __m64 type

[rnk]

Sounds good.

This change reminded me of D46042, https://crbug.com/849251, and rL333791, which I had to revert, and I don't think it relanded. I think your change to add the *_u typedefs fixes the issues encountered there, and perhaps @rjmccall will be able to reland his change after this lands.

[rnk]

Hm, looks like this broke a bunch of code in chromium:
bad build: https://ci.chromium.org/p/chromium/builders/luci.chromium.ci/ToTWin%28dbg%29/2342
good build: https://ci.chromium.org/p/chromium/builders/luci.chromium.ci/ToTWin%28dbg%29/2341

A trace from one of the crashing tests:

[ RUN      ] AudioDecoderTestScenarios/AudioDecoderTest.DecodesFramesWithVaryingDuration/3
Received fatal exception EXCEPTION_ACCESS_VIOLATION
Backtrace:
	xcorr_kernel_sse [0x00D4CEF7+199] (C:\b\s\w\ir\cache\builder\src\third_party\opus\src\celt\x86\pitch_sse.c:54)
	celt_pitch_xcorr_c [0x00DBC09C+156] (C:\b\s\w\ir\cache\builder\src\third_party\opus\src\celt\pitch.c:264)
	_celt_autocorr [0x00DBAF20+368] (C:\b\s\w\ir\cache\builder\src\third_party\opus\src\celt\celt_lpc.c:266)
	pitch_downsample [0x00DBBC98+536] (C:\b\s\w\ir\cache\builder\src\third_party\opus\src\celt\pitch.c:182)
	run_prefilter [0x00D4093F+703] (C:\b\s\w\ir\cache\builder\src\third_party\opus\src\celt\celt_encoder.c:1144)
	celt_encode_with_ec [0x00D3C915+3701] (C:\b\s\w\ir\cache\builder\src\third_party\opus\src\celt\celt_encoder.c:1606)
	opus_encode_native [0x00CBBF8A+16810] (C:\b\s\w\ir\cache\builder\src\third_party\opus\src\src\opus_encoder.c:2068)
	opus_encode [0x00CBE1C6+374] (C:\b\s\w\ir\cache\builder\src\third_party\opus\src\src\opus_encoder.c:2241)
	media::cast::AudioDecoderTest::FeedMoreAudio [0x0017F0B8+1064] (C:\b\s\w\ir\cache\builder\src\media\cast\receiver\audio_decoder_unittest.cc:111)

The code in question is calling _mm_loadu_ps:
https://cs.chromium.org/chromium/src/third_party/opus/src/celt/x86/pitch_sse.c?type=cs&q=%22opus/src/celt/x86/pitch_sse.c%22&sq=package:chromium&g=0&l=54

for (j = 0; j < len-3; j += 4)
{
   __m128 x0 = _mm_loadu_ps(x+j);
   __m128 yj = _mm_loadu_ps(y+j);
   __m128 y3 = _mm_loadu_ps(y+j+3); // crashes

lib/Headers/xmmintrin.h

1754

I guess we just missed this.

[rnk]

I audited all the intrin header uses of packed, and I think this was the only one that was missing, so I'll try to fix forward.

[craig.topper]

Thanks for the fix @rnk, I've added some MSVC command lines to some of the other test files too.

lib/Headers/avx512bwintrin.h

_mm512_kunpackw (__mmask32 __A, __mmask32 __B)
   return (__mmask32) __builtin_ia32_kunpcksi ((__mmask32) __A,
                 (__mmask32) __B);
 }
 
 static __inline __m512i __DEFAULT_FN_ATTRS512
 _mm512_loadu_epi16 (void const *__P)
 {
   struct __loadu_epi16 {
-    __m512i __v;
+    __m512i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_epi16*)__P)->__v;
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 _mm512_mask_loadu_epi16 (__m512i __W, __mmask32 __U, void const *__P)
 {
   return (__m512i) __builtin_ia32_loaddquhi512_mask ((__v32hi *) __P,
return (__m512i) __builtin_ia32_loaddquhi512_mask ((__v32hi *) __P,
                  _mm512_setzero_si512 (),
                  (__mmask32) __U);
 }
 
 static __inline __m512i __DEFAULT_FN_ATTRS512
 _mm512_loadu_epi8 (void const *__P)
 {
   struct __loadu_epi8 {
-    __m512i __v;
+    __m512i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_epi8*)__P)->__v;
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 _mm512_mask_loadu_epi8 (__m512i __W, __mmask64 __U, void const *__P)
 {
   return (__m512i) __builtin_ia32_loaddquqi512_mask ((__v64qi *) __P,
return (__m512i) __builtin_ia32_loaddquqi512_mask ((__v64qi *) __P,
                  _mm512_setzero_si512 (),
                  (__mmask64) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS512
 _mm512_storeu_epi16 (void *__P, __m512i __A)
 {
   struct __storeu_epi16 {
-    __m512i __v;
+    __m512i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_epi16*)__P)->__v = __A;
 }
 
 static __inline__ void __DEFAULT_FN_ATTRS512
 _mm512_mask_storeu_epi16 (void *__P, __mmask32 __U, __m512i __A)
 {
   __builtin_ia32_storedquhi512_mask ((__v32hi *) __P,
              (__v32hi) __A,
              (__mmask32) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS512
 _mm512_storeu_epi8 (void *__P, __m512i __A)
 {
   struct __storeu_epi8 {
-    __m512i __v;
+    __m512i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_epi8*)__P)->__v = __A;
 }
 
 static __inline__ void __DEFAULT_FN_ATTRS512
 _mm512_mask_storeu_epi8 (void *__P, __mmask64 __U, __m512i __A)
 {
   __builtin_ia32_storedquqi512_mask ((__v64qi *) __P,

lib/Headers/avx512fintrin.h

 typedef int __v16si __attribute__((__vector_size__(64)));
 
