[X86] Merge the 3 different flavors of masked vpermi2var/vpermt2var builtins to a single version without masking. Use select builtins with appropriate operand instead.

llvm-svn: 333387

6 files changed, 350 insertions, 532 deletions

diff --git a/clang/lib/Headers/avx512bwintrin.h b/clang/lib/Headers/avx512bwintrin.h
index ed04a55fd64..0bf98e947f9 100644
--- a/clang/lib/Headers/avx512bwintrin.h
+++ b/clang/lib/Headers/avx512bwintrin.h
@@ -917,43 +917,37 @@ _mm512_maskz_subs_epu16 (__mmask32 __U, __m512i __A, __m512i __B)
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask2_permutex2var_epi16 (__m512i __A, __m512i __I,
-         __mmask32 __U, __m512i __B)
+_mm512_permutex2var_epi16(__m512i __A, __m512i __I, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermi2varhi512_mask ((__v32hi) __A,
-              (__v32hi) __I /* idx */ ,
-              (__v32hi) __B,
-              (__mmask32) __U);
+  return (__m512i)__builtin_ia32_vpermi2varhi512((__v32hi)__A, (__v32hi)__I,
+                                                 (__v32hi)__B);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_permutex2var_epi16 (__m512i __A, __m512i __I, __m512i __B)
+_mm512_mask_permutex2var_epi16(__m512i __A, __mmask32 __U, __m512i __I,
+                               __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermt2varhi512_mask ((__v32hi) __I /* idx */,
-              (__v32hi) __A,
-              (__v32hi) __B,
-              (__mmask32) -1);
+  return (__m512i)__builtin_ia32_selectw_512(__U,
+                              (__v32hi)_mm512_permutex2var_epi16(__A, __I, __B),
+                              (__v32hi)__A);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask_permutex2var_epi16 (__m512i __A, __mmask32 __U,
-        __m512i __I, __m512i __B)
+_mm512_mask2_permutex2var_epi16(__m512i __A, __m512i __I, __mmask32 __U,
+                                __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermt2varhi512_mask ((__v32hi) __I /* idx */,
-              (__v32hi) __A,
-              (__v32hi) __B,
-              (__mmask32) __U);
+  return (__m512i)__builtin_ia32_selectw_512(__U,
+                              (__v32hi)_mm512_permutex2var_epi16(__A, __I, __B),
+                              (__v32hi)__I);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_maskz_permutex2var_epi16 (__mmask32 __U, __m512i __A,
-         __m512i __I, __m512i __B)
+_mm512_maskz_permutex2var_epi16(__mmask32 __U, __m512i __A, __m512i __I,
+                                __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermt2varhi512_maskz ((__v32hi) __I
-              /* idx */ ,
-              (__v32hi) __A,
-              (__v32hi) __B,
-              (__mmask32) __U);
+  return (__m512i)__builtin_ia32_selectw_512(__U,
+                              (__v32hi)_mm512_permutex2var_epi16(__A, __I, __B),
+                              (__v32hi)_mm512_setzero_hi());
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
diff --git a/clang/lib/Headers/avx512fintrin.h b/clang/lib/Headers/avx512fintrin.h
index 0a3cf301b68..fc030d87a07 100644
--- a/clang/lib/Headers/avx512fintrin.h
+++ b/clang/lib/Headers/avx512fintrin.h
@@ -3399,66 +3399,69 @@ _mm512_mask3_fnmsub_ps(__m512 __A, __m512 __B, __m512 __C, __mmask16 __U)
 static __inline __m512i __DEFAULT_FN_ATTRS
 _mm512_permutex2var_epi32(__m512i __A, __m512i __I, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermt2vard512_mask ((__v16si) __I
-                                                       /* idx */ ,
-                                                       (__v16si) __A,
-                                                       (__v16si) __B,
-                                                       (__mmask16) -1);
+  return (__m512i)__builtin_ia32_vpermi2vard512((__v16si)__A, (__v16si) __I,
+                                                (__v16si) __B);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask_permutex2var_epi32 (__m512i __A, __mmask16 __U,
-                                __m512i __I, __m512i __B)
+_mm512_mask_permutex2var_epi32(__m512i __A, __mmask16 __U, __m512i __I,
+                               __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermt2vard512_mask ((__v16si) __I
-                                                        /* idx */ ,
-                                                        (__v16si) __A,
-                                                        (__v16si) __B,
-                                                        (__mmask16) __U);
+  return (__m512i)__builtin_ia32_selectd_512(__U,
+                              (__v16si)_mm512_permutex2var_epi32(__A, __I, __B),
+                              (__v16si)__A);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_maskz_permutex2var_epi32 (__mmask16 __U, __m512i __A,
-                                 __m512i __I, __m512i __B)
+_mm512_mask2_permutex2var_epi32(__m512i __A, __m512i __I, __mmask16 __U,
+                                __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermt2vard512_maskz ((__v16si) __I
-                                                        /* idx */ ,
-                                                        (__v16si) __A,
-                                                        (__v16si) __B,
-                                                        (__mmask16) __U);
+  return (__m512i)__builtin_ia32_selectd_512(__U,
+                              (__v16si)_mm512_permutex2var_epi32(__A, __I, __B),
+                              (__v16si)__I);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_permutex2var_epi32(__mmask16 __U, __m512i __A, __m512i __I,
+                                __m512i __B)
+{
+  return (__m512i)__builtin_ia32_selectd_512(__U,
+                              (__v16si)_mm512_permutex2var_epi32(__A, __I, __B),
+                              (__v16si)_mm512_setzero_si512());
 }
 
