MFC r309124:

Upgrade our copies of clang, llvm, lldb, compiler-rt and libc++ to 3.9.0
release, and add lld 3.9.0.  Also completely revamp the build system for
clang, llvm, lldb and their related tools.

Please note that from 3.5.0 onwards, clang, llvm and lldb require C++11
support to build; see UPDATING for more information.

Release notes for llvm, clang and lld are available here:
<http://llvm.org/releases/3.9.0/docs/ReleaseNotes.html>
<http://llvm.org/releases/3.9.0/tools/clang/docs/ReleaseNotes.html>
<http://llvm.org/releases/3.9.0/tools/lld/docs/ReleaseNotes.html>

Thanks to Ed Maste, Bryan Drewery, Andrew Turner, Antoine Brodin and Jan
Beich for their help.

Relnotes:	yes

MFC r309147:

Pull in r282174 from upstream llvm trunk (by Krzysztof Parzyszek):

  [PPC] Set SP after loading data from stack frame, if no red zone is
  present

  Follow-up to r280705: Make sure that the SP is only restored after
  all data is loaded from the stack frame, if there is no red zone.

  This completes the fix for
  https://llvm.org/bugs/show_bug.cgi?id=26519.

  Differential Revision: https://reviews.llvm.org/D24466

Reported by:    Mark Millard
PR:             214433

MFC r309149:

Pull in r283060 from upstream llvm trunk (by Hal Finkel):

  [PowerPC] Refactor soft-float support, and enable PPC64 soft float

  This change enables soft-float for PowerPC64, and also makes
  soft-float disable all vector instruction sets for both 32-bit and
  64-bit modes. This latter part is necessary because the PPC backend
  canonicalizes many Altivec vector types to floating-point types, and
  so soft-float breaks scalarization support for many operations. Both
  for embedded targets and for operating-system kernels desiring
  soft-float support, it seems reasonable that disabling hardware
  floating-point also disables vector instructions (embedded targets
  without hardware floating point support are unlikely to have Altivec,
  etc. and operating system kernels desiring not to use floating-point
  registers to lower syscall cost are unlikely to want to use vector
  registers either). If someone needs this to work, we'll need to
  change the fact that we promote many Altivec operations to act on
  v4f32. To make it possible to disable Altivec when soft-float is
  enabled, hardware floating-point support needs to be expressed as a
  positive feature, like the others, and not a negative feature,
  because target features cannot have dependencies on the disabling of
  some other feature. So +soft-float has now become -hard-float.

  Fixes PR26970.

Pull in r283061 from upstream clang trunk (by Hal Finkel):

  [PowerPC] Enable soft-float for PPC64, and +soft-float -> -hard-float

  Enable soft-float support on PPC64, as the backend now supports it.
  Also, the backend now uses -hard-float instead of +soft-float, so set
  the target features accordingly.

  Fixes PR26970.

Reported by:	Mark Millard
PR:		214433

MFC r309212:

Add a few missed clang 3.9.0 files to OptionalObsoleteFiles.

MFC r309262:

Fix packaging for clang, lldb and lld 3.9.0

During the upgrade of clang/llvm etc to 3.9.0 in r309124, the PACKAGE
directive in the usr.bin/clang/*.mk files got dropped accidentally.

Restore it, with a few minor changes and additions:
* Correct license in clang.ucl to NCSA
* Add PACKAGE=clang for clang and most of the "ll" tools
* Put lldb in its own package
* Put lld in its own package

Reviewed by:	gjb, jmallett
Differential Revision: https://reviews.freebsd.org/D8666

MFC r309656:

During the bootstrap phase, when building the minimal llvm library on
PowerPC, add lib/Support/Atomic.cpp.  This is needed because upstream
llvm revision r271821 disabled the use of std::call_once, which causes
some fallback functions from Atomic.cpp to be used instead.

Reported by:	Mark Millard
PR:		214902

MFC r309835:

Tentatively apply https://reviews.llvm.org/D18730 to work around gcc PR
70528 (bogus error: constructor required before non-static data member).
This should fix buildworld with the external gcc package.

Reported by:	https://jenkins.freebsd.org/job/FreeBSD_HEAD_amd64_gcc/

MFC r310194:

Upgrade our copies of clang, llvm, lld, lldb, compiler-rt and libc++ to
3.9.1 release.

Please note that from 3.5.0 onwards, clang, llvm and lldb require C++11
support to build; see UPDATING for more information.

Release notes for llvm, clang and lld will be available here:
<http://releases.llvm.org/3.9.1/docs/ReleaseNotes.html>
<http://releases.llvm.org/3.9.1/tools/clang/docs/ReleaseNotes.html>
<http://releases.llvm.org/3.9.1/tools/lld/docs/ReleaseNotes.html>

Relnotes:	yes

1 files changed, 5019 insertions, 455 deletions

diff --git a/contrib/llvm/tools/clang/lib/Headers/avx512vlintrin.h b/contrib/llvm/tools/clang/lib/Headers/avx512vlintrin.h
index 8f13536..295ce29 100644
--- a/contrib/llvm/tools/clang/lib/Headers/avx512vlintrin.h
+++ b/contrib/llvm/tools/clang/lib/Headers/avx512vlintrin.h
@@ -29,17 +29,22 @@
 #define __AVX512VLINTRIN_H
 
 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vl")))
-#define __DEFAULT_FN_ATTRS_BOTH __attribute__((__always_inline__, __nodebug__, __target__("avx512vl, avx512bw")))
+
+/* Doesn't require avx512vl, used in avx512dqintrin.h */
+static  __inline __m128i __attribute__((__always_inline__, __nodebug__, __target__("avx512f")))
+_mm_setzero_di(void) {
+  return (__m128i)(__v2di){ 0LL, 0LL};
+}
 
 /* Integer compare */
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm_cmpeq_epi32_mask(__m128i __a, __m128i __b) {
   return (__mmask8)__builtin_ia32_pcmpeqd128_mask((__v4si)__a, (__v4si)__b,
                                                   (__mmask8)-1);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm_mask_cmpeq_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
   return (__mmask8)__builtin_ia32_pcmpeqd128_mask((__v4si)__a, (__v4si)__b,
                                                   __u);
@@ -57,13 +62,13 @@ _mm_mask_cmpeq_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
                                                 __u);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm256_cmpeq_epi32_mask(__m256i __a, __m256i __b) {
   return (__mmask8)__builtin_ia32_pcmpeqd256_mask((__v8si)__a, (__v8si)__b,
                                                   (__mmask8)-1);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm256_mask_cmpeq_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
   return (__mmask8)__builtin_ia32_pcmpeqd256_mask((__v8si)__a, (__v8si)__b,
                                                   __u);
@@ -81,13 +86,13 @@ _mm256_mask_cmpeq_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
                                                 __u);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm_cmpeq_epi64_mask(__m128i __a, __m128i __b) {
   return (__mmask8)__builtin_ia32_pcmpeqq128_mask((__v2di)__a, (__v2di)__b,
                                                   (__mmask8)-1);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm_mask_cmpeq_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
   return (__mmask8)__builtin_ia32_pcmpeqq128_mask((__v2di)__a, (__v2di)__b,
                                                   __u);
@@ -105,13 +110,13 @@ _mm_mask_cmpeq_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
                                                 __u);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm256_cmpeq_epi64_mask(__m256i __a, __m256i __b) {
   return (__mmask8)__builtin_ia32_pcmpeqq256_mask((__v4di)__a, (__v4di)__b,
                                                   (__mmask8)-1);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm256_mask_cmpeq_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
   return (__mmask8)__builtin_ia32_pcmpeqq256_mask((__v4di)__a, (__v4di)__b,
                                                   __u);
@@ -226,13 +231,13 @@ _mm256_mask_cmpge_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
                                                 __u);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm_cmpgt_epi32_mask(__m128i __a, __m128i __b) {
   return (__mmask8)__builtin_ia32_pcmpgtd128_mask((__v4si)__a, (__v4si)__b,
                                                   (__mmask8)-1);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm_mask_cmpgt_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
   return (__mmask8)__builtin_ia32_pcmpgtd128_mask((__v4si)__a, (__v4si)__b,
                                                   __u);
@@ -250,13 +255,13 @@ _mm_mask_cmpgt_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) {
                                                 __u);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm256_cmpgt_epi32_mask(__m256i __a, __m256i __b) {
   return (__mmask8)__builtin_ia32_pcmpgtd256_mask((__v8si)__a, (__v8si)__b,
                                                   (__mmask8)-1);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm256_mask_cmpgt_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
   return (__mmask8)__builtin_ia32_pcmpgtd256_mask((__v8si)__a, (__v8si)__b,
                                                   __u);
@@ -274,13 +279,13 @@ _mm256_mask_cmpgt_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) {
                                                 __u);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm_cmpgt_epi64_mask(__m128i __a, __m128i __b) {
   return (__mmask8)__builtin_ia32_pcmpgtq128_mask((__v2di)__a, (__v2di)__b,
                                                   (__mmask8)-1);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm_mask_cmpgt_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
   return (__mmask8)__builtin_ia32_pcmpgtq128_mask((__v2di)__a, (__v2di)__b,
                                                   __u);
@@ -298,13 +303,13 @@ _mm_mask_cmpgt_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) {
                                                 __u);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm256_cmpgt_epi64_mask(__m256i __a, __m256i __b) {
   return (__mmask8)__builtin_ia32_pcmpgtq256_mask((__v4di)__a, (__v4di)__b,
                                                   (__mmask8)-1);
 }
 
-static __inline__ __mmask8 __DEFAULT_FN_ATTRS_BOTH
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 _mm256_mask_cmpgt_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) {
   return (__mmask8)__builtin_ia32_pcmpgtq256_mask((__v4di)__a, (__v4di)__b,
                                                   __u);
@@ -885,437 +890,352 @@ _mm_mask_mullo_epi32 (__m128i __W, __mmask16 __M, __m128i __A,
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_and_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
-           __m256i __B)
+_mm256_mask_and_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pandd256_mask ((__v8si) __A,
-             (__v8si) __B,
-             (__v8si) __W,
-             (__mmask8) __U);
+  return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
+                                             (__v8si)_mm256_and_si256(__A, __B),
+                                             (__v8si)__W);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_and_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
+_mm256_maskz_and_epi32(__mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pandd256_mask ((__v8si) __A,
-             (__v8si) __B,
-             (__v8si)
-             _mm256_setzero_si256 (),
-             (__mmask8) __U);
+  return (__m256i)_mm256_mask_and_epi32(_mm256_setzero_si256(), __U, __A, __B);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_and_epi32 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
+_mm_mask_and_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pandd128_mask ((__v4si) __A,
-             (__v4si) __B,
-             (__v4si) __W,
-             (__mmask8) __U);
+  return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
+                                             (__v4si)_mm_and_si128(__A, __B),
+                                             (__v4si)__W);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_maskz_and_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
+_mm_maskz_and_epi32(__mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pandd128_mask ((__v4si) __A,
-             (__v4si) __B,
-             (__v4si)
-             _mm_setzero_si128 (),
-             (__mmask8) __U);
+  return (__m128i)_mm_mask_and_epi32(_mm_setzero_si128(), __U, __A, __B);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_andnot_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
-        __m256i __B)
+_mm256_mask_andnot_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pandnd256_mask ((__v8si) __A,
-              (__v8si) __B,
-              (__v8si) __W,
-              (__mmask8) __U);
+  return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
+                                          (__v8si)_mm256_andnot_si256(__A, __B),
+                                          (__v8si)__W);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_andnot_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
+_mm256_maskz_andnot_epi32(__mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pandnd256_mask ((__v8si) __A,
-              (__v8si) __B,
-              (__v8si)
-              _mm256_setzero_si256 (),
-              (__mmask8) __U);
+  return (__m256i)_mm256_mask_andnot_epi32(_mm256_setzero_si256(),
+                                           __U, __A, __B);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_andnot_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
-           __m128i __B)
+_mm_mask_andnot_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pandnd128_mask ((__v4si) __A,
-              (__v4si) __B,
-              (__v4si) __W,
-              (__mmask8) __U);
+  return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
+                                             (__v4si)_mm_andnot_si128(__A, __B),
+                                             (__v4si)__W);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_maskz_andnot_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pandnd128_mask ((__v4si) __A,
-              (__v4si) __B,
-              (__v4si)
-              _mm_setzero_si128 (),
-              (__mmask8) __U);
+  return (__m128i)_mm_mask_andnot_epi32(_mm_setzero_si128(), __U, __A, __B);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_or_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
-          __m256i __B)
+_mm256_mask_or_epi32 (__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pord256_mask ((__v8si) __A,
-            (__v8si) __B,
-            (__v8si) __W,
-            (__mmask8) __U);
+  return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
+                                             (__v8si)_mm256_or_si256(__A, __B),
+                                             (__v8si)__W);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_or_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
+_mm256_maskz_or_epi32(__mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pord256_mask ((__v8si) __A,
-            (__v8si) __B,
-            (__v8si)
-            _mm256_setzero_si256 (),
-            (__mmask8) __U);
+  return (__m256i)_mm256_mask_or_epi32(_mm256_setzero_si256(), __U, __A, __B);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_or_epi32 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
+_mm_mask_or_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pord128_mask ((__v4si) __A,
-            (__v4si) __B,
-            (__v4si) __W,
-            (__mmask8) __U);
+  return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
+                                             (__v4si)_mm_or_si128(__A, __B),
+                                             (__v4si)__W);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_maskz_or_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
+_mm_maskz_or_epi32(__mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pord128_mask ((__v4si) __A,
-            (__v4si) __B,
-            (__v4si)
-            _mm_setzero_si128 (),
-            (__mmask8) __U);
+  return (__m128i)_mm_mask_or_epi32(_mm_setzero_si128(), __U, __A, __B);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_xor_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
-           __m256i __B)
+_mm256_mask_xor_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pxord256_mask ((__v8si) __A,
-             (__v8si) __B,
-             (__v8si) __W,
-             (__mmask8) __U);
+  return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
+                                             (__v8si)_mm256_xor_si256(__A, __B),
+                                             (__v8si)__W);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_xor_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
+_mm256_maskz_xor_epi32(__mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pxord256_mask ((__v8si) __A,
-             (__v8si) __B,
-             (__v8si)
-             _mm256_setzero_si256 (),
-             (__mmask8) __U);
+  return (__m256i)_mm256_mask_xor_epi32(_mm256_setzero_si256(), __U, __A, __B);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_xor_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
+_mm_mask_xor_epi32(__m128i __W, __mmask8 __U, __m128i __A,
         __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pxord128_mask ((__v4si) __A,
-             (__v4si) __B,
-             (__v4si) __W,
-             (__mmask8) __U);
+  return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
+                                             (__v4si)_mm_xor_si128(__A, __B),
+                                             (__v4si)__W);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_maskz_xor_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
+_mm_maskz_xor_epi32(__mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pxord128_mask ((__v4si) __A,
-             (__v4si) __B,
-             (__v4si)
-             _mm_setzero_si128 (),
-             (__mmask8) __U);
+  return (__m128i)_mm_mask_xor_epi32(_mm_setzero_si128(), __U, __A, __B);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_and_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
-           __m256i __B)
+_mm256_mask_and_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pandq256_mask ((__v4di) __A,
-             (__v4di) __B,
-             (__v4di) __W, __U);
+  return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
+                                             (__v4di)_mm256_and_si256(__A, __B),
+                                             (__v4di)__W);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_and_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
+_mm256_maskz_and_epi64(__mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pandq256_mask ((__v4di) __A,
-             (__v4di) __B,
-             (__v4di)
-             _mm256_setzero_pd (),
-             __U);
+  return (__m256i)_mm256_mask_and_epi64(_mm256_setzero_si256(), __U, __A, __B);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_and_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
-        __m128i __B)
+_mm_mask_and_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pandq128_mask ((__v2di) __A,
-             (__v2di) __B,
-             (__v2di) __W, __U);
+  return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
+                                             (__v2di)_mm_and_si128(__A, __B),
+                                             (__v2di)__W);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_maskz_and_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
+_mm_maskz_and_epi64(__mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pandq128_mask ((__v2di) __A,
-             (__v2di) __B,
-             (__v2di)
-             _mm_setzero_pd (),
-             __U);
+  return (__m128i)_mm_mask_and_epi64(_mm_setzero_si128(), __U, __A, __B);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_andnot_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
-        __m256i __B)
+_mm256_mask_andnot_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pandnq256_mask ((__v4di) __A,
-              (__v4di) __B,
-              (__v4di) __W, __U);
+  return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
+                                          (__v4di)_mm256_andnot_si256(__A, __B),
+                                          (__v4di)__W);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_andnot_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
+_mm256_maskz_andnot_epi64(__mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pandnq256_mask ((__v4di) __A,
-              (__v4di) __B,
-              (__v4di)
-              _mm256_setzero_pd (),
-              __U);
+  return (__m256i)_mm256_mask_andnot_epi64(_mm256_setzero_si256(),
+                                           __U, __A, __B);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_andnot_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
-           __m128i __B)
+_mm_mask_andnot_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pandnq128_mask ((__v2di) __A,
-              (__v2di) __B,
-              (__v2di) __W, __U);
+  return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
+                                             (__v2di)_mm_andnot_si128(__A, __B),
+                                             (__v2di)__W);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_maskz_andnot_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
+_mm_maskz_andnot_epi64(__mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pandnq128_mask ((__v2di) __A,
-              (__v2di) __B,
-              (__v2di)
-              _mm_setzero_pd (),
-              __U);
+  return (__m128i)_mm_mask_andnot_epi64(_mm_setzero_si128(), __U, __A, __B);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_or_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
-          __m256i __B)
+_mm256_mask_or_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_porq256_mask ((__v4di) __A,
-            (__v4di) __B,
-            (__v4di) __W,
-            (__mmask8) __U);
+  return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
+                                             (__v4di)_mm256_or_si256(__A, __B),
+                                             (__v4di)__W);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_or_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
+_mm256_maskz_or_epi64(__mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_porq256_mask ((__v4di) __A,
-            (__v4di) __B,
-            (__v4di)
-            _mm256_setzero_si256 (),
-            (__mmask8) __U);
+  return (__m256i)_mm256_mask_or_epi64(_mm256_setzero_si256(), __U, __A, __B);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_or_epi64 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
+_mm_mask_or_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_porq128_mask ((__v2di) __A,
-            (__v2di) __B,
-            (__v2di) __W,
-            (__mmask8) __U);
+  return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
+                                             (__v2di)_mm_or_si128(__A, __B),
+                                             (__v2di)__W);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_maskz_or_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
+_mm_maskz_or_epi64(__mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_porq128_mask ((__v2di) __A,
-            (__v2di) __B,
-            (__v2di)
-            _mm_setzero_si128 (),
-            (__mmask8) __U);
+  return (__m128i)_mm_mask_or_epi64(_mm_setzero_si128(), __U, __A, __B);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_xor_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
-           __m256i __B)
+_mm256_mask_xor_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pxorq256_mask ((__v4di) __A,
-             (__v4di) __B,
-             (__v4di) __W,
-             (__mmask8) __U);
+  return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
+                                             (__v4di)_mm256_xor_si256(__A, __B),
+                                             (__v4di)__W);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_xor_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
+_mm256_maskz_xor_epi64(__mmask8 __U, __m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_pxorq256_mask ((__v4di) __A,
-             (__v4di) __B,
-             (__v4di)
-             _mm256_setzero_si256 (),
-             (__mmask8) __U);
+  return (__m256i)_mm256_mask_xor_epi64(_mm256_setzero_si256(), __U, __A, __B);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_xor_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
+_mm_mask_xor_epi64(__m128i __W, __mmask8 __U, __m128i __A,
         __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pxorq128_mask ((__v2di) __A,
-             (__v2di) __B,
-             (__v2di) __W,
-             (__mmask8) __U);
+  return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
+                                             (__v2di)_mm_xor_si128(__A, __B),
+                                             (__v2di)__W);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_maskz_xor_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
+_mm_maskz_xor_epi64(__mmask8 __U, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_pxorq128_mask ((__v2di) __A,
-             (__v2di) __B,
-             (__v2di)
-             _mm_setzero_si128 (),
-             (__mmask8) __U);
+  return (__m128i)_mm_mask_xor_epi64(_mm_setzero_si128(), __U, __A, __B);
 }
 