 /* Unsigned types */
 typedef unsigned char __v64qu __attribute__((__vector_size__(64)));
 typedef unsigned short __v32hu __attribute__((__vector_size__(64)));
 typedef unsigned long long __v8du __attribute__((__vector_size__(64)));
 typedef unsigned int __v16su __attribute__((__vector_size__(64)));
 
-typedef float __m512 __attribute__((__vector_size__(64)));
-typedef double __m512d __attribute__((__vector_size__(64)));
-typedef long long __m512i __attribute__((__vector_size__(64)));
+typedef float __m512 __attribute__((__vector_size__(64), __aligned__(64)));
+typedef double __m512d __attribute__((__vector_size__(64), __aligned__(64)));
+typedef long long __m512i __attribute__((__vector_size__(64), __aligned__(64)));
+
+typedef float __m512_u __attribute__((__vector_size__(64), __aligned__(1)));
+typedef double __m512d_u __attribute__((__vector_size__(64), __aligned__(1)));
+typedef long long __m512i_u __attribute__((__vector_size__(64), __aligned__(1)));
 
 typedef unsigned char __mmask8;
 typedef unsigned short __mmask16;
 
 /* Rounding mode macros.  */
 #define _MM_FROUND_TO_NEAREST_INT   0x00
 #define _MM_FROUND_TO_NEG_INF       0x01
 #define _MM_FROUND_TO_POS_INF       0x02

 
 
 /* SIMD load ops */
 
 static __inline __m512i __DEFAULT_FN_ATTRS512
 _mm512_loadu_si512 (void const *__P)
 {
   struct __loadu_si512 {
-    __m512i __v;
+    __m512i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_si512*)__P)->__v;
 }
 
 static __inline __m512i __DEFAULT_FN_ATTRS512
 _mm512_loadu_epi32 (void const *__P)
 {
   struct __loadu_epi32 {
-    __m512i __v;
+    __m512i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_epi32*)__P)->__v;
 }
 
 static __inline __m512i __DEFAULT_FN_ATTRS512
 _mm512_mask_loadu_epi32 (__m512i __W, __mmask16 __U, void const *__P)
 {
   return (__m512i) __builtin_ia32_loaddqusi512_mask ((const int *) __P,
return (__m512i) __builtin_ia32_loaddqusi512_mask ((const int *)__P,
                                                      _mm512_setzero_si512 (),
                                                      (__mmask16) __U);
 }
 
 static __inline __m512i __DEFAULT_FN_ATTRS512
 _mm512_loadu_epi64 (void const *__P)
 {
   struct __loadu_epi64 {
-    __m512i __v;
+    __m512i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_epi64*)__P)->__v;
 }
 
 static __inline __m512i __DEFAULT_FN_ATTRS512
 _mm512_mask_loadu_epi64 (__m512i __W, __mmask8 __U, void const *__P)
 {
   return (__m512i) __builtin_ia32_loaddqudi512_mask ((const long long *) __P,
return (__m512d) __builtin_ia32_loadupd512_mask ((const double *)__P,
                                                    _mm512_setzero_pd (),
                                                    (__mmask8) __U);
 }
 
 static __inline __m512d __DEFAULT_FN_ATTRS512
 _mm512_loadu_pd(void const *__p)
 {
   struct __loadu_pd {
-    __m512d __v;
+    __m512d_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_pd*)__p)->__v;
 }
 
 static __inline __m512 __DEFAULT_FN_ATTRS512
 _mm512_loadu_ps(void const *__p)
 {
   struct __loadu_ps {
-    __m512 __v;
+    __m512_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_ps*)__p)->__v;
 }
 
 static __inline __m512 __DEFAULT_FN_ATTRS512
 _mm512_load_ps(void const *__p)
 {
   return *(__m512*)__p;

 }
 
 /* SIMD store ops */
 
 static __inline void __DEFAULT_FN_ATTRS512
 _mm512_storeu_epi64 (void *__P, __m512i __A)
 {
   struct __storeu_epi64 {
-    __m512i __v;
+    __m512i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_epi64*)__P)->__v = __A;
 }
 
 static __inline void __DEFAULT_FN_ATTRS512
 _mm512_mask_storeu_epi64(void *__P, __mmask8 __U, __m512i __A)
 {
   __builtin_ia32_storedqudi512_mask ((long long *)__P, (__v8di) __A,
                                      (__mmask8) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS512
 _mm512_storeu_si512 (void *__P, __m512i __A)
 {
   struct __storeu_si512 {
-    __m512i __v;
+    __m512i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_si512*)__P)->__v = __A;
 }
 
 static __inline void __DEFAULT_FN_ATTRS512
 _mm512_storeu_epi32 (void *__P, __m512i __A)
 {
   struct __storeu_epi32 {
-    __m512i __v;
+    __m512i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_epi32*)__P)->__v = __A;
 }
 
 static __inline void __DEFAULT_FN_ATTRS512
 _mm512_mask_storeu_epi32(void *__P, __mmask16 __U, __m512i __A)
 {
   __builtin_ia32_storedqusi512_mask ((int *)__P, (__v16si) __A,
                                      (__mmask16) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS512
 _mm512_mask_storeu_pd(void *__P, __mmask8 __U, __m512d __A)
 {
   __builtin_ia32_storeupd512_mask ((double *)__P, (__v8df) __A, (__mmask8) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS512
 _mm512_storeu_pd(void *__P, __m512d __A)
 {
   struct __storeu_pd {
-    __m512d __v;
+    __m512d_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_pd*)__P)->__v = __A;
 }
 
 static __inline void __DEFAULT_FN_ATTRS512
 _mm512_mask_storeu_ps(void *__P, __mmask16 __U, __m512 __A)
 {
   __builtin_ia32_storeups512_mask ((float *)__P, (__v16sf) __A,
                                    (__mmask16) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS512
 _mm512_storeu_ps(void *__P, __m512 __A)
 {
   struct __storeu_ps {
-    __m512 __v;
+    __m512_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_ps*)__P)->__v = __A;
 }
 
 static __inline void __DEFAULT_FN_ATTRS512
 _mm512_mask_store_pd(void *__P, __mmask8 __U, __m512d __A)
 {
   __builtin_ia32_storeapd512_mask ((__v8df *)__P, (__v8df) __A, (__mmask8) __U);

lib/Headers/avx512vlbwintrin.h

return (__m256i) __builtin_ia32_selectb_256(__M,
                                               (__v32qi) _mm256_set1_epi8(__A),
                                               (__v32qi) _mm256_setzero_si256());
 }
 
 static __inline __m128i __DEFAULT_FN_ATTRS128
 _mm_loadu_epi16 (void const *__P)
 {
   struct __loadu_epi16 {
-    __m128i __v;
+    __m128i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_epi16*)__P)->__v;
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_loadu_epi16 (__m128i __W, __mmask8 __U, void const *__P)
 {
   return (__m128i) __builtin_ia32_loaddquhi128_mask ((__v8hi *) __P,
return (__m128i) __builtin_ia32_loaddquhi128_mask ((__v8hi *) __P,
                  _mm_setzero_si128 (),
                  (__mmask8) __U);
 }
 
 static __inline __m256i __DEFAULT_FN_ATTRS256
 _mm256_loadu_epi16 (void const *__P)
 {
   struct __loadu_epi16 {
-    __m256i __v;
+    __m256i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_epi16*)__P)->__v;
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_mask_loadu_epi16 (__m256i __W, __mmask16 __U, void const *__P)
 {
   return (__m256i) __builtin_ia32_loaddquhi256_mask ((__v16hi *) __P,
return (__m256i) __builtin_ia32_loaddquhi256_mask ((__v16hi *) __P,
                  _mm256_setzero_si256 (),
                  (__mmask16) __U);
 }
 
 static __inline __m128i __DEFAULT_FN_ATTRS128
 _mm_loadu_epi8 (void const *__P)
 {
   struct __loadu_epi8 {
-    __m128i __v;
+    __m128i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_epi8*)__P)->__v;
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_loadu_epi8 (__m128i __W, __mmask16 __U, void const *__P)
 {
   return (__m128i) __builtin_ia32_loaddquqi128_mask ((__v16qi *) __P,
return (__m128i) __builtin_ia32_loaddquqi128_mask ((__v16qi *) __P,
                  _mm_setzero_si128 (),
                  (__mmask16) __U);
 }
 
 static __inline __m256i __DEFAULT_FN_ATTRS256
 _mm256_loadu_epi8 (void const *__P)
 {
   struct __loadu_epi8 {
-    __m256i __v;
+    __m256i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_epi8*)__P)->__v;
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_mask_loadu_epi8 (__m256i __W, __mmask32 __U, void const *__P)
 {
   return (__m256i) __builtin_ia32_loaddquqi256_mask ((__v32qi *) __P,
return (__m256i) __builtin_ia32_loaddquqi256_mask ((__v32qi *) __P,
                  _mm256_setzero_si256 (),
                  (__mmask32) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS128
 _mm_storeu_epi16 (void *__P, __m128i __A)
 {
   struct __storeu_epi16 {
-    __m128i __v;
+    __m128i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_epi16*)__P)->__v = __A;
 }
 
 static __inline__ void __DEFAULT_FN_ATTRS128
 _mm_mask_storeu_epi16 (void *__P, __mmask8 __U, __m128i __A)
 {
   __builtin_ia32_storedquhi128_mask ((__v8hi *) __P,
              (__v8hi) __A,
              (__mmask8) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS256
 _mm256_storeu_epi16 (void *__P, __m256i __A)
 {
   struct __storeu_epi16 {
-    __m256i __v;
+    __m256i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_epi16*)__P)->__v = __A;
 }
 