 #define _mm_cmp_epi32_mask(a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)(__m128i)(a), \
-                                        (__v4si)(__m128i)(b), \
-                                        (p), (__mmask8)-1); })
+                                        (__v4si)(__m128i)(b), (int)(p), \
+                                        (__mmask8)-1); })
 
 #define _mm_mask_cmp_epi32_mask(m, a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)(__m128i)(a), \
-                                        (__v4si)(__m128i)(b), \
-                                        (p), (__mmask8)(m)); })
+                                        (__v4si)(__m128i)(b), (int)(p), \
+                                        (__mmask8)(m)); })
 
 #define _mm_cmp_epu32_mask(a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)(__m128i)(a), \
-                                         (__v4si)(__m128i)(b), \
-                                         (p), (__mmask8)-1); })
+                                         (__v4si)(__m128i)(b), (int)(p), \
+                                         (__mmask8)-1); })
 
 #define _mm_mask_cmp_epu32_mask(m, a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)(__m128i)(a), \
-                                         (__v4si)(__m128i)(b), \
-                                         (p), (__mmask8)(m)); })
+                                         (__v4si)(__m128i)(b), (int)(p), \
+                                         (__mmask8)(m)); })
 
 #define _mm256_cmp_epi32_mask(a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)(__m256i)(a), \
-                                        (__v8si)(__m256i)(b), \
-                                        (p), (__mmask8)-1); })
+                                        (__v8si)(__m256i)(b), (int)(p), \
+                                        (__mmask8)-1); })
 
 #define _mm256_mask_cmp_epi32_mask(m, a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)(__m256i)(a), \
-                                        (__v8si)(__m256i)(b), \
-                                        (p), (__mmask8)(m)); })
+                                        (__v8si)(__m256i)(b), (int)(p), \
+                                        (__mmask8)(m)); })
 
 #define _mm256_cmp_epu32_mask(a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)(__m256i)(a), \
-                                         (__v8si)(__m256i)(b), \
-                                         (p), (__mmask8)-1); })
+                                         (__v8si)(__m256i)(b), (int)(p), \
+                                         (__mmask8)-1); })
 
 #define _mm256_mask_cmp_epu32_mask(m, a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)(__m256i)(a), \
-                                         (__v8si)(__m256i)(b), \
-                                         (p), (__mmask8)(m)); })
+                                         (__v8si)(__m256i)(b), (int)(p), \
+                                         (__mmask8)(m)); })
 
 #define _mm_cmp_epi64_mask(a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)(__m128i)(a), \
-                                        (__v2di)(__m128i)(b), \
-                                        (p), (__mmask8)-1); })
+                                        (__v2di)(__m128i)(b), (int)(p), \
+                                        (__mmask8)-1); })
 
 #define _mm_mask_cmp_epi64_mask(m, a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)(__m128i)(a), \
-                                        (__v2di)(__m128i)(b), \
-                                        (p), (__mmask8)(m)); })
+                                        (__v2di)(__m128i)(b), (int)(p), \
+                                        (__mmask8)(m)); })
 
 #define _mm_cmp_epu64_mask(a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)(__m128i)(a), \
-                                         (__v2di)(__m128i)(b), \
-                                         (p), (__mmask8)-1); })
+                                         (__v2di)(__m128i)(b), (int)(p), \
+                                         (__mmask8)-1); })
 
 #define _mm_mask_cmp_epu64_mask(m, a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)(__m128i)(a), \
-                                         (__v2di)(__m128i)(b), \
-                                         (p), (__mmask8)(m)); })
+                                         (__v2di)(__m128i)(b), (int)(p), \
+                                         (__mmask8)(m)); })
 
 #define _mm256_cmp_epi64_mask(a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)(__m256i)(a), \
-                                        (__v4di)(__m256i)(b), \
-                                        (p), (__mmask8)-1); })
+                                        (__v4di)(__m256i)(b), (int)(p), \
+                                        (__mmask8)-1); })
 
 #define _mm256_mask_cmp_epi64_mask(m, a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)(__m256i)(a), \
-                                        (__v4di)(__m256i)(b), \
-                                        (p), (__mmask8)(m)); })
+                                        (__v4di)(__m256i)(b), (int)(p), \
+                                        (__mmask8)(m)); })
 
 #define _mm256_cmp_epu64_mask(a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)(__m256i)(a), \
-                                         (__v4di)(__m256i)(b), \
-                                         (p), (__mmask8)-1); })
+                                         (__v4di)(__m256i)(b), (int)(p), \
+                                         (__mmask8)-1); })
 
 #define _mm256_mask_cmp_epu64_mask(m, a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)(__m256i)(a), \
-                                         (__v4di)(__m256i)(b), \
-                                         (p), (__mmask8)(m)); })
+                                         (__v4di)(__m256i)(b), (int)(p), \
+                                         (__mmask8)(m)); })
 
 #define _mm256_cmp_ps_mask(a, b, p)  __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpps256_mask((__v8sf)(__m256)(a), \
-                                         (__v8sf)(__m256)(b), \
-                                         (p), (__mmask8)-1); })
+                                         (__v8sf)(__m256)(b), (int)(p), \
+                                         (__mmask8)-1); })
 
 #define _mm256_mask_cmp_ps_mask(m, a, b, p)  __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpps256_mask((__v8sf)(__m256)(a), \
-                                         (__v8sf)(__m256)(b), \
-                                         (p), (__mmask8)(m)); })
+                                         (__v8sf)(__m256)(b), (int)(p), \
+                                         (__mmask8)(m)); })
 
 #define _mm256_cmp_pd_mask(a, b, p)  __extension__ ({ \
-  (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256)(a), \
-                                         (__v4df)(__m256)(b), \
-                                         (p), (__mmask8)-1); })
+  (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256d)(a), \
+                                         (__v4df)(__m256d)(b), (int)(p), \
+                                         (__mmask8)-1); })
 
 #define _mm256_mask_cmp_pd_mask(m, a, b, p)  __extension__ ({ \
-  (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256)(a), \
-                                         (__v4df)(__m256)(b), \
-                                         (p), (__mmask8)(m)); })
+  (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256d)(a), \
+                                         (__v4df)(__m256d)(b), (int)(p), \
+                                         (__mmask8)(m)); })
 
-#define _mm128_cmp_ps_mask(a, b, p)  __extension__ ({ \
+#define _mm_cmp_ps_mask(a, b, p)  __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \
-                                         (__v4sf)(__m128)(b), \
-                                         (p), (__mmask8)-1); })
+                                         (__v4sf)(__m128)(b), (int)(p), \
+                                         (__mmask8)-1); })
 
-#define _mm128_mask_cmp_ps_mask(m, a, b, p)  __extension__ ({ \
+#define _mm_mask_cmp_ps_mask(m, a, b, p)  __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \
-                                         (__v4sf)(__m128)(b), \
-                                         (p), (__mmask8)(m)); })
+                                         (__v4sf)(__m128)(b), (int)(p), \
+                                         (__mmask8)(m)); })
 
-#define _mm128_cmp_pd_mask(a, b, p)  __extension__ ({ \
-  (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128)(a), \
-                                         (__v2df)(__m128)(b), \
-                                         (p), (__mmask8)-1); })
+#define _mm_cmp_pd_mask(a, b, p)  __extension__ ({ \
+  (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128d)(a), \
+                                         (__v2df)(__m128d)(b), (int)(p), \
+                                         (__mmask8)-1); })
 
-#define _mm128_mask_cmp_pd_mask(m, a, b, p)  __extension__ ({ \
-  (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128)(a), \
-                                         (__v2df)(__m128)(b), \
-                                         (p), (__mmask8)(m)); })
+#define _mm_mask_cmp_pd_mask(m, a, b, p)  __extension__ ({ \
+  (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128d)(a), \
+                                         (__v2df)(__m128d)(b), (int)(p), \
+                                         (__mmask8)(m)); })
 
 static __inline__ __m128d __DEFAULT_FN_ATTRS
 _mm_mask_fmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
@@ -2044,58 +1964,58 @@ _mm256_maskz_add_ps (__mmask16 __U, __m256 __A, __m256 __B) {
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_mask_blend_epi32 (__mmask8 __U, __m128i __A, __m128i __W) {
-  return (__m128i) __builtin_ia32_blendmd_128_mask ((__v4si) __A,
+  return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U,
                 (__v4si) __W,
-                (__mmask8) __U);
+                (__v4si) __A);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
 _mm256_mask_blend_epi32 (__mmask8 __U, __m256i __A, __m256i __W) {
-  return (__m256i) __builtin_ia32_blendmd_256_mask ((__v8si) __A,
+  return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U,
                 (__v8si) __W,
-                (__mmask8) __U);
+                (__v8si) __A);
 }
 
 static __inline__ __m128d __DEFAULT_FN_ATTRS
 _mm_mask_blend_pd (__mmask8 __U, __m128d __A, __m128d __W) {
-  return (__m128d) __builtin_ia32_blendmpd_128_mask ((__v2df) __A,
+  return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U,
                  (__v2df) __W,
-                 (__mmask8) __U);
+                 (__v2df) __A);
 }
 
 static __inline__ __m256d __DEFAULT_FN_ATTRS
 _mm256_mask_blend_pd (__mmask8 __U, __m256d __A, __m256d __W) {
-  return (__m256d) __builtin_ia32_blendmpd_256_mask ((__v4df) __A,
+  return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U,
                  (__v4df) __W,
-                 (__mmask8) __U);
+                 (__v4df) __A);
 }
 
 static __inline__ __m128 __DEFAULT_FN_ATTRS
 _mm_mask_blend_ps (__mmask8 __U, __m128 __A, __m128 __W) {
-  return (__m128) __builtin_ia32_blendmps_128_mask ((__v4sf) __A,
+  return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U,
                 (__v4sf) __W,
-                (__mmask8) __U);
+                (__v4sf) __A);
 }
 
 static __inline__ __m256 __DEFAULT_FN_ATTRS
 _mm256_mask_blend_ps (__mmask8 __U, __m256 __A, __m256 __W) {
-  return (__m256) __builtin_ia32_blendmps_256_mask ((__v8sf) __A,
+  return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U,
                 (__v8sf) __W,
-                (__mmask8) __U);
+                (__v8sf) __A);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_mask_blend_epi64 (__mmask8 __U, __m128i __A, __m128i __W) {
-  return (__m128i) __builtin_ia32_blendmq_128_mask ((__v2di) __A,
+  return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U,
                 (__v2di) __W,
-                (__mmask8) __U);
+                (__v2di) __A);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
 _mm256_mask_blend_epi64 (__mmask8 __U, __m256i __A, __m256i __W) {
-  return (__m256i) __builtin_ia32_blendmq_256_mask ((__v4di) __A,
+  return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U,
                 (__v4di) __W,
-                (__mmask8) __U);
+                (__v4di) __A);
 }
 
 static __inline__ __m128d __DEFAULT_FN_ATTRS
@@ -3833,61 +3753,79 @@ _mm256_maskz_min_epu64 (__mmask8 __M, __m256i __A, __m256i __B) {
               __M);
 }
 
-#define _mm_roundscale_pd(__A, __imm) __extension__ ({ \
-  (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A, \
-                   __imm, (__v2df) _mm_setzero_pd (), (__mmask8) -1); })
+#define _mm_roundscale_pd(A, imm) __extension__ ({ \
+  (__m128d)__builtin_ia32_rndscalepd_128_mask((__v2df)(__m128d)(A), \
+                                              (int)(imm), \
+                                              (__v2df)_mm_setzero_pd(), \
+                                              (__mmask8)-1); })
 
 
-#define _mm_mask_roundscale_pd(__W, __U, __A, __imm) __extension__ ({ \
-  (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A, __imm, \
-                   (__v2df) __W, (__mmask8) __U); })
+#define _mm_mask_roundscale_pd(W, U, A, imm) __extension__ ({ \
+  (__m128d)__builtin_ia32_rndscalepd_128_mask((__v2df)(__m128d)(A), \
+                                              (int)(imm), \
+                                              (__v2df)(__m128d)(W), \
+                                              (__mmask8)(U)); })
 
 
-#define _mm_maskz_roundscale_pd(__U, __A, __imm) __extension__ ({ \
-  (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A, __imm, \
-                   (__v2df) _mm_setzero_pd (), (__mmask8) __U); })
+#define _mm_maskz_roundscale_pd(U, A, imm) __extension__ ({ \
+  (__m128d)__builtin_ia32_rndscalepd_128_mask((__v2df)(__m128d)(A), \
+                                              (int)(imm), \
+                                              (__v2df)_mm_setzero_pd(), \
+                                              (__mmask8)(U)); })
 
 
-#define _mm256_roundscale_pd(__A, __imm) __extension__ ({ \
-  (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A, __imm, \
-                   (__v4df) _mm256_setzero_pd (), (__mmask8) -1); })
+#define _mm256_roundscale_pd(A, imm) __extension__ ({ \
+  (__m256d)__builtin_ia32_rndscalepd_256_mask((__v4df)(__m256d)(A), \
+                                              (int)(imm), \
+                                              (__v4df)_mm256_setzero_pd(), \
+                                              (__mmask8)-1); })
 
 
-#define _mm256_mask_roundscale_pd(__W, __U, __A, __imm) __extension__ ({ \
-  (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A, __imm, \
-                   (__v4df) __W, (__mmask8) __U); })
+#define _mm256_mask_roundscale_pd(W, U, A, imm) __extension__ ({ \
+  (__m256d)__builtin_ia32_rndscalepd_256_mask((__v4df)(__m256d)(A), \
+                                              (int)(imm), \
+                                              (__v4df)(__m256d)(W), \
+                                              (__mmask8)(U)); })
 
 
-#define _mm256_maskz_roundscale_pd(__U, __A, __imm)  __extension__ ({ \
-  (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A, __imm, \
-                   (__v4df) _mm256_setzero_pd(), (__mmask8) __U); })
+#define _mm256_maskz_roundscale_pd(U, A, imm)  __extension__ ({ \
+  (__m256d)__builtin_ia32_rndscalepd_256_mask((__v4df)(__m256d)(A), \
+                                              (int)(imm), \
+                                              (__v4df)_mm256_setzero_pd(), \
+                                              (__mmask8)(U)); })
 
-#define _mm_roundscale_ps(__A, __imm)  __extension__ ({ \
-  (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A, __imm, \
-                  (__v4sf) _mm_setzero_ps(), (__mmask8) -1); })
+#define _mm_roundscale_ps(A, imm)  __extension__ ({ \
+  (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \
+                                             (__v4sf)_mm_setzero_ps(), \
+                                             (__mmask8)-1); })
 
 
-#define _mm_mask_roundscale_ps(__W, __U, __A, __imm)  __extension__ ({ \
-  (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A, __imm, \
-                  (__v4sf) __W, (__mmask8) __U); })
+#define _mm_mask_roundscale_ps(W, U, A, imm)  __extension__ ({ \
+  (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \
+                                             (__v4sf)(__m128)(W), \
+                                             (__mmask8)(U)); })
 
 
-#define _mm_maskz_roundscale_ps(__U, __A, __imm)  __extension__ ({ \
-  (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A, __imm, \
-                  (__v4sf) _mm_setzero_ps(), (__mmask8) __U); })
+#define _mm_maskz_roundscale_ps(U, A, imm)  __extension__ ({ \
+  (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \
+                                             (__v4sf)_mm_setzero_ps(), \
+                                             (__mmask8)(U)); })
 
-#define _mm256_roundscale_ps(__A, __imm)  __extension__ ({ \
-  (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A,__imm, \
-                  (__v8sf) _mm256_setzero_ps(), (__mmask8) -1); })
+#define _mm256_roundscale_ps(A, imm)  __extension__ ({ \
+  (__m256)__builtin_ia32_rndscaleps_256_mask((__v8sf)(__m256)(A), (int)(imm), \
+                                             (__v8sf)_mm256_setzero_ps(), \
+                                             (__mmask8)-1); })
 
-#define _mm256_mask_roundscale_ps(__W, __U, __A,__imm)  __extension__ ({ \
-  (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A, __imm, \
-                  (__v8sf) __W, (__mmask8) __U); })
+#define _mm256_mask_roundscale_ps(W, U, A, imm)  __extension__ ({ \
+  (__m256)__builtin_ia32_rndscaleps_256_mask((__v8sf)(__m256)(A), (int)(imm), \
+                                             (__v8sf)(__m256)(W), \
+                                             (__mmask8)(U)); })
 
 
-#define _mm256_maskz_roundscale_ps(__U, __A, __imm)  __extension__ ({ \
-  (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A, __imm, \
-                  (__v8sf) _mm256_setzero_ps(), (__mmask8) __U); })
+#define _mm256_maskz_roundscale_ps(U, A, imm)  __extension__ ({ \
+  (__m256)__builtin_ia32_rndscaleps_256_mask((__v8sf)(__m256)(A), (int)(imm), \
+                                             (__v8sf)_mm256_setzero_ps(), \
+                                             (__mmask8)(U)); })
 
 static __inline__ __m128d __DEFAULT_FN_ATTRS
 _mm_scalef_pd (__m128d __A, __m128d __B) {
@@ -3996,153 +3934,165 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
                (__mmask8) __U);
 }
 