 static __inline__ void __DEFAULT_FN_ATTRS256
 _mm256_mask_storeu_epi16 (void *__P, __mmask16 __U, __m256i __A)
 {
   __builtin_ia32_storedquhi256_mask ((__v16hi *) __P,
              (__v16hi) __A,
              (__mmask16) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS128
 _mm_storeu_epi8 (void *__P, __m128i __A)
 {
   struct __storeu_epi8 {
-    __m128i __v;
+    __m128i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_epi8*)__P)->__v = __A;
 }
 
 static __inline__ void __DEFAULT_FN_ATTRS128
 _mm_mask_storeu_epi8 (void *__P, __mmask16 __U, __m128i __A)
 {
   __builtin_ia32_storedquqi128_mask ((__v16qi *) __P,
              (__v16qi) __A,
              (__mmask16) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS256
 _mm256_storeu_epi8 (void *__P, __m256i __A)
 {
   struct __storeu_epi8 {
-    __m256i __v;
+    __m256i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_epi8*)__P)->__v = __A;
 }
 
 static __inline__ void __DEFAULT_FN_ATTRS256
 _mm256_mask_storeu_epi8 (void *__P, __mmask32 __U, __m256i __A)
 {
   __builtin_ia32_storedquqi256_mask ((__v32qi *) __P,

lib/Headers/avx512vlintrin.h

return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P,
               _mm256_setzero_ps (),
               (__mmask8) __U);
 }
 
 static __inline __m128i __DEFAULT_FN_ATTRS128
 _mm_loadu_epi64 (void const *__P)
 {
   struct __loadu_epi64 {
-    __m128i __v;
+    __m128i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_epi64*)__P)->__v;
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_loadu_epi64 (__m128i __W, __mmask8 __U, void const *__P)
 {
   return (__m128i) __builtin_ia32_loaddqudi128_mask ((__v2di *) __P,
return (__m128i) __builtin_ia32_loaddqudi128_mask ((__v2di *) __P,
                  _mm_setzero_si128 (),
                  (__mmask8) __U);
 }
 
 static __inline __m256i __DEFAULT_FN_ATTRS256
 _mm256_loadu_epi64 (void const *__P)
 {
   struct __loadu_epi64 {
-    __m256i __v;
+    __m256i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_epi64*)__P)->__v;
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_mask_loadu_epi64 (__m256i __W, __mmask8 __U, void const *__P)
 {
   return (__m256i) __builtin_ia32_loaddqudi256_mask ((__v4di *) __P,
return (__m256i) __builtin_ia32_loaddqudi256_mask ((__v4di *) __P,
                  _mm256_setzero_si256 (),
                  (__mmask8) __U);
 }
 
 static __inline __m128i __DEFAULT_FN_ATTRS128
 _mm_loadu_epi32 (void const *__P)
 {
   struct __loadu_epi32 {
-    __m128i __v;
+    __m128i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_epi32*)__P)->__v;
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_loadu_epi32 (__m128i __W, __mmask8 __U, void const *__P)
 {
   return (__m128i) __builtin_ia32_loaddqusi128_mask ((__v4si *) __P,
return (__m128i) __builtin_ia32_loaddqusi128_mask ((__v4si *) __P,
                  _mm_setzero_si128 (),
                  (__mmask8) __U);
 }
 
 static __inline __m256i __DEFAULT_FN_ATTRS256
 _mm256_loadu_epi32 (void const *__P)
 {
   struct __loadu_epi32 {
-    __m256i __v;
+    __m256i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_epi32*)__P)->__v;
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_mask_loadu_epi32 (__m256i __W, __mmask8 __U, void const *__P)
 {
   return (__m256i) __builtin_ia32_loaddqusi256_mask ((__v8si *) __P,
__builtin_ia32_storeaps256_mask ((__v8sf *) __P,
            (__v8sf) __A,
            (__mmask8) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS128
 _mm_storeu_epi64 (void *__P, __m128i __A)
 {
   struct __storeu_epi64 {
-    __m128i __v;
+    __m128i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_epi64*)__P)->__v = __A;
 }
 
 static __inline__ void __DEFAULT_FN_ATTRS128
 _mm_mask_storeu_epi64 (void *__P, __mmask8 __U, __m128i __A)
 {
   __builtin_ia32_storedqudi128_mask ((__v2di *) __P,
              (__v2di) __A,
              (__mmask8) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS256
 _mm256_storeu_epi64 (void *__P, __m256i __A)
 {
   struct __storeu_epi64 {
-    __m256i __v;
+    __m256i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_epi64*)__P)->__v = __A;
 }
 
 static __inline__ void __DEFAULT_FN_ATTRS256
 _mm256_mask_storeu_epi64 (void *__P, __mmask8 __U, __m256i __A)
 {
   __builtin_ia32_storedqudi256_mask ((__v4di *) __P,
              (__v4di) __A,
              (__mmask8) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS128
 _mm_storeu_epi32 (void *__P, __m128i __A)
 {
   struct __storeu_epi32 {
-    __m128i __v;
+    __m128i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_epi32*)__P)->__v = __A;
 }
 
 static __inline__ void __DEFAULT_FN_ATTRS128
 _mm_mask_storeu_epi32 (void *__P, __mmask8 __U, __m128i __A)
 {
   __builtin_ia32_storedqusi128_mask ((__v4si *) __P,
              (__v4si) __A,
              (__mmask8) __U);
 }
 
 static __inline void __DEFAULT_FN_ATTRS256
 _mm256_storeu_epi32 (void *__P, __m256i __A)
 {
   struct __storeu_epi32 {
-    __m256i __v;
+    __m256i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_epi32*)__P)->__v = __A;
 }
 
 static __inline__ void __DEFAULT_FN_ATTRS256
 _mm256_mask_storeu_epi32 (void *__P, __mmask8 __U, __m256i __A)
 {
   __builtin_ia32_storedqusi256_mask ((__v8si *) __P,

lib/Headers/avxintrin.h

 typedef unsigned int __v8su __attribute__ ((__vector_size__ (32)));
 typedef unsigned short __v16hu __attribute__ ((__vector_size__ (32)));
 typedef unsigned char __v32qu __attribute__ ((__vector_size__ (32)));
 
 /* We need an explicitly signed variant for char. Note that this shouldn't
  * appear in the interface though. */
 typedef signed char __v32qs __attribute__((__vector_size__(32)));
 
-typedef float __m256 __attribute__ ((__vector_size__ (32)));
-typedef double __m256d __attribute__((__vector_size__(32)));
-typedef long long __m256i __attribute__((__vector_size__(32)));
+typedef float __m256 __attribute__ ((__vector_size__ (32), __aligned__(32)));
+typedef double __m256d __attribute__((__vector_size__(32), __aligned__(32)));
+typedef long long __m256i __attribute__((__vector_size__(32), __aligned__(32)));
+
+typedef float __m256_u __attribute__ ((__vector_size__ (32), __aligned__(1)));
+typedef double __m256d_u __attribute__((__vector_size__(32), __aligned__(1)));
+typedef long long __m256i_u __attribute__((__vector_size__(32), __aligned__(1)));
 
 /* Define the default attributes for the functions in this file. */
 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx"), __min_vector_width__(256)))
 #define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx"), __min_vector_width__(128)))
 
 /* Arithmetic */
 /// Adds two 256-bit vectors of [4 x double].
 ///

 /// \param __p
 ///    A pointer to a memory location containing double-precision floating
 ///    point values.
 /// \returns A 256-bit vector of [4 x double] containing the moved values.
 static __inline __m256d __DEFAULT_FN_ATTRS
 _mm256_loadu_pd(double const *__p)
 {
   struct __loadu_pd {
-    __m256d __v;
+    __m256d_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_pd*)__p)->__v;
 }
 
 /// Loads 8 single-precision floating point values from an unaligned
 ///    memory location pointed to by \a __p into a vector of [8 x float].
 ///
 /// \headerfile <x86intrin.h>
 ///
 /// This intrinsic corresponds to the <c> VMOVUPS </c> instruction.
 ///
 /// \param __p
 ///    A pointer to a memory location containing single-precision floating
 ///    point values.
 /// \returns A 256-bit vector of [8 x float] containing the moved values.
 static __inline __m256 __DEFAULT_FN_ATTRS
 _mm256_loadu_ps(float const *__p)
 {
   struct __loadu_ps {
-    __m256 __v;
+    __m256_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_ps*)__p)->__v;
 }
 
 /// Loads 256 bits of integer data from a 32-byte aligned memory
 ///    location pointed to by \a __p into elements of a 256-bit integer vector.
 ///
 /// \headerfile <x86intrin.h>

 ///
 /// \param __p
 ///    A pointer to a 256-bit integer vector containing integer values.
 /// \returns A 256-bit integer vector containing the moved values.
 static __inline __m256i __DEFAULT_FN_ATTRS
 _mm256_loadu_si256(__m256i const *__p)
 {
   struct __loadu_si256 {
-    __m256i __v;
+    __m256i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_si256*)__p)->__v;
 }
 
 /// Loads 256 bits of integer data from an unaligned memory location
 ///    pointed to by \a __p into a 256-bit integer vector. This intrinsic may
 ///    perform better than \c _mm256_loadu_si256 when the data crosses a cache
 ///    line boundary.