-#define _mm_i64scatter_pd(__addr,__index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv2df(__addr, (__mmask8) 0xFF, (__v2di) __index, \
-                              (__v2df) __v1, __scale); })
-
-#define _mm_mask_i64scatter_pd(__addr, __mask, __index, __v1, \
-                               __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv2df (__addr, __mask, (__v2di) __index, \
-                               (__v2df) __v1, __scale); })
-
-
-#define _mm_i64scatter_epi64(__addr, __index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv2di (__addr, (__mmask8) 0xFF, \
-        (__v2di) __index, (__v2di) __v1, __scale); })
-
-#define _mm_mask_i64scatter_epi64(__addr, __mask, __index, __v1,\
-                                  __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv2di (__addr, __mask, (__v2di) __index,\
-        (__v2di) __v1, __scale); })
-
-#define _mm256_i64scatter_pd(__addr, __index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv4df (__addr, (__mmask8) 0xFF,\
-        (__v4di) __index, (__v4df) __v1, __scale); })
-
-#define _mm256_mask_i64scatter_pd(__addr, __mask, __index, __v1,\
-                                   __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv4df (__addr, __mask, (__v4di) __index,\
-        (__v4df) __v1, __scale); })
-
-#define _mm256_i64scatter_epi64(__addr, __index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv4di (__addr, (__mmask8) 0xFF, (__v4di) __index,\
-                               (__v4di) __v1, __scale); })
-
-#define _mm256_mask_i64scatter_epi64(__addr, __mask, __index, __v1,\
-                                      __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv4di (__addr, __mask, (__v4di) __index,\
-        (__v4di) __v1, __scale); })
-
-#define _mm_i64scatter_ps(__addr, __index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv4sf (__addr, (__mmask8) 0xFF,\
-        (__v2di) __index, (__v4sf) __v1, __scale); })
-
-#define _mm_mask_i64scatter_ps(__addr, __mask, __index, __v1, \
-                                __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv4sf (__addr, __mask, (__v2di) __index,\
-        (__v4sf) __v1, __scale); })
-
-#define _mm_i64scatter_epi32(__addr, __index, __v1, \
-                              __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv4si (__addr, (__mmask8) 0xFF,\
-        (__v2di) __index, (__v4si) __v1, __scale); })
-
-#define _mm_mask_i64scatter_epi32(__addr, __mask, __index, __v1,\
-         __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv4si (__addr, __mask, (__v2di) __index,\
-        (__v4si) __v1, __scale); })
-
-#define _mm256_i64scatter_ps(__addr, __index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv8sf (__addr, (__mmask8) 0xFF, (__v4di) __index, \
-                              (__v4sf) __v1, __scale); })
-
-#define _mm256_mask_i64scatter_ps(__addr, __mask, __index, __v1, \
-                                   __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv8sf (__addr, __mask, (__v4di) __index, \
-        (__v4sf) __v1, __scale); })
-
-#define _mm256_i64scatter_epi32(__addr, __index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scatterdiv8si (__addr, (__mmask8) 0xFF, \
-        (__v4di) __index, (__v4si) __v1, __scale); })
-
-#define _mm256_mask_i64scatter_epi32(__addr, __mask, __index, __v1, \
-                                      __scale) __extension__ ({  \
-  __builtin_ia32_scatterdiv8si(__addr, __mask, (__v4di) __index, \
-        (__v4si) __v1, __scale); })
-
-#define _mm_i32scatter_pd(__addr, __index, __v1,         \
-                          __scale) __extension__ ({      \
-  __builtin_ia32_scattersiv2df (__addr, (__mmask8) 0xFF, \
-        (__v4si) __index, (__v2df) __v1, __scale); })
-
-#define _mm_mask_i32scatter_pd(__addr, __mask, __index, __v1,    \
-                                __scale) __extension__ ({        \
-  __builtin_ia32_scattersiv2df (__addr, __mask, (__v4si) __index,\
-         (__v2df) __v1, __scale); })
-
-#define _mm_i32scatter_epi64(__addr, __index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scattersiv2di (__addr, (__mmask8) 0xFF,                       \
-        (__v4si) __index, (__v2di) __v1, __scale); })
-
-#define _mm_mask_i32scatter_epi64(__addr, __mask, __index, __v1, \
-         __scale) __extension__ ({                                \
-  __builtin_ia32_scattersiv2di (__addr, __mask, (__v4si) __index, \
-        (__v2di) __v1, __scale); })
-
-#define _mm256_i32scatter_pd(__addr, __index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scattersiv4df (__addr, (__mmask8) 0xFF,                      \
-        (__v4si) __index, (__v4df) __v1, __scale); })
-
-#define _mm256_mask_i32scatter_pd(__addr, __mask, __index, __v1, \
-         __scale) __extension__ ({                                \
-  __builtin_ia32_scattersiv4df (__addr, __mask, (__v4si) __index, \
-        (__v4df) __v1, __scale); })
-
-#define _mm256_i32scatter_epi64(__addr, __index, __v1,    \
-                                __scale) __extension__ ({ \
-  __builtin_ia32_scattersiv4di (__addr, (__mmask8) 0xFF,  \
-        (__v4si) __index, (__v4di) __v1, __scale); })
-
-#define _mm256_mask_i32scatter_epi64(__addr, __mask, __index, __v1, \
-            __scale) __extension__ ({                               \
-  __builtin_ia32_scattersiv4di (__addr, __mask, (__v4si) __index,   \
-        (__v4di) __v1, __scale); })
-
-#define _mm_i32scatter_ps(__addr, __index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scattersiv4sf (__addr, (__mmask8) 0xFF,                   \
-        (__v4si) __index, (__v4sf) __v1, __scale); })
-
-#define _mm_mask_i32scatter_ps(__addr, __mask, __index, __v1,     \
-                               __scale) __extension__ ({          \
-  __builtin_ia32_scattersiv4sf (__addr, __mask, (__v4si) __index, \
-        (__v4sf) __v1, __scale); })
-
-#define _mm_i32scatter_epi32(__addr, __index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scattersiv4si (__addr, (__mmask8) 0xFF,                       \
-        (__v4si) __index, (__v4si) __v1, __scale); })
-
-#define _mm_mask_i32scatter_epi32(__addr, __mask, __index, __v1, \
-                                  __scale) __extension__ ({      \
-  __builtin_ia32_scattersiv4si (__addr, __mask, (__v4si) __index,\
-        (__v4si) __v1, __scale); })
-
-#define _mm256_i32scatter_ps(__addr, __index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scattersiv8sf (__addr, (__mmask8) 0xFF,                      \
-        (__v8si) __index, (__v8sf) __v1, __scale); })
-
-#define _mm256_mask_i32scatter_ps(__addr, __mask, __index, __v1, \
-                                   __scale) __extension__ ({     \
-  __builtin_ia32_scattersiv8sf (__addr, __mask, (__v8si) __index,\
-        (__v8sf) __v1, __scale); })
-
-#define _mm256_i32scatter_epi32(__addr, __index, __v1, __scale) __extension__ ({ \
-  __builtin_ia32_scattersiv8si (__addr, (__mmask8) 0xFF,                         \
-        (__v8si) __index, (__v8si) __v1, __scale); })
-
-#define _mm256_mask_i32scatter_epi32(__addr, __mask, __index, __v1, \
-            __scale) __extension__ ({                                \
-  __builtin_ia32_scattersiv8si (__addr, __mask, (__v8si) __index,    \
-        (__v8si) __v1, __scale); })
+#define _mm_i64scatter_pd(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv2df((double *)(addr), (__mmask8)-1, \
+                               (__v2di)(__m128i)(index), \
+                               (__v2df)(__m128d)(v1), (int)(scale)); })
+
+#define _mm_mask_i64scatter_pd(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv2df((double *)(addr), (__mmask8)(mask), \
+                               (__v2di)(__m128i)(index), \
+                               (__v2df)(__m128d)(v1), (int)(scale)); })
+
+#define _mm_i64scatter_epi64(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv2di((long long *)(addr), (__mmask8)-1, \
+                               (__v2di)(__m128i)(index), \
+                               (__v2di)(__m128i)(v1), (int)(scale)); })
+
+#define _mm_mask_i64scatter_epi64(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv2di((long long *)(addr), (__mmask8)(mask), \
+                               (__v2di)(__m128i)(index), \
+                               (__v2di)(__m128i)(v1), (int)(scale)); })
+
+#define _mm256_i64scatter_pd(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv4df((double *)(addr), (__mmask8)-1, \
+                               (__v4di)(__m256i)(index), \
+                               (__v4df)(__m256d)(v1), (int)(scale)); })
+
+#define _mm256_mask_i64scatter_pd(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv4df((double *)(addr), (__mmask8)(mask), \
+                               (__v4di)(__m256i)(index), \
+                               (__v4df)(__m256d)(v1), (int)(scale)); })
+
+#define _mm256_i64scatter_epi64(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv4di((long long *)(addr), (__mmask8)-1, \
+                               (__v4di)(__m256i)(index), \
+                               (__v4di)(__m256i)(v1), (int)(scale)); })
+
+#define _mm256_mask_i64scatter_epi64(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv4di((long long *)(addr), (__mmask8)(mask), \
+                               (__v4di)(__m256i)(index), \
+                               (__v4di)(__m256i)(v1), (int)(scale)); })
+
+#define _mm_i64scatter_ps(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv4sf((float *)(addr), (__mmask8)-1, \
+                               (__v2di)(__m128i)(index), (__v4sf)(__m128)(v1), \
+                               (int)(scale)); })
+
+#define _mm_mask_i64scatter_ps(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv4sf((float *)(addr), (__mmask8)(mask), \
+                               (__v2di)(__m128i)(index), (__v4sf)(__m128)(v1), \
+                               (int)(scale)); })
+
+#define _mm_i64scatter_epi32(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv4si((int *)(addr), (__mmask8)-1, \
+                               (__v2di)(__m128i)(index), \
+                               (__v4si)(__m128i)(v1), (int)(scale)); })
+
+#define _mm_mask_i64scatter_epi32(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv4si((int *)(addr), (__mmask8)(mask), \
+                               (__v2di)(__m128i)(index), \
+                               (__v4si)(__m128i)(v1), (int)(scale)); })
+
+#define _mm256_i64scatter_ps(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv8sf((float *)(addr), (__mmask8)-1, \
+                               (__v4di)(__m256i)(index), (__v4sf)(__m128)(v1), \
+                               (int)(scale)); })
+
+#define _mm256_mask_i64scatter_ps(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv8sf((float *)(addr), (__mmask8)(mask), \
+                               (__v4di)(__m256i)(index), (__v4sf)(__m128)(v1), \
+                               (int)(scale)); })
+
+#define _mm256_i64scatter_epi32(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scatterdiv8si((int *)(addr), (__mmask8)-1, \
+                               (__v4di)(__m256i)(index), \
+                               (__v4si)(__m128i)(v1), (int)(scale)); })
+
+#define _mm256_mask_i64scatter_epi32(addr, mask, index, v1, scale) __extension__ ({  \
+  __builtin_ia32_scatterdiv8si((int *)(addr), (__mmask8)(mask), \
+                               (__v4di)(__m256i)(index), \
+                               (__v4si)(__m128i)(v1), (int)(scale)); })
+
+#define _mm_i32scatter_pd(addr, index, v1, scale) __extension__ ({      \
+  __builtin_ia32_scattersiv2df((double *)(addr), (__mmask8)-1, \
+                               (__v4si)(__m128i)(index), \
+                               (__v2df)(__m128d)(v1), (int)(scale)); })
+
+#define _mm_mask_i32scatter_pd(addr, mask, index, v1, scale) __extension__ ({        \
+  __builtin_ia32_scattersiv2df((double *)(addr), (__mmask8)(mask), \
+                               (__v4si)(__m128i)(index), \
+                               (__v2df)(__m128d)(v1), (int)(scale)); })
+
+#define _mm_i32scatter_epi64(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv2di((long long *)(addr), (__mmask8)-1, \
+                               (__v4si)(__m128i)(index), \
+                               (__v2di)(__m128i)(v1), (int)(scale)); })
+
+#define _mm_mask_i32scatter_epi64(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv2di((long long *)(addr), (__mmask8)(mask), \
+                               (__v4si)(__m128i)(index), \
+                               (__v2di)(__m128i)(v1), (int)(scale)); })
+
+#define _mm256_i32scatter_pd(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv4df((double *)(addr), (__mmask8)-1, \
+                               (__v4si)(__m128i)(index), \
+                               (__v4df)(__m256d)(v1), (int)(scale)); })
+
+#define _mm256_mask_i32scatter_pd(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv4df((double *)(addr), (__mmask8)(mask), \
+                               (__v4si)(__m128i)(index), \
+                               (__v4df)(__m256d)(v1), (int)(scale)); })
+
+#define _mm256_i32scatter_epi64(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv4di((long long *)(addr), (__mmask8)-1, \
+                               (__v4si)(__m128i)(index), \
+                               (__v4di)(__m256i)(v1), (int)(scale)); })
+
+#define _mm256_mask_i32scatter_epi64(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv4di((long long *)(addr), (__mmask8)(mask), \
+                               (__v4si)(__m128i)(index), \
+                               (__v4di)(__m256i)(v1), (int)(scale)); })
+
+#define _mm_i32scatter_ps(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv4sf((float *)(addr), (__mmask8)-1, \
+                               (__v4si)(__m128i)(index), (__v4sf)(__m128)(v1), \
+                               (int)(scale)); })
+
+#define _mm_mask_i32scatter_ps(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv4sf((float *)(addr), (__mmask8)(mask), \
+                               (__v4si)(__m128i)(index), (__v4sf)(__m128)(v1), \
+                               (int)(scale)); })
+
+#define _mm_i32scatter_epi32(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv4si((int *)(addr), (__mmask8)-1, \
+                               (__v4si)(__m128i)(index), \
+                               (__v4si)(__m128i)(v1), (int)(scale)); })
+
+#define _mm_mask_i32scatter_epi32(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv4si((int *)(addr), (__mmask8)(mask), \
+                               (__v4si)(__m128i)(index), \
+                               (__v4si)(__m128i)(v1), (int)(scale)); })
+
+#define _mm256_i32scatter_ps(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv8sf((float *)(addr), (__mmask8)-1, \
+                               (__v8si)(__m256i)(index), (__v8sf)(__m256)(v1), \
+                               (int)(scale)); })
+
+#define _mm256_mask_i32scatter_ps(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv8sf((float *)(addr), (__mmask8)(mask), \
+                               (__v8si)(__m256i)(index), (__v8sf)(__m256)(v1), \
+                               (int)(scale)); })
+
+#define _mm256_i32scatter_epi32(addr, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv8si((int *)(addr), (__mmask8)-1, \
+                               (__v8si)(__m256i)(index), \
+                               (__v8si)(__m256i)(v1), (int)(scale)); })
+
+#define _mm256_mask_i32scatter_epi32(addr, mask, index, v1, scale) __extension__ ({ \
+  __builtin_ia32_scattersiv8si((int *)(addr), (__mmask8)(mask), \
+                               (__v8si)(__m256i)(index), \
+                               (__v8si)(__m256i)(v1), (int)(scale)); })
 
 static __inline__ __m128d __DEFAULT_FN_ATTRS
 _mm_mask_sqrt_pd (__m128d __W, __mmask8 __U, __m128d __A) {
@@ -4600,7 +4550,4621 @@ _mm256_maskz_permutex2var_epi64 (__mmask8 __U, __m256i __A,
               __U);
 }
 
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi8_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A,
+                (__v4si) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A,
+                (__v4si)
+                _mm_setzero_si128 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi8_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A,
+                (__v8si) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A,
+                (__v8si)
+                _mm256_setzero_si256 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi8_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A,
+                (__v2di) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A,
+                (__v2di)
+                _mm_setzero_si128 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi8_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A,
+                (__v4di) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A,
+                (__v4di)
+                _mm256_setzero_si256 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi32_epi64 (__m128i __W, __mmask8 __U, __m128i __X)
+{
+  return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X,
+                (__v2di) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X)
+{
+  return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X,
+                (__v2di)
+                _mm_setzero_si128 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi32_epi64 (__m256i __W, __mmask8 __U, __m128i __X)
+{
+  return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X,
+                (__v4di) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X)
+{
+  return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X,
+                (__v4di)
+                _mm256_setzero_si256 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi16_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A,
+                (__v4si) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A,
+                (__v4si)
+                _mm_setzero_si128 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi16_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A,
+                (__v8si) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A,
+                (__v8si)
+                _mm256_setzero_si256 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi16_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A,
+                (__v2di) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A,
+                (__v2di)
+                _mm_setzero_si128 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi16_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A,
+                (__v4di) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A,
+                (__v4di)
+                _mm256_setzero_si256 (),
+                (__mmask8) __U);
+}
+
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepu8_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A,
+                (__v4si) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A,
+                (__v4si)
+                _mm_setzero_si128 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepu8_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A,
+                (__v8si) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A,
+                (__v8si)
+                _mm256_setzero_si256 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepu8_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A,
+                (__v2di) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A,
+                (__v2di)
+                _mm_setzero_si128 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepu8_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A,
+                (__v4di) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A,
+                (__v4di)
+                _mm256_setzero_si256 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepu32_epi64 (__m128i __W, __mmask8 __U, __m128i __X)
+{
+  return (__m128i) __builtin_ia32_pmovzxdq128_mask ((__v4si) __X,
+                (__v2di) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepu32_epi64 (__mmask8 __U, __m128i __X)
+{
+  return (__m128i) __builtin_ia32_pmovzxdq128_mask ((__v4si) __X,
+                (__v2di)
+                _mm_setzero_si128 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepu32_epi64 (__m256i __W, __mmask8 __U, __m128i __X)
+{
+  return (__m256i) __builtin_ia32_pmovzxdq256_mask ((__v4si) __X,
+                (__v4di) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepu32_epi64 (__mmask8 __U, __m128i __X)
+{
+  return (__m256i) __builtin_ia32_pmovzxdq256_mask ((__v4si) __X,
+                (__v4di)
+                _mm256_setzero_si256 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepu16_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovzxwd128_mask ((__v8hi) __A,
+                (__v4si) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovzxwd128_mask ((__v8hi) __A,
+                (__v4si)
+                _mm_setzero_si128 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepu16_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A,
+                (__v8si) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A,
+                (__v8si)
+                _mm256_setzero_si256 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepu16_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovzxwq128_mask ((__v8hi) __A,
+                (__v2di) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepu16_epi64 (__mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovzxwq128_mask ((__v8hi) __A,
+                (__v2di)
+                _mm_setzero_si128 (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepu16_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovzxwq256_mask ((__v8hi) __A,
+                (__v4di) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepu16_epi64 (__mmask8 __U, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_pmovzxwq256_mask ((__v8hi) __A,
+                (__v4di)
+                _mm256_setzero_si256 (),
+                (__mmask8) __U);
+}
+
+
+#define _mm_rol_epi32(a, b) __extension__ ({\
+  (__m128i)__builtin_ia32_prold128_mask((__v4si)(__m128i)(a), (int)(b), \
+                                        (__v4si)_mm_setzero_si128(), \
+                                        (__mmask8)-1); })
+
+#define _mm_mask_rol_epi32(w, u, a, b) __extension__ ({\
+  (__m128i)__builtin_ia32_prold128_mask((__v4si)(__m128i)(a), (int)(b), \
+                                        (__v4si)(__m128i)(w), (__mmask8)(u)); })
+
+#define _mm_maskz_rol_epi32(u, a, b) __extension__ ({\
+  (__m128i)__builtin_ia32_prold128_mask((__v4si)(__m128i)(a), (int)(b), \
+                                        (__v4si)_mm_setzero_si128(), \
+                                        (__mmask8)(u)); })
+
+#define _mm256_rol_epi32(a, b) __extension__ ({\
+  (__m256i)__builtin_ia32_prold256_mask((__v8si)(__m256i)(a), (int)(b), \
+                                        (__v8si)_mm256_setzero_si256(), \
+                                        (__mmask8)-1); })
+
+#define _mm256_mask_rol_epi32(w, u, a, b) __extension__ ({\
+  (__m256i)__builtin_ia32_prold256_mask((__v8si)(__m256i)(a), (int)(b), \
+                                        (__v8si)(__m256i)(w), (__mmask8)(u)); })
+
+#define _mm256_maskz_rol_epi32(u, a, b) __extension__ ({\
+  (__m256i)__builtin_ia32_prold256_mask((__v8si)(__m256i)(a), (int)(b), \
+                                        (__v8si)_mm256_setzero_si256(), \
+                                        (__mmask8)(u)); })
+
+#define _mm_rol_epi64(a, b) __extension__ ({\
+  (__m128i)__builtin_ia32_prolq128_mask((__v2di)(__m128i)(a), (int)(b), \
+                                        (__v2di)_mm_setzero_di(), \
+                                        (__mmask8)-1); })
+
+#define _mm_mask_rol_epi64(w, u, a, b) __extension__ ({\
+  (__m128i)__builtin_ia32_prolq128_mask((__v2di)(__m128i)(a), (int)(b), \
+                                        (__v2di)(__m128i)(w), (__mmask8)(u)); })
+
+#define _mm_maskz_rol_epi64(u, a, b) __extension__ ({\
+  (__m128i)__builtin_ia32_prolq128_mask((__v2di)(__m128i)(a), (int)(b), \
+                                        (__v2di)_mm_setzero_di(), \
+                                        (__mmask8)(u)); })
+
+#define _mm256_rol_epi64(a, b) __extension__ ({\
+  (__m256i)__builtin_ia32_prolq256_mask((__v4di)(__m256i)(a), (int)(b), \
+                                        (__v4di)_mm256_setzero_si256(), \
+                                        (__mmask8)-1); })
+
+#define _mm256_mask_rol_epi64(w, u, a, b) __extension__ ({\
+  (__m256i)__builtin_ia32_prolq256_mask((__v4di)(__m256i)(a), (int)(b), \
+                                        (__v4di)(__m256i)(w), (__mmask8)(u)); })
+
+#define _mm256_maskz_rol_epi64(u, a, b) __extension__ ({\
+  (__m256i)__builtin_ia32_prolq256_mask((__v4di)(__m256i)(a), (int)(b), \
+                                        (__v4di)_mm256_setzero_si256(), \
+                                        (__mmask8)(u)); })
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_rolv_epi32 (__m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A,
+              (__v4si) __B,
+              (__v4si)
+              _mm_setzero_si128 (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_rolv_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
+         __m128i __B)
+{
+  return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A,
+              (__v4si) __B,
+              (__v4si) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_rolv_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A,
+              (__v4si) __B,
+              (__v4si)
+              _mm_setzero_si128 (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_rolv_epi32 (__m256i __A, __m256i __B)
+{
+  return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A,
+              (__v8si) __B,
+              (__v8si)
+              _mm256_setzero_si256 (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_rolv_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
+      __m256i __B)
+{
+  return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A,
+              (__v8si) __B,
+              (__v8si) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_rolv_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A,
+              (__v8si) __B,
+              (__v8si)
+              _mm256_setzero_si256 (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_rolv_epi64 (__m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A,
+              (__v2di) __B,
+              (__v2di)
+              _mm_setzero_di (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_rolv_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
+         __m128i __B)
+{
+  return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A,
+              (__v2di) __B,
+              (__v2di) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_rolv_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A,
+              (__v2di) __B,
+              (__v2di)
+              _mm_setzero_di (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_rolv_epi64 (__m256i __A, __m256i __B)
+{
+  return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A,
+              (__v4di) __B,
+              (__v4di)
+              _mm256_setzero_si256 (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_rolv_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
+      __m256i __B)
+{
+  return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A,
+              (__v4di) __B,
+              (__v4di) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_rolv_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A,
+              (__v4di) __B,
+              (__v4di)
+              _mm256_setzero_si256 (),
+              (__mmask8) __U);
+}
+
+#define _mm_ror_epi32(A, B) __extension__ ({ \
+  (__m128i)__builtin_ia32_prord128_mask((__v4si)(__m128i)(A), (int)(B), \
+                                        (__v4si)_mm_setzero_si128(), \
+                                        (__mmask8)-1); })
+
+#define _mm_mask_ror_epi32(W, U, A, B) __extension__ ({ \
+  (__m128i)__builtin_ia32_prord128_mask((__v4si)(__m128i)(A), (int)(B), \
+                                        (__v4si)(__m128i)(W), (__mmask8)(U)); })
+
+#define _mm_maskz_ror_epi32(U, A, B) __extension__ ({ \
+  (__m128i)__builtin_ia32_prord128_mask((__v4si)(__m128i)(A), (int)(B), \
+                                        (__v4si)_mm_setzero_si128(), \
+                                        (__mmask8)(U)); })
+
+#define _mm256_ror_epi32(A, B) __extension__ ({ \
+  (__m256i)__builtin_ia32_prord256_mask((__v8si)(__m256i)(A), (int)(B), \
+                                        (__v8si)_mm256_setzero_si256(), \
+                                        (__mmask8)-1); })
+
+#define _mm256_mask_ror_epi32(W, U, A, B) __extension__ ({ \
+  (__m256i)__builtin_ia32_prord256_mask((__v8si)(__m256i)(A), (int)(B), \
+                                        (__v8si)(__m256i)(W), (__mmask8)(U)); })
+
+#define _mm256_maskz_ror_epi32(U, A, B) __extension__ ({ \
+  (__m256i)__builtin_ia32_prord256_mask((__v8si)(__m256i)(A), (int)(B), \
+                                        (__v8si)_mm256_setzero_si256(), \
+                                        (__mmask8)(U)); })
+
+#define _mm_ror_epi64(A, B) __extension__ ({ \
+  (__m128i)__builtin_ia32_prorq128_mask((__v2di)(__m128i)(A), (int)(B), \
+                                        (__v2di)_mm_setzero_di(), \
+                                        (__mmask8)-1); })
+
+#define _mm_mask_ror_epi64(W, U, A, B) __extension__ ({ \
+  (__m128i)__builtin_ia32_prorq128_mask((__v2di)(__m128i)(A), (int)(B), \
+                                        (__v2di)(__m128i)(W), (__mmask8)(U)); })
+
+#define _mm_maskz_ror_epi64(U, A, B) __extension__ ({ \
+  (__m128i)__builtin_ia32_prorq128_mask((__v2di)(__m128i)(A), (int)(B), \
+                                        (__v2di)_mm_setzero_di(), \
+                                        (__mmask8)(U)); })
+
+#define _mm256_ror_epi64(A, B) __extension__ ({ \
+  (__m256i)__builtin_ia32_prorq256_mask((__v4di)(__m256i)(A), (int)(B), \
+                                        (__v4di)_mm256_setzero_si256(), \
+                                        (__mmask8)-1); })
+
+#define _mm256_mask_ror_epi64(W, U, A, B) __extension__ ({ \
+  (__m256i)__builtin_ia32_prorq256_mask((__v4di)(__m256i)(A), (int)(B), \
+                                        (__v4di)(__m256i)(W), (__mmask8)(U)); })
+
+#define _mm256_maskz_ror_epi64(U, A, B) __extension__ ({ \
+  (__m256i)__builtin_ia32_prorq256_mask((__v4di)(__m256i)(A), (int)(B), \
+                                        (__v4di)_mm256_setzero_si256(), \
+                                        (__mmask8)(U)); })
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_sll_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
+        __m128i __B)
+{
+  return (__m128i) __builtin_ia32_pslld128_mask ((__v4si) __A,
+             (__v4si) __B,
+             (__v4si) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_sll_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_pslld128_mask ((__v4si) __A,
+             (__v4si) __B,
+             (__v4si)
+             _mm_setzero_si128 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_sll_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
+           __m128i __B)
+{
+  return (__m256i) __builtin_ia32_pslld256_mask ((__v8si) __A,
+             (__v4si) __B,
+             (__v8si) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_sll_epi32 (__mmask8 __U, __m256i __A, __m128i __B)
+{
+  return (__m256i) __builtin_ia32_pslld256_mask ((__v8si) __A,
+             (__v4si) __B,
+             (__v8si)
+             _mm256_setzero_si256 (),
+             (__mmask8) __U);
+}
+
+#define _mm_mask_slli_epi32(W, U, A, B) __extension__ ({ \
+  (__m128i)__builtin_ia32_pslldi128_mask((__v4si)(__m128i)(A), (int)(B), \
+                                         (__v4si)(__m128i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm_maskz_slli_epi32(U, A, B) __extension__ ({ \
+  (__m128i)__builtin_ia32_pslldi128_mask((__v4si)(__m128i)(A), (int)(B), \
+                                         (__v4si)_mm_setzero_si128(), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_mask_slli_epi32(W, U, A, B) __extension__ ({ \
+  (__m256i)__builtin_ia32_pslldi256_mask((__v8si)(__m256i)(A), (int)(B), \
+                                         (__v8si)(__m256i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_maskz_slli_epi32(U, A, B) __extension__ ({ \
+  (__m256i)__builtin_ia32_pslldi256_mask((__v8si)(__m256i)(A), (int)(B), \
+                                         (__v8si)_mm256_setzero_si256(), \
+                                         (__mmask8)(U)); })
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_sll_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
+        __m128i __B)
+{
+  return (__m128i) __builtin_ia32_psllq128_mask ((__v2di) __A,
+             (__v2di) __B,
+             (__v2di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_sll_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_psllq128_mask ((__v2di) __A,
+             (__v2di) __B,
+             (__v2di)
+             _mm_setzero_di (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_sll_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
+           __m128i __B)
+{
+  return (__m256i) __builtin_ia32_psllq256_mask ((__v4di) __A,
+             (__v2di) __B,
+             (__v4di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_sll_epi64 (__mmask8 __U, __m256i __A, __m128i __B)
+{
+  return (__m256i) __builtin_ia32_psllq256_mask ((__v4di) __A,
+             (__v2di) __B,
+             (__v4di)
+             _mm256_setzero_si256 (),
+             (__mmask8) __U);
+}
+
+#define _mm_mask_slli_epi64(W, U, A, B) __extension__ ({ \
+  (__m128i)__builtin_ia32_psllqi128_mask((__v2di)(__m128i)(A), (int)(B), \
+                                         (__v2di)(__m128i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm_maskz_slli_epi64(U, A, B) __extension__ ({ \
+  (__m128i)__builtin_ia32_psllqi128_mask((__v2di)(__m128i)(A), (int)(B), \
+                                         (__v2di)_mm_setzero_di(), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_mask_slli_epi64(W, U, A, B) __extension__ ({ \
+  (__m256i)__builtin_ia32_psllqi256_mask((__v4di)(__m256i)(A), (int)(B), \
+                                         (__v4di)(__m256i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_maskz_slli_epi64(U, A, B) __extension__ ({ \
+  (__m256i)__builtin_ia32_psllqi256_mask((__v4di)(__m256i)(A), (int)(B), \
+                                         (__v4di)_mm256_setzero_si256(), \
+                                         (__mmask8)(U)); })
+
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_rorv_epi32 (__m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A,
+              (__v4si) __B,
+              (__v4si)
+              _mm_setzero_si128 (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_rorv_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
+         __m128i __B)
+{
+  return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A,
+              (__v4si) __B,
+              (__v4si) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_rorv_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A,
+              (__v4si) __B,
+              (__v4si)
+              _mm_setzero_si128 (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_rorv_epi32 (__m256i __A, __m256i __B)
+{
+  return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A,
+              (__v8si) __B,
+              (__v8si)
+              _mm256_setzero_si256 (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_rorv_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
+      __m256i __B)
+{
+  return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A,
+              (__v8si) __B,
+              (__v8si) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_rorv_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A,
+              (__v8si) __B,
+              (__v8si)
+              _mm256_setzero_si256 (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_rorv_epi64 (__m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A,
+              (__v2di) __B,
+              (__v2di)
+              _mm_setzero_di (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_rorv_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
+         __m128i __B)
+{
+  return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A,
+              (__v2di) __B,
+              (__v2di) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_rorv_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A,
+              (__v2di) __B,
+              (__v2di)
+              _mm_setzero_di (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_rorv_epi64 (__m256i __A, __m256i __B)
+{
+  return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A,
+              (__v4di) __B,
+              (__v4di)
+              _mm256_setzero_si256 (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_rorv_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
+      __m256i __B)
+{
+  return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A,
+              (__v4di) __B,
+              (__v4di) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_rorv_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A,
+              (__v4di) __B,
+              (__v4di)
+              _mm256_setzero_si256 (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_sllv_epi64 (__m128i __W, __mmask8 __U, __m128i __X,
+         __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psllv2di_mask ((__v2di) __X,
+             (__v2di) __Y,
+             (__v2di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_sllv_epi64 (__mmask8 __U, __m128i __X, __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psllv2di_mask ((__v2di) __X,
+             (__v2di) __Y,
+             (__v2di)
+             _mm_setzero_di (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_sllv_epi64 (__m256i __W, __mmask8 __U, __m256i __X,
+      __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psllv4di_mask ((__v4di) __X,
+             (__v4di) __Y,
+             (__v4di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_sllv_epi64 (__mmask8 __U, __m256i __X, __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psllv4di_mask ((__v4di) __X,
+             (__v4di) __Y,
+             (__v4di)
+             _mm256_setzero_si256 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_sllv_epi32 (__m128i __W, __mmask8 __U, __m128i __X,
+         __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psllv4si_mask ((__v4si) __X,
+             (__v4si) __Y,
+             (__v4si) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_sllv_epi32 (__mmask8 __U, __m128i __X, __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psllv4si_mask ((__v4si) __X,
+             (__v4si) __Y,
+             (__v4si)
+             _mm_setzero_si128 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_sllv_epi32 (__m256i __W, __mmask8 __U, __m256i __X,
+      __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psllv8si_mask ((__v8si) __X,
+             (__v8si) __Y,
+             (__v8si) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_sllv_epi32 (__mmask8 __U, __m256i __X, __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psllv8si_mask ((__v8si) __X,
+             (__v8si) __Y,
+             (__v8si)
+             _mm256_setzero_si256 (),
+             (__mmask8) __U);
+}
+
+
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_srlv_epi64 (__m128i __W, __mmask8 __U, __m128i __X,
+         __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psrlv2di_mask ((__v2di) __X,
+             (__v2di) __Y,
+             (__v2di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_srlv_epi64 (__mmask8 __U, __m128i __X, __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psrlv2di_mask ((__v2di) __X,
+             (__v2di) __Y,
+             (__v2di)
+             _mm_setzero_di (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_srlv_epi64 (__m256i __W, __mmask8 __U, __m256i __X,
+      __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psrlv4di_mask ((__v4di) __X,
+             (__v4di) __Y,
+             (__v4di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_srlv_epi64 (__mmask8 __U, __m256i __X, __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psrlv4di_mask ((__v4di) __X,
+             (__v4di) __Y,
+             (__v4di)
+             _mm256_setzero_si256 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_srlv_epi32 (__m128i __W, __mmask8 __U, __m128i __X,
+         __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psrlv4si_mask ((__v4si) __X,
+             (__v4si) __Y,
+             (__v4si) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_srlv_epi32 (__mmask8 __U, __m128i __X, __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psrlv4si_mask ((__v4si) __X,
+             (__v4si) __Y,
+             (__v4si)
+             _mm_setzero_si128 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_srlv_epi32 (__m256i __W, __mmask8 __U, __m256i __X,
+      __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psrlv8si_mask ((__v8si) __X,
+             (__v8si) __Y,
+             (__v8si) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_srlv_epi32 (__mmask8 __U, __m256i __X, __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psrlv8si_mask ((__v8si) __X,
+             (__v8si) __Y,
+             (__v8si)
+             _mm256_setzero_si256 (),
+             (__mmask8) __U);
+}
+
+
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_srl_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
+        __m128i __B)
+{
+  return (__m128i) __builtin_ia32_psrld128_mask ((__v4si) __A,
+             (__v4si) __B,
+             (__v4si) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_srl_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_psrld128_mask ((__v4si) __A,
+             (__v4si) __B,
+             (__v4si)
+             _mm_setzero_si128 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_srl_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
+           __m128i __B)
+{
+  return (__m256i) __builtin_ia32_psrld256_mask ((__v8si) __A,
+             (__v4si) __B,
+             (__v8si) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_srl_epi32 (__mmask8 __U, __m256i __A, __m128i __B)
+{
+  return (__m256i) __builtin_ia32_psrld256_mask ((__v8si) __A,
+             (__v4si) __B,
+             (__v8si)
+             _mm256_setzero_si256 (),
+             (__mmask8) __U);
+}
+
+#define _mm_mask_srli_epi32(W, U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psrldi128_mask((__v4si)(__m128i)(A), (int)(imm), \
+                                         (__v4si)(__m128i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm_maskz_srli_epi32(U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psrldi128_mask((__v4si)(__m128i)(A), (int)(imm), \
+                                         (__v4si)_mm_setzero_si128(), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_mask_srli_epi32(W, U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psrldi256_mask((__v8si)(__m256i)(A), (int)(imm), \
+                                         (__v8si)(__m256i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_maskz_srli_epi32(U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psrldi256_mask((__v8si)(__m256i)(A), (int)(imm), \
+                                         (__v8si)_mm256_setzero_si256(), \
+                                         (__mmask8)(U)); })
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_srl_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
+        __m128i __B)
+{
+  return (__m128i) __builtin_ia32_psrlq128_mask ((__v2di) __A,
+             (__v2di) __B,
+             (__v2di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_srl_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_psrlq128_mask ((__v2di) __A,
+             (__v2di) __B,
+             (__v2di)