 ///    A pointer to a memory location that will receive the double-precision
 ///    floating point values.
 /// \param __a
 ///    A 256-bit vector of [4 x double] containing the values to be moved.
 static __inline void __DEFAULT_FN_ATTRS
 _mm256_storeu_pd(double *__p, __m256d __a)
 {
   struct __storeu_pd {
-    __m256d __v;
+    __m256d_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_pd*)__p)->__v = __a;
 }
 
 /// Stores single-precision floating point values from a 256-bit vector
 ///    of [8 x float] to an unaligned memory location pointed to by \a __p.
 ///
 /// \headerfile <x86intrin.h>
 ///
 /// This intrinsic corresponds to the <c> VMOVUPS </c> instruction.
 ///
 /// \param __p
 ///    A pointer to a memory location that will receive the float values.
 /// \param __a
 ///    A 256-bit vector of [8 x float] containing the values to be moved.
 static __inline void __DEFAULT_FN_ATTRS
 _mm256_storeu_ps(float *__p, __m256 __a)
 {
   struct __storeu_ps {
-    __m256 __v;
+    __m256_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_ps*)__p)->__v = __a;
 }
 
 /// Stores integer values from a 256-bit integer vector to a 32-byte
 ///    aligned memory location pointed to by \a __p.
 ///
 /// \headerfile <x86intrin.h>

 /// \param __p
 ///    A pointer to a memory location that will receive the integer values.
 /// \param __a
 ///    A 256-bit integer vector containing the values to be moved.
 static __inline void __DEFAULT_FN_ATTRS
 _mm256_storeu_si256(__m256i *__p, __m256i __a)
 {
   struct __storeu_si256 {
-    __m256i __v;
+    __m256i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_si256*)__p)->__v = __a;
 }
 
 /* Conditional load ops */
 /// Conditionally loads double-precision floating point elements from a
 ///    memory location pointed to by \a __p into a 128-bit vector of
 ///    [2 x double], depending on the mask bits associated with each data

lib/Headers/emmintrin.h

  *===-----------------------------------------------------------------------===
  */
 
 #ifndef __EMMINTRIN_H
 #define __EMMINTRIN_H
 
 #include <xmmintrin.h>
 
-typedef double __m128d __attribute__((__vector_size__(16)));
-typedef long long __m128i __attribute__((__vector_size__(16)));
+typedef double __m128d __attribute__((__vector_size__(16), __aligned__(16)));
+typedef long long __m128i __attribute__((__vector_size__(16), __aligned__(16)));
+
+typedef double __m128d_u __attribute__((__vector_size__(16), __aligned__(1)));
+typedef long long __m128i_u __attribute__((__vector_size__(16), __aligned__(1)));
 
 /* Type defines.  */
 typedef double __v2df __attribute__ ((__vector_size__ (16)));
 typedef long long __v2di __attribute__ ((__vector_size__ (16)));
 typedef short __v8hi __attribute__((__vector_size__(16)));
 typedef char __v16qi __attribute__((__vector_size__(16)));
 
 /* Unsigned types */

 /// \param __dp
 ///    A pointer to a 128-bit memory location. The address of the memory
 ///    location does not have to be aligned.
 /// \returns A 128-bit vector of [2 x double] containing the loaded values.
 static __inline__ __m128d __DEFAULT_FN_ATTRS
 _mm_loadu_pd(double const *__dp)
 {
   struct __loadu_pd {
-    __m128d __v;
+    __m128d_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_pd*)__dp)->__v;
 }
 
 /// Loads a 64-bit integer value to the low element of a 128-bit integer
 ///    vector and clears the upper element.
 ///
 /// \headerfile <x86intrin.h>

 ///    A pointer to a 128-bit memory location. The address of the memory
 ///    location does not have to be aligned.
 /// \param __a
 ///    A 128-bit vector of [2 x double] containing the values to be stored.
 static __inline__ void __DEFAULT_FN_ATTRS
 _mm_storeu_pd(double *__dp, __m128d __a)
 {
   struct __storeu_pd {
-    __m128d __v;
+    __m128d_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_pd*)__dp)->__v = __a;
 }
 
 /// Stores two double-precision values, in reverse order, from a 128-bit
 ///    vector of [2 x double] to a 16-byte aligned memory location.
 ///
 /// \headerfile <x86intrin.h>

 ///
 /// \param __p
 ///    A pointer to a memory location containing integer values.
 /// \returns A 128-bit integer vector containing the moved values.
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_loadu_si128(__m128i const *__p)
 {
   struct __loadu_si128 {
-    __m128i __v;
+    __m128i_u __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_si128*)__p)->__v;
 }
 
 /// Returns a vector of [2 x i64] where the lower element is taken from
 ///    the lower element of the operand, and the upper element is zero.
 ///
 /// \headerfile <x86intrin.h>

 /// \param __p
 ///    A pointer to a memory location that will receive the integer values.
 /// \param __b
 ///    A 128-bit integer vector containing the values to be moved.
 static __inline__ void __DEFAULT_FN_ATTRS
 _mm_storeu_si128(__m128i *__p, __m128i __b)
 {
   struct __storeu_si128 {
-    __m128i __v;
+    __m128i_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_si128*)__p)->__v = __b;
 }
 
 /// Stores a 64-bit integer value from the low element of a 128-bit integer
 ///    vector.
 ///
 /// \headerfile <x86intrin.h>

lib/Headers/xmmintrin.h

 
 #ifndef __XMMINTRIN_H
 #define __XMMINTRIN_H
 
 #include <mmintrin.h>
 
 typedef int __v4si __attribute__((__vector_size__(16)));
 typedef float __v4sf __attribute__((__vector_size__(16)));
-typedef float __m128 __attribute__((__vector_size__(16)));
+typedef float __m128 __attribute__((__vector_size__(16), __aligned__(16)));
+
+typedef float __m128_u __attribute__((__vector_size__(16), __aligned__(1)));
 