+             _mm_setzero_di (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_srl_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
+           __m128i __B)
+{
+  return (__m256i) __builtin_ia32_psrlq256_mask ((__v4di) __A,
+             (__v2di) __B,
+             (__v4di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_srl_epi64 (__mmask8 __U, __m256i __A, __m128i __B)
+{
+  return (__m256i) __builtin_ia32_psrlq256_mask ((__v4di) __A,
+             (__v2di) __B,
+             (__v4di)
+             _mm256_setzero_si256 (),
+             (__mmask8) __U);
+}
+
+#define _mm_mask_srli_epi64(W, U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psrlqi128_mask((__v2di)(__m128i)(A), (int)(imm), \
+                                         (__v2di)(__m128i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm_maskz_srli_epi64(U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psrlqi128_mask((__v2di)(__m128i)(A), (int)(imm), \
+                                         (__v2di)_mm_setzero_si128(), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_mask_srli_epi64(W, U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psrlqi256_mask((__v4di)(__m256i)(A), (int)(imm), \
+                                         (__v4di)(__m256i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_maskz_srli_epi64(U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psrlqi256_mask((__v4di)(__m256i)(A), (int)(imm), \
+                                         (__v4di)_mm256_setzero_si256(), \
+                                         (__mmask8)(U)); })
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_srav_epi32 (__m128i __W, __mmask8 __U, __m128i __X,
+         __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X,
+             (__v4si) __Y,
+             (__v4si) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_srav_epi32 (__mmask8 __U, __m128i __X, __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X,
+             (__v4si) __Y,
+             (__v4si)
+             _mm_setzero_si128 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_srav_epi32 (__m256i __W, __mmask8 __U, __m256i __X,
+      __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X,
+             (__v8si) __Y,
+             (__v8si) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_srav_epi32 (__mmask8 __U, __m256i __X, __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X,
+             (__v8si) __Y,
+             (__v8si)
+             _mm256_setzero_si256 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_srav_epi64 (__m128i __X, __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X,
+              (__v2di) __Y,
+              (__v2di)
+              _mm_setzero_di (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_srav_epi64 (__m128i __W, __mmask8 __U, __m128i __X,
+         __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X,
+              (__v2di) __Y,
+              (__v2di) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_srav_epi64 (__mmask8 __U, __m128i __X, __m128i __Y)
+{
+  return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X,
+              (__v2di) __Y,
+              (__v2di)
+              _mm_setzero_di (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_srav_epi64 (__m256i __X, __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X,
+              (__v4di) __Y,
+              (__v4di)
+              _mm256_setzero_si256 (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_srav_epi64 (__m256i __W, __mmask8 __U, __m256i __X,
+      __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X,
+              (__v4di) __Y,
+              (__v4di) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_srav_epi64 (__mmask8 __U, __m256i __X, __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X,
+              (__v4di) __Y,
+              (__v4di)
+              _mm256_setzero_si256 (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_mov_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U,
+                 (__v4si) __A,
+                 (__v4si) __W);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_mov_epi32 (__mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U,
+                 (__v4si) __A,
+                 (__v4si) _mm_setzero_si128 ());
+}
+
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_mov_epi32 (__m256i __W, __mmask8 __U, __m256i __A)
+{
+  return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U,
+                 (__v8si) __A,
+                 (__v8si) __W);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_mov_epi32 (__mmask8 __U, __m256i __A)
+{
+  return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U,
+                 (__v8si) __A,
+                 (__v8si) _mm256_setzero_si256 ());
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_load_epi32 (__m128i __W, __mmask8 __U, void const *__P)
+{
+  return (__m128i) __builtin_ia32_movdqa32load128_mask ((__v4si *) __P,
+              (__v4si) __W,
+              (__mmask8)
+              __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_load_epi32 (__mmask8 __U, void const *__P)
+{
+  return (__m128i) __builtin_ia32_movdqa32load128_mask ((__v4si *) __P,
+              (__v4si)
+              _mm_setzero_si128 (),
+              (__mmask8)
+              __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_load_epi32 (__m256i __W, __mmask8 __U, void const *__P)
+{
+  return (__m256i) __builtin_ia32_movdqa32load256_mask ((__v8si *) __P,
+              (__v8si) __W,
+              (__mmask8)
+              __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_load_epi32 (__mmask8 __U, void const *__P)
+{
+  return (__m256i) __builtin_ia32_movdqa32load256_mask ((__v8si *) __P,
+              (__v8si)
+              _mm256_setzero_si256 (),
+              (__mmask8)
+              __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_store_epi32 (void *__P, __mmask8 __U, __m128i __A)
+{
+  __builtin_ia32_movdqa32store128_mask ((__v4si *) __P,
+          (__v4si) __A,
+          (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_store_epi32 (void *__P, __mmask8 __U, __m256i __A)
+{
+  __builtin_ia32_movdqa32store256_mask ((__v8si *) __P,
+          (__v8si) __A,
+          (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_mov_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U,
+                 (__v2di) __A,
+                 (__v2di) __W);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_mov_epi64 (__mmask8 __U, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U,
+                 (__v2di) __A,
+                 (__v2di) _mm_setzero_di ());
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_mov_epi64 (__m256i __W, __mmask8 __U, __m256i __A)
+{
+  return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U,
+                 (__v4di) __A,
+                 (__v4di) __W);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_mov_epi64 (__mmask8 __U, __m256i __A)
+{
+  return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U,
+                 (__v4di) __A,
+                 (__v4di) _mm256_setzero_si256 ());
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_load_epi64 (__m128i __W, __mmask8 __U, void const *__P)
+{
+  return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P,
+              (__v2di) __W,
+              (__mmask8)
+              __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_load_epi64 (__mmask8 __U, void const *__P)
+{
+  return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P,
+              (__v2di)
+              _mm_setzero_di (),
+              (__mmask8)
+              __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_load_epi64 (__m256i __W, __mmask8 __U, void const *__P)
+{
+  return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P,
+              (__v4di) __W,
+              (__mmask8)
+              __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_load_epi64 (__mmask8 __U, void const *__P)
+{
+  return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P,
+              (__v4di)
+              _mm256_setzero_si256 (),
+              (__mmask8)
+              __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_store_epi64 (void *__P, __mmask8 __U, __m128i __A)
+{
+  __builtin_ia32_movdqa64store128_mask ((__v2di *) __P,
+          (__v2di) __A,
+          (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_store_epi64 (void *__P, __mmask8 __U, __m256i __A)
+{
+  __builtin_ia32_movdqa64store256_mask ((__v4di *) __P,
+          (__v4di) __A,
+          (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_movedup_pd (__m128d __W, __mmask8 __U, __m128d __A)
+{
+  return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
+                                              (__v2df)_mm_movedup_pd(__A),
+                                              (__v2df)__W);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_movedup_pd (__mmask8 __U, __m128d __A)
+{
+  return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
+                                              (__v2df)_mm_movedup_pd(__A),
+                                              (__v2df)_mm_setzero_pd());
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_movedup_pd (__m256d __W, __mmask8 __U, __m256d __A)
+{
+  return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
+                                              (__v4df)_mm256_movedup_pd(__A),
+                                              (__v4df)__W);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_movedup_pd (__mmask8 __U, __m256d __A)
+{
+  return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
+                                              (__v4df)_mm256_movedup_pd(__A),
+                                              (__v4df)_mm256_setzero_pd());
+}
+
+
+#define _mm_mask_set1_epi32(O, M, A) __extension__ ({ \
+  (__m128i)__builtin_ia32_pbroadcastd128_gpr_mask((int)(A), \
+                                                  (__v4si)(__m128i)(O), \
+                                                  (__mmask8)(M)); })
+
+#define _mm_maskz_set1_epi32(M, A) __extension__ ({ \
+  (__m128i)__builtin_ia32_pbroadcastd128_gpr_mask((int)(A), \
+                                                  (__v4si)_mm_setzero_si128(), \
+                                                  (__mmask8)(M)); })
+
+#define _mm256_mask_set1_epi32(O, M, A) __extension__ ({ \
+  (__m256i)__builtin_ia32_pbroadcastd256_gpr_mask((int)(A), \
+                                                  (__v8si)(__m256i)(O), \
+                                                  (__mmask8)(M)); })
+
+#define _mm256_maskz_set1_epi32(M, A) __extension__ ({ \
+  (__m256i)__builtin_ia32_pbroadcastd256_gpr_mask((int)(A), \
+                                                  (__v8si)_mm256_setzero_si256(), \
+                                                  (__mmask8)(M)); })
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_set1_epi64 (__m128i __O, __mmask8 __M, long long __A)
+{
+  return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A, (__v2di) __O,
+                 __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_set1_epi64 (__mmask8 __M, long long __A)
+{
+  return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A,
+                 (__v2di)
+                 _mm_setzero_si128 (),
+                 __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_set1_epi64 (__m256i __O, __mmask8 __M, long long __A)
+{
+  return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A, (__v4di) __O,
+                 __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_set1_epi64 (__mmask8 __M, long long __A)
+{
+  return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A,
+                 (__v4di)
+                 _mm256_setzero_si256 (),
+                 __M);
+}
+
+#define _mm_fixupimm_pd(A, B, C, imm) __extension__ ({ \
+  (__m128d)__builtin_ia32_fixupimmpd128_mask((__v2df)(__m128d)(A), \
+                                             (__v2df)(__m128d)(B), \
+                                             (__v2di)(__m128i)(C), (int)(imm), \
+                                             (__mmask8)-1); })
+
+#define _mm_mask_fixupimm_pd(A, U, B, C, imm) __extension__ ({ \
+  (__m128d)__builtin_ia32_fixupimmpd128_mask((__v2df)(__m128d)(A), \
+                                             (__v2df)(__m128d)(B), \
+                                             (__v2di)(__m128i)(C), (int)(imm), \
+                                             (__mmask8)(U)); })
+
+#define _mm_maskz_fixupimm_pd(U, A, B, C, imm) __extension__ ({ \
+  (__m128d)__builtin_ia32_fixupimmpd128_maskz((__v2df)(__m128d)(A), \
+                                              (__v2df)(__m128d)(B), \
+                                              (__v2di)(__m128i)(C), \
+                                              (int)(imm), (__mmask8)(U)); })
+
+#define _mm256_fixupimm_pd(A, B, C, imm) __extension__ ({ \
+  (__m256d)__builtin_ia32_fixupimmpd256_mask((__v4df)(__m256d)(A), \
+                                             (__v4df)(__m256d)(B), \
+                                             (__v4di)(__m256i)(C), (int)(imm), \
+                                             (__mmask8)-1); })
+
+#define _mm256_mask_fixupimm_pd(A, U, B, C, imm) __extension__ ({ \
+  (__m256d)__builtin_ia32_fixupimmpd256_mask((__v4df)(__m256d)(A), \
+                                             (__v4df)(__m256d)(B), \
+                                             (__v4di)(__m256i)(C), (int)(imm), \
+                                             (__mmask8)(U)); })
+
+#define _mm256_maskz_fixupimm_pd(U, A, B, C, imm) __extension__ ({ \
+  (__m256d)__builtin_ia32_fixupimmpd256_maskz((__v4df)(__m256d)(A), \
+                                              (__v4df)(__m256d)(B), \
+                                              (__v4di)(__m256i)(C), \
+                                              (int)(imm), (__mmask8)(U)); })
+
+#define _mm_fixupimm_ps(A, B, C, imm) __extension__ ({ \
+  (__m128)__builtin_ia32_fixupimmps128_mask((__v4sf)(__m128)(A), \
+                                            (__v4sf)(__m128)(B), \
+                                            (__v4si)(__m128i)(C), (int)(imm), \
+                                            (__mmask8)-1); })
+
+#define _mm_mask_fixupimm_ps(A, U, B, C, imm) __extension__ ({ \
+  (__m128)__builtin_ia32_fixupimmps128_mask((__v4sf)(__m128)(A), \
+                                            (__v4sf)(__m128)(B), \
+                                            (__v4si)(__m128i)(C), (int)(imm), \
+                                            (__mmask8)(U)); })
+
+#define _mm_maskz_fixupimm_ps(U, A, B, C, imm) __extension__ ({ \
+  (__m128)__builtin_ia32_fixupimmps128_maskz((__v4sf)(__m128)(A), \
+                                             (__v4sf)(__m128)(B), \
+                                             (__v4si)(__m128i)(C), (int)(imm), \
+                                             (__mmask8)(U)); })
+
+#define _mm256_fixupimm_ps(A, B, C, imm) __extension__ ({ \
+  (__m256)__builtin_ia32_fixupimmps256_mask((__v8sf)(__m256)(A), \
+                                            (__v8sf)(__m256)(B), \
+                                            (__v8si)(__m256i)(C), (int)(imm), \
+                                            (__mmask8)-1); })
+
+#define _mm256_mask_fixupimm_ps(A, U, B, C, imm) __extension__ ({ \
+  (__m256)__builtin_ia32_fixupimmps256_mask((__v8sf)(__m256)(A), \
+                                            (__v8sf)(__m256)(B), \
+                                            (__v8si)(__m256i)(C), (int)(imm), \
+                                            (__mmask8)(U)); })
+
+#define _mm256_maskz_fixupimm_ps(U, A, B, C, imm) __extension__ ({ \
+  (__m256)__builtin_ia32_fixupimmps256_maskz((__v8sf)(__m256)(A), \
+                                             (__v8sf)(__m256)(B), \
+                                             (__v8si)(__m256i)(C), (int)(imm), \
+                                             (__mmask8)(U)); })
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_load_pd (__m128d __W, __mmask8 __U, void const *__P)
+{
+  return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P,
+               (__v2df) __W,
+               (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_load_pd (__mmask8 __U, void const *__P)
+{
+  return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P,
+               (__v2df)
+               _mm_setzero_pd (),
+               (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_load_pd (__m256d __W, __mmask8 __U, void const *__P)
+{
+  return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P,
+               (__v4df) __W,
+               (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_load_pd (__mmask8 __U, void const *__P)
+{
+  return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P,
+               (__v4df)
+               _mm256_setzero_pd (),
+               (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_load_ps (__m128 __W, __mmask8 __U, void const *__P)
+{
+  return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P,
+              (__v4sf) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_load_ps (__mmask8 __U, void const *__P)
+{
+  return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P,
+              (__v4sf)
+              _mm_setzero_ps (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_load_ps (__m256 __W, __mmask8 __U, void const *__P)
+{
+  return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P,
+              (__v8sf) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_load_ps (__mmask8 __U, void const *__P)
+{
+  return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P,
+              (__v8sf)
+              _mm256_setzero_ps (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_loadu_epi64 (__m128i __W, __mmask8 __U, void const *__P)
+{
+  return (__m128i) __builtin_ia32_loaddqudi128_mask ((__v2di *) __P,
+                 (__v2di) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_loadu_epi64 (__mmask8 __U, void const *__P)
+{
+  return (__m128i) __builtin_ia32_loaddqudi128_mask ((__v2di *) __P,
+                 (__v2di)
+                 _mm_setzero_si128 (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_loadu_epi64 (__m256i __W, __mmask8 __U, void const *__P)
+{
+  return (__m256i) __builtin_ia32_loaddqudi256_mask ((__v4di *) __P,
+                 (__v4di) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_loadu_epi64 (__mmask8 __U, void const *__P)
+{
+  return (__m256i) __builtin_ia32_loaddqudi256_mask ((__v4di *) __P,
+                 (__v4di)
+                 _mm256_setzero_si256 (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_loadu_epi32 (__m128i __W, __mmask8 __U, void const *__P)
+{
+  return (__m128i) __builtin_ia32_loaddqusi128_mask ((__v4si *) __P,
+                 (__v4si) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_loadu_epi32 (__mmask8 __U, void const *__P)
+{
+  return (__m128i) __builtin_ia32_loaddqusi128_mask ((__v4si *) __P,
+                 (__v4si)
+                 _mm_setzero_si128 (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_loadu_epi32 (__m256i __W, __mmask8 __U, void const *__P)
+{
+  return (__m256i) __builtin_ia32_loaddqusi256_mask ((__v8si *) __P,
+                 (__v8si) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_loadu_epi32 (__mmask8 __U, void const *__P)
+{
+  return (__m256i) __builtin_ia32_loaddqusi256_mask ((__v8si *) __P,
+                 (__v8si)
+                 _mm256_setzero_si256 (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_loadu_pd (__m128d __W, __mmask8 __U, void const *__P)
+{
+  return (__m128d) __builtin_ia32_loadupd128_mask ((__v2df *) __P,
+               (__v2df) __W,
+               (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_loadu_pd (__mmask8 __U, void const *__P)
+{
+  return (__m128d) __builtin_ia32_loadupd128_mask ((__v2df *) __P,
+               (__v2df)
+               _mm_setzero_pd (),
+               (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_loadu_pd (__m256d __W, __mmask8 __U, void const *__P)
+{
+  return (__m256d) __builtin_ia32_loadupd256_mask ((__v4df *) __P,
+               (__v4df) __W,
+               (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_loadu_pd (__mmask8 __U, void const *__P)
+{
+  return (__m256d) __builtin_ia32_loadupd256_mask ((__v4df *) __P,
+               (__v4df)
+               _mm256_setzero_pd (),
+               (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_loadu_ps (__m128 __W, __mmask8 __U, void const *__P)
+{
+  return (__m128) __builtin_ia32_loadups128_mask ((__v4sf *) __P,
+              (__v4sf) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_loadu_ps (__mmask8 __U, void const *__P)
+{
+  return (__m128) __builtin_ia32_loadups128_mask ((__v4sf *) __P,
+              (__v4sf)
+              _mm_setzero_ps (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_loadu_ps (__m256 __W, __mmask8 __U, void const *__P)
+{
+  return (__m256) __builtin_ia32_loadups256_mask ((__v8sf *) __P,
+              (__v8sf) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_loadu_ps (__mmask8 __U, void const *__P)
+{
+  return (__m256) __builtin_ia32_loadups256_mask ((__v8sf *) __P,
+              (__v8sf)
+              _mm256_setzero_ps (),
+              (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_store_pd (void *__P, __mmask8 __U, __m128d __A)
+{
+  __builtin_ia32_storeapd128_mask ((__v2df *) __P,
+           (__v2df) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_store_pd (void *__P, __mmask8 __U, __m256d __A)
+{
+  __builtin_ia32_storeapd256_mask ((__v4df *) __P,
+           (__v4df) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_store_ps (void *__P, __mmask8 __U, __m128 __A)
+{
+  __builtin_ia32_storeaps128_mask ((__v4sf *) __P,
+           (__v4sf) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_store_ps (void *__P, __mmask8 __U, __m256 __A)
+{
+  __builtin_ia32_storeaps256_mask ((__v8sf *) __P,
+           (__v8sf) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_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_ATTRS
+_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_ATTRS
+_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_ATTRS
+_mm256_mask_storeu_epi32 (void *__P, __mmask8 __U, __m256i __A)
+{
+  __builtin_ia32_storedqusi256_mask ((__v8si *) __P,
+             (__v8si) __A,
+             (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_storeu_pd (void *__P, __mmask8 __U, __m128d __A)
+{
+  __builtin_ia32_storeupd128_mask ((__v2df *) __P,
+           (__v2df) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_storeu_pd (void *__P, __mmask8 __U, __m256d __A)
+{
+  __builtin_ia32_storeupd256_mask ((__v4df *) __P,
+           (__v4df) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_storeu_ps (void *__P, __mmask8 __U, __m128 __A)
+{
+  __builtin_ia32_storeups128_mask ((__v4sf *) __P,
+           (__v4sf) __A,
+           (__mmask8) __U);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_storeu_ps (void *__P, __mmask8 __U, __m256 __A)
+{
+  __builtin_ia32_storeups256_mask ((__v8sf *) __P,
+           (__v8sf) __A,
+           (__mmask8) __U);
+}
+
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_unpackhi_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
+{
+  return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
+                                              (__v2df)_mm_unpackhi_pd(__A, __B),
+                                              (__v2df)__W);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_unpackhi_pd(__mmask8 __U, __m128d __A, __m128d __B)
+{
+  return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
+                                              (__v2df)_mm_unpackhi_pd(__A, __B),
+                                              (__v2df)_mm_setzero_pd());
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_unpackhi_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B)
+{
+  return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
+                                           (__v4df)_mm256_unpackhi_pd(__A, __B),
+                                           (__v4df)__W);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_unpackhi_pd(__mmask8 __U, __m256d __A, __m256d __B)
+{
+  return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
+                                           (__v4df)_mm256_unpackhi_pd(__A, __B),
+                                           (__v4df)_mm256_setzero_pd());
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_unpackhi_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
+{
+  return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
+                                             (__v4sf)_mm_unpackhi_ps(__A, __B),
+                                             (__v4sf)__W);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_unpackhi_ps(__mmask8 __U, __m128 __A, __m128 __B)
+{
+  return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
+                                             (__v4sf)_mm_unpackhi_ps(__A, __B),
+                                             (__v4sf)_mm_setzero_ps());
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_unpackhi_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B)
+{
+  return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
+                                           (__v8sf)_mm256_unpackhi_ps(__A, __B),
+                                           (__v8sf)__W);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_unpackhi_ps(__mmask8 __U, __m256 __A, __m256 __B)
+{
+  return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
+                                           (__v8sf)_mm256_unpackhi_ps(__A, __B),
+                                           (__v8sf)_mm256_setzero_ps());
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_unpacklo_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
+{
+  return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
+                                              (__v2df)_mm_unpacklo_pd(__A, __B),
+                                              (__v2df)__W);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_unpacklo_pd(__mmask8 __U, __m128d __A, __m128d __B)
+{
+  return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
+                                              (__v2df)_mm_unpacklo_pd(__A, __B),
+                                              (__v2df)_mm_setzero_pd());
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_unpacklo_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B)
+{
+  return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
+                                           (__v4df)_mm256_unpacklo_pd(__A, __B),
+                                           (__v4df)__W);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_unpacklo_pd(__mmask8 __U, __m256d __A, __m256d __B)
+{
+  return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
+                                           (__v4df)_mm256_unpacklo_pd(__A, __B),
+                                           (__v4df)_mm256_setzero_pd());
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_unpacklo_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
+{
+  return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
+                                             (__v4sf)_mm_unpacklo_ps(__A, __B),
+                                             (__v4sf)__W);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_unpacklo_ps(__mmask8 __U, __m128 __A, __m128 __B)
+{
+  return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
+                                             (__v4sf)_mm_unpacklo_ps(__A, __B),
+                                             (__v4sf)_mm_setzero_ps());
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_unpacklo_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B)
+{
+  return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
+                                           (__v8sf)_mm256_unpacklo_ps(__A, __B),
+                                           (__v8sf)__W);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_unpacklo_ps(__mmask8 __U, __m256 __A, __m256 __B)
+{
+  return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
+                                           (__v8sf)_mm256_unpacklo_ps(__A, __B),
+                                           (__v8sf)_mm256_setzero_ps());
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_rcp14_pd (__m128d __A)
+{
+  return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A,
+                (__v2df)
+                _mm_setzero_pd (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_rcp14_pd (__m128d __W, __mmask8 __U, __m128d __A)
+{
+  return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A,
+                (__v2df) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_rcp14_pd (__mmask8 __U, __m128d __A)
+{
+  return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A,
+                (__v2df)
+                _mm_setzero_pd (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_rcp14_pd (__m256d __A)
+{
+  return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A,
+                (__v4df)
+                _mm256_setzero_pd (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_rcp14_pd (__m256d __W, __mmask8 __U, __m256d __A)
+{
+  return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A,
+                (__v4df) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_rcp14_pd (__mmask8 __U, __m256d __A)
+{
+  return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A,
+                (__v4df)
+                _mm256_setzero_pd (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_rcp14_ps (__m128 __A)
+{
+  return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A,
+               (__v4sf)
+               _mm_setzero_ps (),
+               (__mmask8) -1);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_rcp14_ps (__m128 __W, __mmask8 __U, __m128 __A)
+{
+  return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A,
+               (__v4sf) __W,
+               (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_rcp14_ps (__mmask8 __U, __m128 __A)
+{
+  return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A,
+               (__v4sf)
+               _mm_setzero_ps (),
+               (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_rcp14_ps (__m256 __A)
+{
+  return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A,
+               (__v8sf)
+               _mm256_setzero_ps (),
+               (__mmask8) -1);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_rcp14_ps (__m256 __W, __mmask8 __U, __m256 __A)
+{
+  return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A,
+               (__v8sf) __W,
+               (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_rcp14_ps (__mmask8 __U, __m256 __A)
+{
+  return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A,
+               (__v8sf)
+               _mm256_setzero_ps (),
+               (__mmask8) __U);
+}
+
+#define _mm_mask_permute_pd(W, U, X, C) __extension__ ({ \
+  (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \
+                                       (__v2df)_mm_permute_pd((X), (C)), \
+                                       (__v2df)(__m128d)(W)); })
+
+#define _mm_maskz_permute_pd(U, X, C) __extension__ ({ \
+  (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \
+                                       (__v2df)_mm_permute_pd((X), (C)), \
+                                       (__v2df)_mm_setzero_pd()); })
+
+#define _mm256_mask_permute_pd(W, U, X, C) __extension__ ({ \
+  (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \
+                                       (__v4df)_mm256_permute_pd((X), (C)), \
+                                       (__v4df)(__m256d)(W)); })
+
+#define _mm256_maskz_permute_pd(U, X, C) __extension__ ({ \
+  (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \
+                                       (__v4df)_mm256_permute_pd((X), (C)), \
+                                       (__v4df)_mm256_setzero_pd()); })
+
+#define _mm_mask_permute_ps(W, U, X, C) __extension__ ({ \
+  (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \
+                                      (__v4sf)_mm_permute_ps((X), (C)), \
+                                      (__v4sf)(__m128)(W)); })
+
+#define _mm_maskz_permute_ps(U, X, C) __extension__ ({ \
+  (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \
+                                      (__v4sf)_mm_permute_ps((X), (C)), \
+                                      (__v4sf)_mm_setzero_ps()); })
+
+#define _mm256_mask_permute_ps(W, U, X, C) __extension__ ({ \
+  (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \
+                                      (__v8sf)_mm256_permute_ps((X), (C)), \
+                                      (__v8sf)(__m256)(W)); })
+
+#define _mm256_maskz_permute_ps(U, X, C) __extension__ ({ \
+  (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \
+                                      (__v8sf)_mm256_permute_ps((X), (C)), \
+                                      (__v8sf)_mm256_setzero_ps()); })
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_permutevar_pd (__m128d __W, __mmask8 __U, __m128d __A,
+      __m128i __C)
+{
+  return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A,
+                 (__v2di) __C,
+                 (__v2df) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_permutevar_pd (__mmask8 __U, __m128d __A, __m128i __C)
+{
+  return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A,
+                 (__v2di) __C,
+                 (__v2df)
+                 _mm_setzero_pd (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_permutevar_pd (__m256d __W, __mmask8 __U, __m256d __A,
+         __m256i __C)
+{
+  return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A,
+              (__v4di) __C,
+              (__v4df) __W,
+              (__mmask8)
+              __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_permutevar_pd (__mmask8 __U, __m256d __A, __m256i __C)
+{
+  return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A,
+              (__v4di) __C,
+              (__v4df)
+              _mm256_setzero_pd (),
+              (__mmask8)
+              __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_permutevar_ps (__m128 __W, __mmask8 __U, __m128 __A,
+      __m128i __C)
+{
+  return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A,
+                (__v4si) __C,
+                (__v4sf) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_permutevar_ps (__mmask8 __U, __m128 __A, __m128i __C)
+{
+  return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A,
+                (__v4si) __C,
+                (__v4sf)
+                _mm_setzero_ps (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_permutevar_ps (__m256 __W, __mmask8 __U, __m256 __A,
+         __m256i __C)
+{
+  return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A,
+                   (__v8si) __C,
+                   (__v8sf) __W,
+                   (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_permutevar_ps (__mmask8 __U, __m256 __A, __m256i __C)
+{
+  return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A,
+                   (__v8si) __C,
+                   (__v8sf)
+                   _mm256_setzero_ps (),
+                   (__mmask8) __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_test_epi32_mask (__m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmd128 ((__v4si) __A,
+                 (__v4si) __B,
+                 (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_mask_test_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmd128 ((__v4si) __A,
+                 (__v4si) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_test_epi32_mask (__m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmd256 ((__v8si) __A,
+                 (__v8si) __B,
+                 (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_mask_test_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmd256 ((__v8si) __A,
+                 (__v8si) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_test_epi64_mask (__m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmq128 ((__v2di) __A,
+                 (__v2di) __B,
+                 (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_mask_test_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmq128 ((__v2di) __A,
+                 (__v2di) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_test_epi64_mask (__m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmq256 ((__v4di) __A,
+                 (__v4di) __B,
+                 (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_mask_test_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestmq256 ((__v4di) __A,
+                 (__v4di) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_testn_epi32_mask (__m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmd128 ((__v4si) __A,
+            (__v4si) __B,
+            (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_mask_testn_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmd128 ((__v4si) __A,
+            (__v4si) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_testn_epi32_mask (__m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmd256 ((__v8si) __A,
+            (__v8si) __B,
+            (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_mask_testn_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmd256 ((__v8si) __A,
+            (__v8si) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_testn_epi64_mask (__m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmq128 ((__v2di) __A,
+            (__v2di) __B,
+            (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_mask_testn_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmq128 ((__v2di) __A,
+            (__v2di) __B, __U);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_testn_epi64_mask (__m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmq256 ((__v4di) __A,
+            (__v4di) __B,
+            (__mmask8) -1);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_mask_testn_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__mmask8) __builtin_ia32_ptestnmq256 ((__v4di) __A,
+            (__v4di) __B, __U);
+}
+
+
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_unpackhi_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
+                                           (__v4si)_mm_unpackhi_epi32(__A, __B),
+                                           (__v4si)__W);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_unpackhi_epi32(__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
+                                           (__v4si)_mm_unpackhi_epi32(__A, __B),
+                                           (__v4si)_mm_setzero_si128());
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_unpackhi_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
+                                        (__v8si)_mm256_unpackhi_epi32(__A, __B),
+                                        (__v8si)__W);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_unpackhi_epi32(__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
+                                        (__v8si)_mm256_unpackhi_epi32(__A, __B),
+                                        (__v8si)_mm256_setzero_si256());
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_unpackhi_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
+                                           (__v2di)_mm_unpackhi_epi64(__A, __B),
+                                           (__v2di)__W);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_unpackhi_epi64(__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
+                                           (__v2di)_mm_unpackhi_epi64(__A, __B),
+                                           (__v2di)_mm_setzero_di());
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_unpackhi_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
+                                        (__v4di)_mm256_unpackhi_epi64(__A, __B),
+                                        (__v4di)__W);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_unpackhi_epi64(__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
+                                        (__v4di)_mm256_unpackhi_epi64(__A, __B),
+                                        (__v4di)_mm256_setzero_si256());
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_unpacklo_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
+                                           (__v4si)_mm_unpacklo_epi32(__A, __B),
+                                           (__v4si)__W);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_unpacklo_epi32(__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
+                                           (__v4si)_mm_unpacklo_epi32(__A, __B),
+                                           (__v4si)_mm_setzero_si128());
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_unpacklo_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
+                                        (__v8si)_mm256_unpacklo_epi32(__A, __B),
+                                        (__v8si)__W);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_unpacklo_epi32(__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
+                                        (__v8si)_mm256_unpacklo_epi32(__A, __B),
+                                        (__v8si)_mm256_setzero_si256());
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_unpacklo_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
+                                           (__v2di)_mm_unpacklo_epi64(__A, __B),
+                                           (__v2di)__W);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_unpacklo_epi64(__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
+                                           (__v2di)_mm_unpacklo_epi64(__A, __B),
+                                           (__v2di)_mm_setzero_di());
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_unpacklo_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
+                                        (__v4di)_mm256_unpacklo_epi64(__A, __B),
+                                        (__v4di)__W);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_unpacklo_epi64(__mmask8 __U, __m256i __A, __m256i __B)
+{
+  return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
+                                        (__v4di)_mm256_unpacklo_epi64(__A, __B),
+                                        (__v4di)_mm256_setzero_si256());
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_sra_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
+        __m128i __B)
+{
+  return (__m128i) __builtin_ia32_psrad128_mask ((__v4si) __A,
+             (__v4si) __B,
+             (__v4si) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_sra_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_psrad128_mask ((__v4si) __A,
+             (__v4si) __B,
+             (__v4si)
+             _mm_setzero_si128 (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_sra_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
+           __m128i __B)
+{
+  return (__m256i) __builtin_ia32_psrad256_mask ((__v8si) __A,
+             (__v4si) __B,
+             (__v8si) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_sra_epi32 (__mmask8 __U, __m256i __A, __m128i __B)
+{
+  return (__m256i) __builtin_ia32_psrad256_mask ((__v8si) __A,
+             (__v4si) __B,
+             (__v8si)
+             _mm256_setzero_si256 (),
+             (__mmask8) __U);
+}
+
+#define _mm_mask_srai_epi32(W, U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psradi128_mask((__v4si)(__m128i)(A), (int)(imm), \
+                                         (__v4si)(__m128i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm_maskz_srai_epi32(U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psradi128_mask((__v4si)(__m128i)(A), (int)(imm), \
+                                         (__v4si)_mm_setzero_si128(), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_mask_srai_epi32(W, U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psradi256_mask((__v8si)(__m256i)(A), (int)(imm), \
+                                         (__v8si)(__m256i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_maskz_srai_epi32(U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psradi256_mask((__v8si)(__m256i)(A), (int)(imm), \
+                                         (__v8si)_mm256_setzero_si256(), \
+                                         (__mmask8)(U)); })
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_sra_epi64 (__m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A,
+             (__v2di) __B,
+             (__v2di)
+             _mm_setzero_di (),
+             (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_sra_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
+        __m128i __B)
+{
+  return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A,
+             (__v2di) __B,
+             (__v2di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_sra_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
+{
+  return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A,
+             (__v2di) __B,
+             (__v2di)
+             _mm_setzero_di (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_sra_epi64 (__m256i __A, __m128i __B)
+{
+  return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A,
+             (__v2di) __B,
+             (__v4di)
+             _mm256_setzero_si256 (),
+             (__mmask8) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_sra_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
+           __m128i __B)
+{
+  return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A,
+             (__v2di) __B,
+             (__v4di) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_sra_epi64 (__mmask8 __U, __m256i __A, __m128i __B)
+{
+  return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A,
+             (__v2di) __B,
+             (__v4di)
+             _mm256_setzero_si256 (),
+             (__mmask8) __U);
+}
+
+#define _mm_srai_epi64(A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psraqi128_mask((__v2di)(__m128i)(A), (int)(imm), \
+                                         (__v2di)_mm_setzero_di(), \
+                                         (__mmask8)-1); })
+
+#define _mm_mask_srai_epi64(W, U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psraqi128_mask((__v2di)(__m128i)(A), (int)(imm), \
+                                         (__v2di)(__m128i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm_maskz_srai_epi64(U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psraqi128_mask((__v2di)(__m128i)(A), (int)(imm), \
+                                         (__v2di)_mm_setzero_si128(), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_srai_epi64(A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psraqi256_mask((__v4di)(__m256i)(A), (int)(imm), \
+                                         (__v4di)_mm256_setzero_si256(), \
+                                         (__mmask8)-1); })
+
+#define _mm256_mask_srai_epi64(W, U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psraqi256_mask((__v4di)(__m256i)(A), (int)(imm), \
+                                         (__v4di)(__m256i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_maskz_srai_epi64(U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psraqi256_mask((__v4di)(__m256i)(A), (int)(imm), \
+                                         (__v4di)_mm256_setzero_si256(), \
+                                         (__mmask8)(U)); })
+
+#define _mm_ternarylogic_epi32(A, B, C, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_pternlogd128_mask((__v4si)(__m128i)(A), \
+                                            (__v4si)(__m128i)(B), \
+                                            (__v4si)(__m128i)(C), (int)(imm), \
+                                            (__mmask8)-1); })
+
+#define _mm_mask_ternarylogic_epi32(A, U, B, C, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_pternlogd128_mask((__v4si)(__m128i)(A), \
+                                            (__v4si)(__m128i)(B), \
+                                            (__v4si)(__m128i)(C), (int)(imm), \
+                                            (__mmask8)(U)); })
+
+#define _mm_maskz_ternarylogic_epi32(U, A, B, C, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_pternlogd128_maskz((__v4si)(__m128i)(A), \
+                                             (__v4si)(__m128i)(B), \
+                                             (__v4si)(__m128i)(C), (int)(imm), \
+                                             (__mmask8)(U)); })
+
+#define _mm256_ternarylogic_epi32(A, B, C, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_pternlogd256_mask((__v8si)(__m256i)(A), \
+                                            (__v8si)(__m256i)(B), \
+                                            (__v8si)(__m256i)(C), (int)(imm), \
+                                            (__mmask8)-1); })
+
+#define _mm256_mask_ternarylogic_epi32(A, U, B, C, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_pternlogd256_mask((__v8si)(__m256i)(A), \
+                                            (__v8si)(__m256i)(B), \
+                                            (__v8si)(__m256i)(C), (int)(imm), \
+                                            (__mmask8)(U)); })
+
+#define _mm256_maskz_ternarylogic_epi32(U, A, B, C, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_pternlogd256_maskz((__v8si)(__m256i)(A), \
+                                             (__v8si)(__m256i)(B), \
+                                             (__v8si)(__m256i)(C), (int)(imm), \
+                                             (__mmask8)(U)); })
+
+#define _mm_ternarylogic_epi64(A, B, C, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_pternlogq128_mask((__v2di)(__m128i)(A), \
+                                            (__v2di)(__m128i)(B), \
+                                            (__v2di)(__m128i)(C), (int)(imm), \
+                                            (__mmask8)-1); })
+
+#define _mm_mask_ternarylogic_epi64(A, U, B, C, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_pternlogq128_mask((__v2di)(__m128i)(A), \
+                                            (__v2di)(__m128i)(B), \
+                                            (__v2di)(__m128i)(C), (int)(imm), \
+                                            (__mmask8)(U)); })
+
+#define _mm_maskz_ternarylogic_epi64(U, A, B, C, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_pternlogq128_maskz((__v2di)(__m128i)(A), \
+                                             (__v2di)(__m128i)(B), \
+                                             (__v2di)(__m128i)(C), (int)(imm), \
+                                             (__mmask8)(U)); })
+
+#define _mm256_ternarylogic_epi64(A, B, C, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_pternlogq256_mask((__v4di)(__m256i)(A), \
+                                            (__v4di)(__m256i)(B), \
+                                            (__v4di)(__m256i)(C), (int)(imm), \
+                                            (__mmask8)-1); })
+
+#define _mm256_mask_ternarylogic_epi64(A, U, B, C, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_pternlogq256_mask((__v4di)(__m256i)(A), \
+                                            (__v4di)(__m256i)(B), \
+                                            (__v4di)(__m256i)(C), (int)(imm), \
+                                            (__mmask8)(U)); })
+
+#define _mm256_maskz_ternarylogic_epi64(U, A, B, C, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_pternlogq256_maskz((__v4di)(__m256i)(A), \
+                                             (__v4di)(__m256i)(B), \
+                                             (__v4di)(__m256i)(C), (int)(imm), \
+                                             (__mmask8)(U)); })
+
+
+
+#define _mm256_shuffle_f32x4(A, B, imm) __extension__ ({ \
+  (__m256)__builtin_ia32_shuf_f32x4_256_mask((__v8sf)(__m256)(A), \
+                                             (__v8sf)(__m256)(B), (int)(imm), \
+                                             (__v8sf)_mm256_setzero_ps(), \
+                                             (__mmask8)-1); })
+
+#define _mm256_mask_shuffle_f32x4(W, U, A, B, imm) __extension__ ({ \
+  (__m256)__builtin_ia32_shuf_f32x4_256_mask((__v8sf)(__m256)(A), \
+                                             (__v8sf)(__m256)(B), (int)(imm), \
+                                             (__v8sf)(__m256)(W), \
+                                             (__mmask8)(U)); })
+
+#define _mm256_maskz_shuffle_f32x4(U, A, B, imm) __extension__ ({ \
+  (__m256)__builtin_ia32_shuf_f32x4_256_mask((__v8sf)(__m256)(A), \
+                                             (__v8sf)(__m256)(B), (int)(imm), \
+                                             (__v8sf)_mm256_setzero_ps(), \
+                                             (__mmask8)(U)); })
+
+#define _mm256_shuffle_f64x2(A, B, imm) __extension__ ({ \
+  (__m256d)__builtin_ia32_shuf_f64x2_256_mask((__v4df)(__m256d)(A), \
+                                              (__v4df)(__m256d)(B), \
+                                              (int)(imm), \
+                                              (__v4df)_mm256_setzero_pd(), \
+                                              (__mmask8)-1); })
+
+#define _mm256_mask_shuffle_f64x2(W, U, A, B, imm) __extension__ ({ \
+  (__m256d)__builtin_ia32_shuf_f64x2_256_mask((__v4df)(__m256d)(A), \
+                                              (__v4df)(__m256d)(B), \
+                                              (int)(imm), \
+                                              (__v4df)(__m256d)(W), \
+                                              (__mmask8)(U)); })
+
+#define _mm256_maskz_shuffle_f64x2(U, A, B, imm) __extension__ ({ \
+  (__m256d)__builtin_ia32_shuf_f64x2_256_mask((__v4df)(__m256d)(A), \
+                                              (__v4df)(__m256d)(B), \
+                                              (int)(imm), \
+                                              (__v4df)_mm256_setzero_pd(), \
+                                              (__mmask8)(U)); })
+
+#define _mm256_shuffle_i32x4(A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_shuf_i32x4_256_mask((__v8si)(__m256i)(A), \
+                                              (__v8si)(__m256i)(B), \
+                                              (int)(imm), \
+                                              (__v8si)_mm256_setzero_si256(), \
+                                              (__mmask8)-1); })
+
+#define _mm256_mask_shuffle_i32x4(W, U, A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_shuf_i32x4_256_mask((__v8si)(__m256i)(A), \
+                                              (__v8si)(__m256i)(B), \
+                                              (int)(imm), \
+                                              (__v8si)(__m256i)(W), \
+                                              (__mmask8)(U)); })
+
+#define _mm256_maskz_shuffle_i32x4(U, A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_shuf_i32x4_256_mask((__v8si)(__m256i)(A), \
+                                              (__v8si)(__m256i)(B), \
+                                              (int)(imm), \
+                                              (__v8si)_mm256_setzero_si256(), \
+                                              (__mmask8)(U)); })
+
+#define _mm256_shuffle_i64x2(A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_shuf_i64x2_256_mask((__v4di)(__m256i)(A), \
+                                              (__v4di)(__m256i)(B), \
+                                              (int)(imm), \
+                                              (__v4di)_mm256_setzero_si256(), \
+                                              (__mmask8)-1); })
+
+#define _mm256_mask_shuffle_i64x2(W, U, A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_shuf_i64x2_256_mask((__v4di)(__m256i)(A), \
+                                              (__v4di)(__m256i)(B), \
+                                              (int)(imm), \
+                                              (__v4di)(__m256i)(W), \
+                                              (__mmask8)(U)); })
+
+#define _mm256_maskz_shuffle_i64x2(U, A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_shuf_i64x2_256_mask((__v4di)(__m256i)(A), \