 /* Unsigned types */
 typedef unsigned int __v4su __attribute__((__vector_size__(16)));
 
 /* This header should only be included in a hosted environment as it depends on
  * a standard library to provide allocation routines. */
 #if __STDC_HOSTED__
 #include <mm_malloc.h>

 ///
 /// This intrinsic corresponds to the <c> VMOVUPS / MOVUPS </c> instruction.
 ///
 /// \param __p
 ///    A pointer to a 128-bit memory location. The address of the memory
 ///    location does not have to be aligned.
 /// \returns A 128-bit vector of [4 x float] containing the loaded values.
 static __inline__ __m128 __DEFAULT_FN_ATTRS
 _mm_loadu_ps(const float *__p)
 {
   struct __loadu_ps {
     __m128 __v;
   } __attribute__((__packed__, __may_alias__));
   return ((struct __loadu_ps*)__p)->__v;
 }
 
 /// Loads four packed float values, in reverse order, from an aligned

 ///    A pointer to a 128-bit memory location. The address of the memory
 ///    location does not have to be aligned.
 /// \param __a
 ///    A 128-bit vector of [4 x float] containing the values to be stored.
 static __inline__ void __DEFAULT_FN_ATTRS
 _mm_storeu_ps(float *__p, __m128 __a)
 {
   struct __storeu_ps {
-    __m128 __v;
+    __m128_u __v;
   } __attribute__((__packed__, __may_alias__));
   ((struct __storeu_ps*)__p)->__v = __a;
 }
 
 /// Stores a 128-bit vector of [4 x float] into an aligned memory
 ///    location.
 ///
 /// \headerfile <x86intrin.h>