+                                              (__v4di)(__m256i)(B), \
+                                              (int)(imm), \
+                                              (__v4di)_mm256_setzero_si256(), \
+                                              (__mmask8)(U)); })
+
+#define _mm_mask_shuffle_pd(W, U, A, B, M) __extension__ ({ \
+  (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \
+                                       (__v2df)_mm_shuffle_pd((A), (B), (M)), \
+                                       (__v2df)(__m128d)(W)); })
+
+#define _mm_maskz_shuffle_pd(U, A, B, M) __extension__ ({ \
+  (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \
+                                       (__v2df)_mm_shuffle_pd((A), (B), (M)), \
+                                       (__v2df)_mm_setzero_pd()); })
+
+#define _mm256_mask_shuffle_pd(W, U, A, B, M) __extension__ ({ \
+  (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \
+                                       (__v4df)_mm256_shuffle_pd((A), (B), (M)), \
+                                       (__v4df)(__m256d)(W)); })
+
+#define _mm256_maskz_shuffle_pd(U, A, B, M) __extension__ ({ \
+  (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \
+                                       (__v4df)_mm256_shuffle_pd((A), (B), (M)), \
+                                       (__v4df)_mm256_setzero_pd()); })
+
+#define _mm_mask_shuffle_ps(W, U, A, B, M) __extension__ ({ \
+  (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \
+                                      (__v4sf)_mm_shuffle_ps((A), (B), (M)), \
+                                      (__v4sf)(__m128)(W)); })
+
+#define _mm_maskz_shuffle_ps(U, A, B, M) __extension__ ({ \
+  (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \
+                                      (__v4sf)_mm_shuffle_ps((A), (B), (M)), \
+                                      (__v4sf)_mm_setzero_ps()); })
+
+#define _mm256_mask_shuffle_ps(W, U, A, B, M) __extension__ ({ \
+  (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \
+                                      (__v8sf)_mm256_shuffle_ps((A), (B), (M)), \
+                                      (__v8sf)(__m256)(W)); })
+
+#define _mm256_maskz_shuffle_ps(U, A, B, M) __extension__ ({ \
+  (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \
+                                      (__v8sf)_mm256_shuffle_ps((A), (B), (M)), \
+                                      (__v8sf)_mm256_setzero_ps()); })
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_rsqrt14_pd (__m128d __A)
+{
+  return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A,
+                 (__v2df)
+                 _mm_setzero_pd (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_rsqrt14_pd (__m128d __W, __mmask8 __U, __m128d __A)
+{
+  return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A,
+                 (__v2df) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_rsqrt14_pd (__mmask8 __U, __m128d __A)
+{
+  return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A,
+                 (__v2df)
+                 _mm_setzero_pd (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_rsqrt14_pd (__m256d __A)
+{
+  return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A,
+                 (__v4df)
+                 _mm256_setzero_pd (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_rsqrt14_pd (__m256d __W, __mmask8 __U, __m256d __A)
+{
+  return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A,
+                 (__v4df) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_rsqrt14_pd (__mmask8 __U, __m256d __A)
+{
+  return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A,
+                 (__v4df)
+                 _mm256_setzero_pd (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_rsqrt14_ps (__m128 __A)
+{
+  return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A,
+                (__v4sf)
+                _mm_setzero_ps (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_rsqrt14_ps (__m128 __W, __mmask8 __U, __m128 __A)
+{
+  return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A,
+                (__v4sf) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_rsqrt14_ps (__mmask8 __U, __m128 __A)
+{
+  return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A,
+                (__v4sf)
+                _mm_setzero_ps (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_rsqrt14_ps (__m256 __A)
+{
+  return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
+                (__v8sf)
+                _mm256_setzero_ps (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_rsqrt14_ps (__m256 __W, __mmask8 __U, __m256 __A)
+{
+  return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
+                (__v8sf) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_rsqrt14_ps (__mmask8 __U, __m256 __A)
+{
+  return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
+                (__v8sf)
+                _mm256_setzero_ps (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_broadcast_f32x4 (__m128 __A)
+{
+  return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A,
+                (__v8sf)_mm256_undefined_pd (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_broadcast_f32x4 (__m256 __O, __mmask8 __M, __m128 __A)
+{
+  return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A,
+                (__v8sf) __O,
+                __M);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_broadcast_f32x4 (__mmask8 __M, __m128 __A)
+{
+  return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A,
+                (__v8sf) _mm256_setzero_ps (),
+                __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_broadcast_i32x4 (__m128i __A)
+{
+  return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si) __A,
+                 (__v8si)_mm256_undefined_si256 (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_broadcast_i32x4 (__m256i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si) __A,
+                 (__v8si)
+                 __O, __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_broadcast_i32x4 (__mmask8 __M, __m128i __A)
+{
+  return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si)
+                 __A,
+                 (__v8si) _mm256_setzero_si256 (),
+                 __M);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_broadcastsd_pd (__m256d __O, __mmask8 __M, __m128d __A)
+{
+  return (__m256d)__builtin_ia32_selectpd_256(__M,
+                                              (__v4df) _mm256_broadcastsd_pd(__A),
+                                              (__v4df) __O);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_broadcastsd_pd (__mmask8 __M, __m128d __A)
+{
+  return (__m256d)__builtin_ia32_selectpd_256(__M,
+                                              (__v4df) _mm256_broadcastsd_pd(__A),
+                                              (__v4df) _mm256_setzero_pd());
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_broadcastss_ps (__m128 __O, __mmask8 __M, __m128 __A)
+{
+  return (__m128)__builtin_ia32_selectps_128(__M,
+                                             (__v4sf) _mm_broadcastss_ps(__A),
+                                             (__v4sf) __O);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_broadcastss_ps (__mmask8 __M, __m128 __A)
+{
+  return (__m128)__builtin_ia32_selectps_128(__M,
+                                             (__v4sf) _mm_broadcastss_ps(__A),
+                                             (__v4sf) _mm_setzero_ps());
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_broadcastss_ps (__m256 __O, __mmask8 __M, __m128 __A)
+{
+  return (__m256)__builtin_ia32_selectps_256(__M,
+                                             (__v8sf) _mm256_broadcastss_ps(__A),
+                                             (__v8sf) __O);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_broadcastss_ps (__mmask8 __M, __m128 __A)
+{
+  return (__m256)__builtin_ia32_selectps_256(__M,
+                                             (__v8sf) _mm256_broadcastss_ps(__A),
+                                             (__v8sf) _mm256_setzero_ps());
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_broadcastd_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i)__builtin_ia32_selectd_128(__M,
+                                             (__v4si) _mm_broadcastd_epi32(__A),
+                                             (__v4si) __O);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i)__builtin_ia32_selectd_128(__M,
+                                             (__v4si) _mm_broadcastd_epi32(__A),
+                                             (__v4si) _mm_setzero_si128());
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_broadcastd_epi32 (__m256i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m256i)__builtin_ia32_selectd_256(__M,
+                                             (__v8si) _mm256_broadcastd_epi32(__A),
+                                             (__v8si) __O);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A)
+{
+  return (__m256i)__builtin_ia32_selectd_256(__M,
+                                             (__v8si) _mm256_broadcastd_epi32(__A),
+                                             (__v8si) _mm256_setzero_si256());
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_broadcastq_epi64 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i)__builtin_ia32_selectq_128(__M,
+                                             (__v2di) _mm_broadcastq_epi64(__A),
+                                             (__v2di) __O);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i)__builtin_ia32_selectq_128(__M,
+                                             (__v2di) _mm_broadcastq_epi64(__A),
+                                             (__v2di) _mm_setzero_si128());
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_broadcastq_epi64 (__m256i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m256i)__builtin_ia32_selectq_256(__M,
+                                             (__v4di) _mm256_broadcastq_epi64(__A),
+                                             (__v4di) __O);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A)
+{
+  return (__m256i)__builtin_ia32_selectq_256(__M,
+                                             (__v4di) _mm256_broadcastq_epi64(__A),
+                                             (__v4di) _mm256_setzero_si256());
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtsepi32_epi8 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A,
+               (__v16qi)_mm_undefined_si128(),
+               (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A,
+               (__v16qi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtsepi32_epi8 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A,
+               (__v16qi) _mm_setzero_si128 (),
+               __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovsdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtsepi32_epi8 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A,
+               (__v16qi)_mm_undefined_si128(),
+               (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A,
+               (__v16qi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtsepi32_epi8 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A,
+               (__v16qi) _mm_setzero_si128 (),
+               __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovsdb256mem_mask ((__v16qi *) __P, (__v8si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtsepi32_epi16 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A,
+               (__v8hi)_mm_setzero_si128 (),
+               (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A,
+               (__v8hi)__O,
+               __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtsepi32_epi16 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A,
+               (__v8hi) _mm_setzero_si128 (),
+               __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovsdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtsepi32_epi16 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A,
+               (__v8hi)_mm_undefined_si128(),
+               (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A,
+               (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtsepi32_epi16 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A,
+               (__v8hi) _mm_setzero_si128 (),
+               __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovsdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtsepi64_epi8 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A,
+               (__v16qi)_mm_undefined_si128(),
+               (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A,
+               (__v16qi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtsepi64_epi8 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A,
+               (__v16qi) _mm_setzero_si128 (),
+               __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovsqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtsepi64_epi8 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A,
+               (__v16qi)_mm_undefined_si128(),
+               (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A,
+               (__v16qi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtsepi64_epi8 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A,
+               (__v16qi) _mm_setzero_si128 (),
+               __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovsqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtsepi64_epi32 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A,
+               (__v4si)_mm_undefined_si128(),
+               (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A,
+               (__v4si) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtsepi64_epi32 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A,
+               (__v4si) _mm_setzero_si128 (),
+               __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovsqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtsepi64_epi32 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A,
+               (__v4si)_mm_undefined_si128(),
+               (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A,
+               (__v4si)__O,
+               __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtsepi64_epi32 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A,
+               (__v4si) _mm_setzero_si128 (),
+               __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovsqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtsepi64_epi16 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A,
+               (__v8hi)_mm_undefined_si128(),
+               (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A,
+               (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtsepi64_epi16 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A,
+               (__v8hi) _mm_setzero_si128 (),
+               __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovsqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtsepi64_epi16 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A,
+               (__v8hi)_mm_undefined_si128(),
+               (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A,
+               (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtsepi64_epi16 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A,
+               (__v8hi) _mm_setzero_si128 (),
+               __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovsqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtusepi32_epi8 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A,
+                (__v16qi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A,
+                (__v16qi) __O,
+                __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtusepi32_epi8 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A,
+                (__v16qi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovusdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtusepi32_epi8 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A,
+                (__v16qi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A,
+                (__v16qi) __O,
+                __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtusepi32_epi8 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A,
+                (__v16qi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovusdb256mem_mask ((__v16qi*) __P, (__v8si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtusepi32_epi16 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A,
+                (__v8hi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A,
+                (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtusepi32_epi16 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A,
+                (__v8hi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovusdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtusepi32_epi16 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A,
+                (__v8hi) _mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A,
+                (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtusepi32_epi16 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A,
+                (__v8hi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovusdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtusepi64_epi8 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A,
+                (__v16qi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A,
+                (__v16qi) __O,
+                __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtusepi64_epi8 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A,
+                (__v16qi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovusqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtusepi64_epi8 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A,
+                (__v16qi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A,
+                (__v16qi) __O,
+                __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtusepi64_epi8 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A,
+                (__v16qi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovusqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtusepi64_epi32 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A,
+                (__v4si)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A,
+                (__v4si) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtusepi64_epi32 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A,
+                (__v4si) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovusqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtusepi64_epi32 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A,
+                (__v4si)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A,
+                (__v4si) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtusepi64_epi32 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A,
+                (__v4si) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovusqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtusepi64_epi16 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A,
+                (__v8hi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A,
+                (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtusepi64_epi16 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A,
+                (__v8hi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovusqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtusepi64_epi16 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A,
+                (__v8hi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A,
+                (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtusepi64_epi16 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A,
+                (__v8hi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
+{
+  return __builtin_ia32_pmovusqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtepi32_epi8 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A,
+              (__v16qi)_mm_undefined_si128(),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A,
+              (__v16qi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepi32_epi8 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A,
+              (__v16qi)
+              _mm_setzero_si128 (),
+              __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtepi32_epi8 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
+              (__v16qi)_mm_undefined_si128(),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
+              (__v16qi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepi32_epi8 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
+              (__v16qi) _mm_setzero_si128 (),
+              __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovdb256mem_mask ((__v16qi *) __P, (__v8si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtepi32_epi16 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A,
+              (__v8hi) _mm_setzero_si128 (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A,
+              (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepi32_epi16 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A,
+              (__v8hi) _mm_setzero_si128 (),
+              __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtepi32_epi16 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A,
+              (__v8hi)_mm_setzero_si128 (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A,
+              (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepi32_epi16 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A,
+              (__v8hi) _mm_setzero_si128 (),
+              __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi32_storeu_epi16 (void *  __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtepi64_epi8 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A,
+              (__v16qi) _mm_undefined_si128(),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A,
+              (__v16qi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepi64_epi8 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A,
+              (__v16qi) _mm_setzero_si128 (),
+              __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtepi64_epi8 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A,
+              (__v16qi) _mm_undefined_si128(),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A,
+              (__v16qi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepi64_epi8 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A,
+              (__v16qi) _mm_setzero_si128 (),
+              __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtepi64_epi32 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A,
+              (__v4si)_mm_undefined_si128(),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A,
+              (__v4si) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepi64_epi32 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A,
+              (__v4si) _mm_setzero_si128 (),
+              __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtepi64_epi32 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A,
+              (__v4si) _mm_undefined_si128(),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A,
+              (__v4si) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepi64_epi32 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A,
+              (__v4si) _mm_setzero_si128 (),
+              __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtepi64_epi16 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A,
+              (__v8hi) _mm_undefined_si128(),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A,
+              (__v8hi)__O,
+              __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtepi64_epi16 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A,
+              (__v8hi) _mm_setzero_si128 (),
+              __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtepi64_epi16 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A,
+              (__v8hi)_mm_undefined_si128(),
+              (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A,
+              (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtepi64_epi16 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A,
+              (__v8hi) _mm_setzero_si128 (),
+              __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M);
+}
+
+#define _mm256_extractf32x4_ps(A, imm) __extension__ ({ \
+  (__m128)__builtin_ia32_extractf32x4_256_mask((__v8sf)(__m256)(A), \
+                                               (int)(imm), \
+                                               (__v4sf)_mm_setzero_ps(), \
+                                               (__mmask8)-1); })
+
+#define _mm256_mask_extractf32x4_ps(W, U, A, imm) __extension__ ({ \
+  (__m128)__builtin_ia32_extractf32x4_256_mask((__v8sf)(__m256)(A), \
+                                               (int)(imm), \
+                                               (__v4sf)(__m128)(W), \
+                                               (__mmask8)(U)); })
+
+#define _mm256_maskz_extractf32x4_ps(U, A, imm) __extension__ ({ \
+  (__m128)__builtin_ia32_extractf32x4_256_mask((__v8sf)(__m256)(A), \
+                                               (int)(imm), \
+                                               (__v4sf)_mm_setzero_ps(), \
+                                               (__mmask8)(U)); })
+
+#define _mm256_extracti32x4_epi32(A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_extracti32x4_256_mask((__v8si)(__m256i)(A), \
+                                                (int)(imm), \
+                                                (__v4si)_mm_setzero_si128(), \
+                                                (__mmask8)-1); })
+
+#define _mm256_mask_extracti32x4_epi32(W, U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_extracti32x4_256_mask((__v8si)(__m256i)(A), \
+                                                (int)(imm), \
+                                                (__v4si)(__m128i)(W), \
+                                                (__mmask8)(U)); })
+
+#define _mm256_maskz_extracti32x4_epi32(U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_extracti32x4_256_mask((__v8si)(__m256i)(A), \
+                                                (int)(imm), \
+                                                (__v4si)_mm_setzero_si128(), \
+                                                (__mmask8)(U)); })
+
+#define _mm256_insertf32x4(A, B, imm) __extension__ ({ \
+  (__m256)__builtin_ia32_insertf32x4_256_mask((__v8sf)(__m256)(A), \
+                                              (__v4sf)(__m128)(B), (int)(imm), \
+                                              (__v8sf)_mm256_setzero_ps(), \
+                                              (__mmask8)-1); })
+
+#define _mm256_mask_insertf32x4(W, U, A, B, imm) __extension__ ({ \
+  (__m256)__builtin_ia32_insertf32x4_256_mask((__v8sf)(__m256)(A), \
+                                              (__v4sf)(__m128)(B), (int)(imm), \
+                                              (__v8sf)(__m256)(W), \
+                                              (__mmask8)(U)); })
+
+#define _mm256_maskz_insertf32x4(U, A, B, imm) __extension__ ({ \
+  (__m256)__builtin_ia32_insertf32x4_256_mask((__v8sf)(__m256)(A), \
+                                              (__v4sf)(__m128)(B), (int)(imm), \
+                                              (__v8sf)_mm256_setzero_ps(), \
+                                              (__mmask8)(U)); })
+
+#define _mm256_inserti32x4(A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_inserti32x4_256_mask((__v8si)(__m256i)(A), \
+                                               (__v4si)(__m128i)(B), \
+                                               (int)(imm), \
+                                               (__v8si)_mm256_setzero_si256(), \
+                                               (__mmask8)-1); })
+
+#define _mm256_mask_inserti32x4(W, U, A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_inserti32x4_256_mask((__v8si)(__m256i)(A), \
+                                               (__v4si)(__m128i)(B), \
+                                               (int)(imm), \
+                                               (__v8si)(__m256i)(W), \
+                                               (__mmask8)(U)); })
+
+#define _mm256_maskz_inserti32x4(U, A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_inserti32x4_256_mask((__v8si)(__m256i)(A), \
+                                               (__v4si)(__m128i)(B), \
+                                               (int)(imm), \
+                                               (__v8si)_mm256_setzero_si256(), \
+                                               (__mmask8)(U)); })
+
+#define _mm_getmant_pd(A, B, C) __extension__({\
+  (__m128d)__builtin_ia32_getmantpd128_mask((__v2df)(__m128d)(A), \
+                                            (int)(((C)<<2) | (B)), \
+                                            (__v2df)_mm_setzero_pd(), \
+                                            (__mmask8)-1); })
+
+#define _mm_mask_getmant_pd(W, U, A, B, C) __extension__({\
+  (__m128d)__builtin_ia32_getmantpd128_mask((__v2df)(__m128d)(A), \
+                                            (int)(((C)<<2) | (B)), \
+                                            (__v2df)(__m128d)(W), \
+                                            (__mmask8)(U)); })
+
+#define _mm_maskz_getmant_pd(U, A, B, C) __extension__({\
+  (__m128d)__builtin_ia32_getmantpd128_mask((__v2df)(__m128d)(A), \
+                                            (int)(((C)<<2) | (B)), \
+                                            (__v2df)_mm_setzero_pd(), \
+                                            (__mmask8)(U)); })
+
+#define _mm256_getmant_pd(A, B, C) __extension__ ({ \
+  (__m256d)__builtin_ia32_getmantpd256_mask((__v4df)(__m256d)(A), \
+                                            (int)(((C)<<2) | (B)), \
+                                            (__v4df)_mm256_setzero_pd(), \
+                                            (__mmask8)-1); })
+
+#define _mm256_mask_getmant_pd(W, U, A, B, C) __extension__ ({ \
+  (__m256d)__builtin_ia32_getmantpd256_mask((__v4df)(__m256d)(A), \
+                                            (int)(((C)<<2) | (B)), \
+                                            (__v4df)(__m256d)(W), \
+                                            (__mmask8)(U)); })
+
+#define _mm256_maskz_getmant_pd(U, A, B, C) __extension__ ({ \
+  (__m256d)__builtin_ia32_getmantpd256_mask((__v4df)(__m256d)(A), \
+                                            (int)(((C)<<2) | (B)), \
+                                            (__v4df)_mm256_setzero_pd(), \
+                                            (__mmask8)(U)); })
+
+#define _mm_getmant_ps(A, B, C) __extension__ ({ \
+  (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \
+                                           (int)(((C)<<2) | (B)), \
+                                           (__v4sf)_mm_setzero_ps(), \
+                                           (__mmask8)-1); })
+
+#define _mm_mask_getmant_ps(W, U, A, B, C) __extension__ ({ \
+  (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \
+                                           (int)(((C)<<2) | (B)), \
+                                           (__v4sf)(__m128)(W), \
+                                           (__mmask8)(U)); })
+
+#define _mm_maskz_getmant_ps(U, A, B, C) __extension__ ({ \
+  (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \
+                                           (int)(((C)<<2) | (B)), \
+                                           (__v4sf)_mm_setzero_ps(), \
+                                           (__mmask8)(U)); })
+
+#define _mm256_getmant_ps(A, B, C) __extension__ ({ \
+  (__m256)__builtin_ia32_getmantps256_mask((__v8sf)(__m256)(A), \
+                                           (int)(((C)<<2) | (B)), \
+                                           (__v8sf)_mm256_setzero_ps(), \
+                                           (__mmask8)-1); })
+
+#define _mm256_mask_getmant_ps(W, U, A, B, C) __extension__ ({ \
+  (__m256)__builtin_ia32_getmantps256_mask((__v8sf)(__m256)(A), \
+                                           (int)(((C)<<2) | (B)), \
+                                           (__v8sf)(__m256)(W), \
+                                           (__mmask8)(U)); })
+
+#define _mm256_maskz_getmant_ps(U, A, B, C) __extension__ ({ \
+  (__m256)__builtin_ia32_getmantps256_mask((__v8sf)(__m256)(A), \
+                                           (int)(((C)<<2) | (B)), \
+                                           (__v8sf)_mm256_setzero_ps(), \
+                                           (__mmask8)(U)); })
+
+#define _mm_mmask_i64gather_pd(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m128d)__builtin_ia32_gather3div2df((__v2df)(__m128d)(v1_old), \
+                                        (double const *)(addr), \
+                                        (__v2di)(__m128i)(index), \
+                                        (__mmask8)(mask), (int)(scale)); })
+
+#define _mm_mmask_i64gather_epi64(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m128i)__builtin_ia32_gather3div2di((__v2di)(__m128i)(v1_old), \
+                                        (long long const *)(addr), \
+                                        (__v2di)(__m128i)(index), \
+                                        (__mmask8)(mask), (int)(scale)); })
+
+#define _mm256_mmask_i64gather_pd(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m256d)__builtin_ia32_gather3div4df((__v4df)(__m256d)(v1_old), \
+                                        (double const *)(addr), \
+                                        (__v4di)(__m256i)(index), \
+                                        (__mmask8)(mask), (int)(scale)); })
+
+#define _mm256_mmask_i64gather_epi64(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m256i)__builtin_ia32_gather3div4di((__v4di)(__m256i)(v1_old), \
+                                        (long long const *)(addr), \
+                                        (__v4di)(__m256i)(index), \
+                                        (__mmask8)(mask), (int)(scale)); })
+
+#define _mm_mmask_i64gather_ps(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m128)__builtin_ia32_gather3div4sf((__v4sf)(__m128)(v1_old), \
+                                       (float const *)(addr), \
+                                       (__v2di)(__m128i)(index), \
+                                       (__mmask8)(mask), (int)(scale)); })
+
+#define _mm_mmask_i64gather_epi32(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m128i)__builtin_ia32_gather3div4si((__v4si)(__m128i)(v1_old), \
+                                        (int const *)(addr), \
+                                        (__v2di)(__m128i)(index), \
+                                        (__mmask8)(mask), (int)(scale)); })
+
+#define _mm256_mmask_i64gather_ps(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m128)__builtin_ia32_gather3div8sf((__v4sf)(__m128)(v1_old), \
+                                       (float const *)(addr), \
+                                       (__v4di)(__m256i)(index), \
+                                       (__mmask8)(mask), (int)(scale)); })
+
+#define _mm256_mmask_i64gather_epi32(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m128i)__builtin_ia32_gather3div8si((__v4si)(__m128i)(v1_old), \
+                                        (int const *)(addr), \
+                                        (__v4di)(__m256i)(index), \
+                                        (__mmask8)(mask), (int)(scale)); })
+
+#define _mm_mmask_i32gather_pd(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m128d)__builtin_ia32_gather3siv2df((__v2df)(__m128d)(v1_old), \
+                                        (double const *)(addr), \
+                                        (__v4si)(__m128i)(index), \
+                                        (__mmask8)(mask), (int)(scale)); })
+
+#define _mm_mmask_i32gather_epi64(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m128i)__builtin_ia32_gather3siv2di((__v2di)(__m128i)(v1_old), \
+                                        (long long const *)(addr), \
+                                        (__v4si)(__m128i)(index), \
+                                        (__mmask8)(mask), (int)(scale)); })
+
+#define _mm256_mmask_i32gather_pd(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m256d)__builtin_ia32_gather3siv4df((__v4df)(__m256d)(v1_old), \
+                                        (double const *)(addr), \
+                                        (__v4si)(__m128i)(index), \
+                                        (__mmask8)(mask), (int)(scale)); })
+
+#define _mm256_mmask_i32gather_epi64(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m256i)__builtin_ia32_gather3siv4di((__v4di)(__m256i)(v1_old), \
+                                        (long long const *)(addr), \
+                                        (__v4si)(__m128i)(index), \
+                                        (__mmask8)(mask), (int)(scale)); })
+
+#define _mm_mmask_i32gather_ps(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m128)__builtin_ia32_gather3siv4sf((__v4sf)(__m128)(v1_old), \
+                                       (float const *)(addr), \
+                                       (__v4si)(__m128i)(index), \
+                                       (__mmask8)(mask), (int)(scale)); })
+
+#define _mm_mmask_i32gather_epi32(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m128i)__builtin_ia32_gather3siv4si((__v4si)(__m128i)(v1_old), \
+                                        (int const *)(addr), \
+                                        (__v4si)(__m128i)(index), \
+                                        (__mmask8)(mask), (int)(scale)); })
+
+#define _mm256_mmask_i32gather_ps(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m256)__builtin_ia32_gather3siv8sf((__v8sf)(__m256)(v1_old), \
+                                       (float const *)(addr), \
+                                       (__v8si)(__m256i)(index), \
+                                       (__mmask8)(mask), (int)(scale)); })
+
+#define _mm256_mmask_i32gather_epi32(v1_old, mask, index, addr, scale) __extension__ ({\
+  (__m256i)__builtin_ia32_gather3siv8si((__v8si)(__m256i)(v1_old), \
+                                        (int const *)(addr), \
+                                        (__v8si)(__m256i)(index), \
+                                        (__mmask8)(mask), (int)(scale)); })
+
+#define _mm256_permutex_pd(X, C) __extension__ ({ \
+  (__m256d)__builtin_shufflevector((__v4df)(__m256d)(X), \
+                                   (__v4df)_mm256_undefined_pd(), \
+                                   ((C) >> 0) & 0x3, ((C) >> 2) & 0x3, \
+                                   ((C) >> 4) & 0x3, ((C) >> 6) & 0x3); })
+
+#define _mm256_mask_permutex_pd(W, U, X, C) __extension__ ({ \
+  (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \
+                                       (__v4df)_mm256_permutex_pd((X), (C)), \
+                                       (__v4df)(__m256d)(W)); })
+
+#define _mm256_maskz_permutex_pd(U, X, C) __extension__ ({ \
+  (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \
+                                       (__v4df)_mm256_permutex_pd((X), (C)), \
+                                       (__v4df)_mm256_setzero_pd()); })
+
+#define _mm256_permutex_epi64(X, C) __extension__ ({ \
+  (__m256i)__builtin_shufflevector((__v4di)(__m256i)(X), \
+                                   (__v4di)_mm256_undefined_si256(), \
+                                   ((C) >> 0) & 0x3, ((C) >> 2) & 0x3, \
+                                   ((C) >> 4) & 0x3, ((C) >> 6) & 0x3); })
+
+#define _mm256_mask_permutex_epi64(W, U, X, C) __extension__ ({ \
+  (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \
+                                      (__v4di)_mm256_permutex_epi64((X), (C)), \
+                                      (__v4di)(__m256i)(W)); })
+
+#define _mm256_maskz_permutex_epi64(U, X, C) __extension__ ({ \
+  (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \
+                                      (__v4di)_mm256_permutex_epi64((X), (C)), \
+                                      (__v4di)_mm256_setzero_si256()); })
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_permutexvar_pd (__m256i __X, __m256d __Y)
+{
+  return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y,
+                 (__v4di) __X,
+                 (__v4df) _mm256_undefined_si256 (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_permutexvar_pd (__m256d __W, __mmask8 __U, __m256i __X,
+          __m256d __Y)
+{
+  return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y,
+                 (__v4di) __X,
+                 (__v4df) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_permutexvar_pd (__mmask8 __U, __m256i __X, __m256d __Y)
+{
+  return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y,
+                 (__v4di) __X,
+                 (__v4df) _mm256_setzero_pd (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_permutexvar_epi64 (__mmask8 __M, __m256i __X, __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y,
+                 (__v4di) __X,
+                 (__v4di) _mm256_setzero_si256 (),
+                 (__mmask8) __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_permutexvar_epi64 ( __m256i __X, __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y,
+                 (__v4di) __X,
+                 (__v4di) _mm256_undefined_si256 (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_permutexvar_epi64 (__m256i __W, __mmask8 __M, __m256i __X,
+             __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y,
+                 (__v4di) __X,
+                 (__v4di) __W,
+                 __M);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_permutexvar_ps (__m256 __W, __mmask8 __U, __m256i __X,
+          __m256 __Y)
+{
+  return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y,
+                (__v8si) __X,
+                (__v8sf) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_permutexvar_ps (__mmask8 __U, __m256i __X, __m256 __Y)
+{
+  return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y,
+                (__v8si) __X,
+                (__v8sf) _mm256_setzero_ps (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_permutexvar_ps (__m256i __X, __m256 __Y)
+{
+  return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y,
+                (__v8si) __X,
+                (__v8sf) _mm256_undefined_si256 (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_permutexvar_epi32 (__mmask8 __M, __m256i __X, __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y,
+                 (__v8si) __X,
+                 (__v8si) _mm256_setzero_si256 (),
+                 __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_permutexvar_epi32 (__m256i __W, __mmask8 __M, __m256i __X,
+             __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y,
+                 (__v8si) __X,
+                 (__v8si) __W,
+                 (__mmask8) __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_permutexvar_epi32 (__m256i __X, __m256i __Y)
+{
+  return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y,
+                 (__v8si) __X,
+                 (__v8si) _mm256_undefined_si256(),
+                 (__mmask8) -1);
+}
+
+#define _mm_alignr_epi32(A, B, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_alignd128_mask((__v4si)(__m128i)(A), \
+                                         (__v4si)(__m128i)(B), (int)(imm), \
+                                         (__v4si)_mm_undefined_si128(), \
+                                         (__mmask8)-1); })
+
+#define _mm_mask_alignr_epi32(W, U, A, B, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_alignd128_mask((__v4si)(__m128i)(A), \
+                                         (__v4si)(__m128i)(B), (int)(imm), \
+                                         (__v4si)(__m128i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm_maskz_alignr_epi32(U, A, B, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_alignd128_mask((__v4si)(__m128i)(A), \
+                                         (__v4si)(__m128i)(B), (int)(imm), \
+                                         (__v4si)_mm_setzero_si128(), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_alignr_epi32(A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_alignd256_mask((__v8si)(__m256i)(A), \
+                                         (__v8si)(__m256i)(B), (int)(imm), \
+                                         (__v8si)_mm256_undefined_si256(), \
+                                         (__mmask8)-1); })
+
+#define _mm256_mask_alignr_epi32(W, U, A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_alignd256_mask((__v8si)(__m256i)(A), \
+                                         (__v8si)(__m256i)(B), (int)(imm), \
+                                         (__v8si)(__m256i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_maskz_alignr_epi32(U, A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_alignd256_mask((__v8si)(__m256i)(A), \
+                                         (__v8si)(__m256i)(B), (int)(imm), \
+                                         (__v8si)_mm256_setzero_si256(), \
+                                         (__mmask8)(U)); })
+
+#define _mm_alignr_epi64(A, B, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_alignq128_mask((__v2di)(__m128i)(A), \
+                                         (__v2di)(__m128i)(B), (int)(imm), \
+                                         (__v2di)_mm_setzero_di(), \
+                                         (__mmask8)-1); })
+
+#define _mm_mask_alignr_epi64(W, U, A, B, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_alignq128_mask((__v2di)(__m128i)(A), \
+                                         (__v2di)(__m128i)(B), (int)(imm), \
+                                         (__v2di)(__m128i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm_maskz_alignr_epi64(U, A, B, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_alignq128_mask((__v2di)(__m128i)(A), \
+                                         (__v2di)(__m128i)(B), (int)(imm), \
+                                         (__v2di)_mm_setzero_di(), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_alignr_epi64(A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_alignq256_mask((__v4di)(__m256i)(A), \
+                                         (__v4di)(__m256i)(B), (int)(imm), \
+                                         (__v4di)_mm256_undefined_pd(), \
+                                         (__mmask8)-1); })
+
+#define _mm256_mask_alignr_epi64(W, U, A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_alignq256_mask((__v4di)(__m256i)(A), \
+                                         (__v4di)(__m256i)(B), (int)(imm), \
+                                         (__v4di)(__m256i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm256_maskz_alignr_epi64(U, A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_alignq256_mask((__v4di)(__m256i)(A), \
+                                         (__v4di)(__m256i)(B), (int)(imm), \
+                                         (__v4di)_mm256_setzero_si256(), \
+                                         (__mmask8)(U)); })
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_movehdup_ps (__m128 __W, __mmask8 __U, __m128 __A)
+{
+  return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
+                                             (__v4sf)_mm_movehdup_ps(__A),
+                                             (__v4sf)__W);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_movehdup_ps (__mmask8 __U, __m128 __A)
+{
+  return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
+                                             (__v4sf)_mm_movehdup_ps(__A),
+                                             (__v4sf)_mm_setzero_ps());
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_movehdup_ps (__m256 __W, __mmask8 __U, __m256 __A)
+{
+  return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
+                                             (__v8sf)_mm256_movehdup_ps(__A),
+                                             (__v8sf)__W);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_movehdup_ps (__mmask8 __U, __m256 __A)
+{
+  return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
+                                             (__v8sf)_mm256_movehdup_ps(__A),
+                                             (__v8sf)_mm256_setzero_ps());
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_moveldup_ps (__m128 __W, __mmask8 __U, __m128 __A)
+{
+  return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
+                                             (__v4sf)_mm_moveldup_ps(__A),
+                                             (__v4sf)__W);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_moveldup_ps (__mmask8 __U, __m128 __A)
+{
+  return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
+                                             (__v4sf)_mm_moveldup_ps(__A),
+                                             (__v4sf)_mm_setzero_ps());
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_moveldup_ps (__m256 __W, __mmask8 __U, __m256 __A)
+{
+  return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
+                                             (__v8sf)_mm256_moveldup_ps(__A),
+                                             (__v8sf)__W);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_moveldup_ps (__mmask8 __U, __m256 __A)
+{
+  return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
+                                             (__v8sf)_mm256_moveldup_ps(__A),
+                                             (__v8sf)_mm256_setzero_ps());
+}
+
+#define _mm256_mask_shuffle_epi32(W, U, A, I) __extension__({\
+  (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \
+                                      (__v8si)_mm256_shuffle_epi32((A), (I)), \
+                                      (__v8si)(__m256i)(W)); })
+
+#define _mm256_maskz_shuffle_epi32(U, A, I) __extension__({\
+  (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \
+                                      (__v8si)_mm256_shuffle_epi32((A), (I)), \
+                                      (__v8si)_mm256_setzero_si256()); })
+
+#define _mm_mask_shuffle_epi32(W, U, A, I) __extension__({\
+  (__m128i)__builtin_ia32_selectd_128((__mmask8)(U), \
+                                      (__v4si)_mm_shuffle_epi32((A), (I)), \
+                                      (__v4si)(__m128i)(W)); })
+
+#define _mm_maskz_shuffle_epi32(U, A, I) __extension__({\
+  (__m128i)__builtin_ia32_selectd_128((__mmask8)(U), \
+                                      (__v4si)_mm_shuffle_epi32((A), (I)), \
+                                      (__v4si)_mm_setzero_si128()); })
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_mov_pd (__m128d __W, __mmask8 __U, __m128d __A)
+{
+  return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U,
+              (__v2df) __A,
+              (__v2df) __W);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_mov_pd (__mmask8 __U, __m128d __A)
+{
+  return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U,
+              (__v2df) __A,
+              (__v2df) _mm_setzero_pd ());
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_mov_pd (__m256d __W, __mmask8 __U, __m256d __A)
+{
+  return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U,
+              (__v4df) __A,
+              (__v4df) __W);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_mov_pd (__mmask8 __U, __m256d __A)
+{
+  return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U,
+              (__v4df) __A,
+              (__v4df) _mm256_setzero_pd ());
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_mov_ps (__m128 __W, __mmask8 __U, __m128 __A)
+{
+  return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U,
+             (__v4sf) __A,
+             (__v4sf) __W);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_mov_ps (__mmask8 __U, __m128 __A)
+{
+  return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U,
+             (__v4sf) __A,
+             (__v4sf) _mm_setzero_ps ());
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_mov_ps (__m256 __W, __mmask8 __U, __m256 __A)
+{
+  return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U,
+             (__v8sf) __A,
+             (__v8sf) __W);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_mov_ps (__mmask8 __U, __m256 __A)
+{
+  return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U,
+             (__v8sf) __A,
+             (__v8sf) _mm256_setzero_ps ());
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_cvtph_ps (__m128 __W, __mmask8 __U, __m128i __A)
+{
+  return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A,
+             (__v4sf) __W,
+             (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_cvtph_ps (__mmask8 __U, __m128i __A)
+{
+  return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A,
+             (__v4sf)
+             _mm_setzero_ps (),
+             (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_cvtph_ps (__m256 __W, __mmask8 __U, __m128i __A)
+{
+  return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A,
+                (__v8sf) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtph_ps (__mmask8 __U, __m128i __A)
+{
+  return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A,
+                (__v8sf)
+                _mm256_setzero_ps (),
+                (__mmask8) __U);
+}
+
+static __inline __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtps_ph (__m128i __W, __mmask8 __U, __m128 __A)
+{
+  return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, _MM_FROUND_CUR_DIRECTION,
+                                                  (__v8hi) __W,
+                                                  (__mmask8) __U);
+}
+
+static __inline __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtps_ph (__mmask8 __U, __m128 __A)
+{
+  return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, _MM_FROUND_CUR_DIRECTION,
+                                                  (__v8hi) _mm_setzero_si128 (),
+                                                  (__mmask8) __U);
+}
+
+#define _mm_mask_cvt_roundps_ph(W, U, A, I) __extension__ ({ \
+  (__m128i)__builtin_ia32_vcvtps2ph_mask((__v4sf)(__m128)(A), (int)(I), \
+                                         (__v8hi)(__m128i)(W), \
+                                         (__mmask8)(U)); })
+
+#define _mm_maskz_cvt_roundps_ph(U, A, I) __extension__ ({ \
+  (__m128i)__builtin_ia32_vcvtps2ph_mask((__v4sf)(__m128)(A), (int)(I), \
+                                         (__v8hi)_mm_setzero_si128(), \
+                                         (__mmask8)(U)); })
+
+static __inline __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtps_ph (__m128i __W, __mmask8 __U, __m256 __A)
+{
+  return (__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf) __A, _MM_FROUND_CUR_DIRECTION,
+                                                      (__v8hi) __W,
+                                                      (__mmask8) __U);
+}
+
+static __inline __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtps_ph ( __mmask8 __U, __m256 __A)
+{
+  return (__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf) __A, _MM_FROUND_CUR_DIRECTION,
+                                                      (__v8hi) _mm_setzero_si128(),
+                                                      (__mmask8) __U);
+}
+#define _mm256_mask_cvt_roundps_ph(W, U, A, I) __extension__ ({ \
+  (__m128i)__builtin_ia32_vcvtps2ph256_mask((__v8sf)(__m256)(A), (int)(I), \
+                                            (__v8hi)(__m128i)(W), \
+                                            (__mmask8)(U)); })
+
+#define _mm256_maskz_cvt_roundps_ph(U, A, I) __extension__ ({ \
+  (__m128i)__builtin_ia32_vcvtps2ph256_mask((__v8sf)(__m256)(A), (int)(I), \
+                                            (__v8hi)_mm_setzero_si128(), \
+                                            (__mmask8)(U)); })
+
+
 #undef __DEFAULT_FN_ATTRS
-#undef __DEFAULT_FN_ATTRS_BOTH
 
 #endif /* __AVX512VLINTRIN_H */