test/CodeGen/x86-vec-struct-packing.c

+// RUN: %clang_cc1 -ffreestanding -emit-llvm-only  -triple x86_64-windows-coff -fdump-record-layouts %s | FileCheck %s --check-prefix=CHECK-MS
+// RUN: %clang_cc1 -ffreestanding -emit-llvm-only  -triple x86_64-apple-darwin -fdump-record-layouts %s | FileCheck %s --check-prefix=CHECK-NOTMS
+#include <x86intrin.h>
+
+#pragma pack(1)
+
+struct s_m128 {
+  int a;
+  __m128 b;
+};
+typedef struct s_m128 m128;
+
+#if defined(_WIN32)
+static int a1[(sizeof(m128) == 32) - 1];
+#else
+static int a1[(sizeof(m128) == 20) - 1];
+#endif
+
+struct s_m128i {
+  int a;
+  __m128i b;
+};
+typedef struct s_m128i m128i;
+
+#if defined(_WIN32)
+static int a1[(sizeof(m128i) == 32) - 1];
+#else
+static int a1[(sizeof(m128i) == 20) - 1];
+#endif
+
+struct s_m128d {
+  int a;
+  __m128d b;
+};
+typedef struct s_m128d m128d;
+
+#if defined(_WIN32)
+static int a1[(sizeof(m128d) == 32) - 1];
+#else
+static int a1[(sizeof(m128d) == 20) - 1];
+#endif
+
+struct s_m256 {
+  int a;
+  __m256 b;
+};
+typedef struct s_m256 m256;
+
+#if defined(_WIN32)
+static int a1[(sizeof(m256) == 64) - 1];
+#else
+static int a1[(sizeof(m256) == 36) - 1];
+#endif
+
+struct s_m256i {
+  int a;
+  __m256i b;
+};
+typedef struct s_m256i m256i;
+
+#if defined(_WIN32)
+static int a1[(sizeof(m256i) == 64) - 1];
+#else
+static int a1[(sizeof(m256i) == 36) - 1];
+#endif
+
+struct s_m256d {
+  int a;
+  __m256d b;
+};
+typedef struct s_m256d m256d;
+
+#if defined(_WIN32)
+static int a1[(sizeof(m256d) == 64) - 1];
+#else
+static int a1[(sizeof(m256d) == 36) - 1];
+#endif
+
+struct s_m512 {
+  int a;
+  __m512 b;
+};
+typedef struct s_m512 m512;
+
+#if defined(_WIN32)
+static int a1[(sizeof(m512) == 128) - 1];
+#else
+static int a1[(sizeof(m512) == 68) - 1];
+#endif
+
+struct s_m512i {
+  int a;
+  __m512i b;
+};
+typedef struct s_m512i m512i;
+
+#if defined(_WIN32)
+static int a1[(sizeof(m512i) == 128) - 1];
+#else
+static int a1[(sizeof(m512i) == 68) - 1];
+#endif
+
+struct s_m512d {
+  int a;
+  __m512d b;
+};
+typedef struct s_m512d m512d;
+
+#if defined(_WIN32)
+static int a1[(sizeof(m512d) == 128) - 1];
+#else
+static int a1[(sizeof(m512d) == 68) - 1];
+#endif
+
+// CHECK-MS: *** Dumping AST Record Layout
+// CHECK-MS:          0 | struct s_m128
+// CHECK-MS:          0 |   int a
+// CHECK-MS:         16 |   __m128 b
+// CHECK-MS:            | [sizeof=32, align=16]
+// CHECK-MS: *** Dumping AST Record Layout
+// CHECK-MS:          0 | struct s_m128i
+// CHECK-MS:          0 |   int a
+// CHECK-MS:         16 |   __m128i b
+// CHECK-MS:            | [sizeof=32, align=16]
+// CHECK-MS: *** Dumping AST Record Layout
+// CHECK-MS:          0 | struct s_m128d
+// CHECK-MS:          0 |   int a
+// CHECK-MS:         16 |   __m128d b
+// CHECK-MS:            | [sizeof=32, align=16]
+// CHECK-MS: *** Dumping AST Record Layout
+// CHECK-MS:          0 | struct s_m256
+// CHECK-MS:          0 |   int a
+// CHECK-MS:         32 |   __m256 b
+// CHECK-MS:            | [sizeof=64, align=32]
+// CHECK-MS: *** Dumping AST Record Layout
+// CHECK-MS:          0 | struct s_m256i
+// CHECK-MS:          0 |   int a
+// CHECK-MS:         32 |   __m256i b
+// CHECK-MS:            | [sizeof=64, align=32]
+// CHECK-MS: *** Dumping AST Record Layout
+// CHECK-MS:          0 | struct s_m256d
+// CHECK-MS:          0 |   int a
+// CHECK-MS:         32 |   __m256d b
+// CHECK-MS:            | [sizeof=64, align=32]
+// CHECK-MS: *** Dumping AST Record Layout
+// CHECK-MS:          0 | struct s_m512
+// CHECK-MS:          0 |   int a
+// CHECK-MS:         64 |   __m512 b
+// CHECK-MS:            | [sizeof=128, align=64]
+// CHECK-MS: *** Dumping AST Record Layout
+// CHECK-MS:          0 | struct s_m512i
+// CHECK-MS:          0 |   int a
+// CHECK-MS:         64 |   __m512i b
+// CHECK-MS:            | [sizeof=128, align=64]
+// CHECK-MS: *** Dumping AST Record Layout
+// CHECK-MS:          0 | struct s_m512d
+// CHECK-MS:          0 |   int a
+// CHECK-MS:         64 |   __m512d b
+// CHECK-MS:            | [sizeof=128, align=64]
+
+// CHECK-NOTMS: *** Dumping AST Record Layout
+// CHECK-NOTMS:          0 | struct s_m128
+// CHECK-NOTMS:          0 |   int a
+// CHECK-NOTMS:          4 |   __m128 b
+// CHECK-NOTMS:            | [sizeof=20, align=1]
+// CHECK-NOTMS: *** Dumping AST Record Layout
+// CHECK-NOTMS:          0 | struct s_m128i
+// CHECK-NOTMS:          0 |   int a
+// CHECK-NOTMS:          4 |   __m128i b
+// CHECK-NOTMS:            | [sizeof=20, align=1]
+// CHECK-NOTMS: *** Dumping AST Record Layout
+// CHECK-NOTMS:          0 | struct s_m128d
+// CHECK-NOTMS:          0 |   int a
+// CHECK-NOTMS:          4 |   __m128d b
+// CHECK-NOTMS:            | [sizeof=20, align=1]
+// CHECK-NOTMS: *** Dumping AST Record Layout
+// CHECK-NOTMS:          0 | struct s_m256
+// CHECK-NOTMS:          0 |   int a
+// CHECK-NOTMS:          4 |   __m256 b
+// CHECK-NOTMS:            | [sizeof=36, align=1]
+// CHECK-NOTMS: *** Dumping AST Record Layout
+// CHECK-NOTMS:          0 | struct s_m256i
+// CHECK-NOTMS:          0 |   int a
+// CHECK-NOTMS:          4 |   __m256i b
+// CHECK-NOTMS:            | [sizeof=36, align=1]
+// CHECK-NOTMS: *** Dumping AST Record Layout
+// CHECK-NOTMS:          0 | struct s_m256d
+// CHECK-NOTMS:          0 |   int a
+// CHECK-NOTMS:          4 |   __m256d b
+// CHECK-NOTMS:            | [sizeof=36, align=1]
+// CHECK-NOTMS: *** Dumping AST Record Layout
+// CHECK-NOTMS:          0 | struct s_m512
+// CHECK-NOTMS:          0 |   int a
+// CHECK-NOTMS:          4 |   __m512 b
+// CHECK-NOTMS:            | [sizeof=68, align=1]
+// CHECK-NOTMS: *** Dumping AST Record Layout
+// CHECK-NOTMS:          0 | struct s_m512i
+// CHECK-NOTMS:          0 |   int a
+// CHECK-NOTMS:          4 |   __m512i b
+// CHECK-NOTMS:            | [sizeof=68, align=1]
+// CHECK-NOTMS: *** Dumping AST Record Layout
+// CHECK-NOTMS:          0 | struct s_m512d
+// CHECK-NOTMS:          0 |   int a
+// CHECK-NOTMS:          4 |   __m512d b
+// CHECK-NOTMS:            | [sizeof=68, align=1]