 static __inline __m512i __DEFAULT_FN_ATTRS
 _mm512_permutex2var_epi64(__m512i __A, __m512i __I, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermt2varq512_mask ((__v8di) __I
-                                                       /* idx */ ,
-                                                       (__v8di) __A,
-                                                       (__v8di) __B,
-                                                       (__mmask8) -1);
+  return (__m512i)__builtin_ia32_vpermi2varq512((__v8di)__A, (__v8di) __I,
+                                                (__v8di) __B);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask_permutex2var_epi64 (__m512i __A, __mmask8 __U, __m512i __I,
-                                __m512i __B)
+_mm512_mask_permutex2var_epi64(__m512i __A, __mmask8 __U, __m512i __I,
+                               __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermt2varq512_mask ((__v8di) __I
-                                                       /* idx */ ,
-                                                       (__v8di) __A,
-                                                       (__v8di) __B,
-                                                       (__mmask8) __U);
+  return (__m512i)__builtin_ia32_selectq_512(__U,
+                               (__v8di)_mm512_permutex2var_epi64(__A, __I, __B),
+                               (__v8di)__A);
 }
 
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask2_permutex2var_epi64(__m512i __A, __m512i __I, __mmask8 __U,
+                                __m512i __B)
+{
+  return (__m512i)__builtin_ia32_selectq_512(__U,
+                               (__v8di)_mm512_permutex2var_epi64(__A, __I, __B),
+                               (__v8di)__I);
+}
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_maskz_permutex2var_epi64 (__mmask8 __U, __m512i __A,
-         __m512i __I, __m512i __B)
+_mm512_maskz_permutex2var_epi64(__mmask8 __U, __m512i __A, __m512i __I,
+                                __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermt2varq512_maskz ((__v8di) __I
-                                                        /* idx */ ,
-                                                        (__v8di) __A,
-                                                        (__v8di) __B,
-                                                        (__mmask8) __U);
+  return (__m512i)__builtin_ia32_selectq_512(__U,
+                               (__v8di)_mm512_permutex2var_epi64(__A, __I, __B),
+                               (__v8di)_mm512_setzero_si512());
 }
 
 #define _mm512_alignr_epi64(A, B, I) __extension__ ({ \
@@ -5844,17 +5847,6 @@ _mm512_kmov (__mmask16 __A)
 #endif
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask2_permutex2var_epi32 (__m512i __A, __m512i __I,
-         __mmask16 __U, __m512i __B)
-{
-  return (__m512i) __builtin_ia32_vpermi2vard512_mask ((__v16si) __A,
-                   (__v16si) __I
-                   /* idx */ ,
-                   (__v16si) __B,
-                   (__mmask16) __U);
-}
-
-static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_sll_epi32(__m512i __A, __m128i __B)
 {
   return (__m512i)__builtin_ia32_pslld512((__v16si) __A, (__v4si)__B);
@@ -6331,39 +6323,6 @@ _mm_cvttss_u64 (__m128 __A)
 }
 #endif
 
-static __inline__ __m512d __DEFAULT_FN_ATTRS
-_mm512_mask2_permutex2var_pd (__m512d __A, __m512i __I, __mmask8 __U,
-            __m512d __B)
-{
-  return (__m512d) __builtin_ia32_vpermi2varpd512_mask ((__v8df) __A,
-              (__v8di) __I
-              /* idx */ ,
-              (__v8df) __B,
-              (__mmask8) __U);
-}
-
-static __inline__ __m512 __DEFAULT_FN_ATTRS
-_mm512_mask2_permutex2var_ps (__m512 __A, __m512i __I, __mmask16 __U,
-            __m512 __B)
-{
-  return (__m512) __builtin_ia32_vpermi2varps512_mask ((__v16sf) __A,
-                   (__v16si) __I
-                   /* idx */ ,
-                   (__v16sf) __B,
-                   (__mmask16) __U);
-}
-
-static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask2_permutex2var_epi64 (__m512i __A, __m512i __I,
-         __mmask8 __U, __m512i __B)
-{
-  return (__m512i) __builtin_ia32_vpermi2varq512_mask ((__v8di) __A,
-                   (__v8di) __I
-                   /* idx */ ,
-                   (__v8di) __B,
-                   (__mmask8) __U);
-}
-
 #define _mm512_permute_pd(X, C) __extension__ ({ \
   (__m512d)__builtin_shufflevector((__v8df)(__m512d)(X), \
                                    (__v8df)_mm512_undefined_pd(), \
@@ -6463,63 +6422,65 @@ _mm512_maskz_permutevar_ps(__mmask16 __U, __m512 __A, __m512i __C)
 static __inline __m512d __DEFAULT_FN_ATTRS
 _mm512_permutex2var_pd(__m512d __A, __m512i __I, __m512d __B)
 {
-  return (__m512d) __builtin_ia32_vpermt2varpd512_mask ((__v8di) __I
-                    /* idx */ ,
-                    (__v8df) __A,
-                    (__v8df) __B,
-                    (__mmask8) -1);
+  return (__m512d)__builtin_ia32_vpermi2varpd512((__v8df)__A, (__v8di)__I,
+                                                 (__v8df)__B);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_mask_permutex2var_pd(__m512d __A, __mmask8 __U, __m512i __I, __m512d __B)
+{
+  return (__m512d)__builtin_ia32_selectpd_512(__U,
+                                  (__v8df)_mm512_permutex2var_pd(__A, __I, __B),
+                                  (__v8df)__A);
 }
 
 static __inline__ __m512d __DEFAULT_FN_ATTRS
-_mm512_mask_permutex2var_pd (__m512d __A, __mmask8 __U, __m512i __I, __m512d __B)
+_mm512_mask2_permutex2var_pd(__m512d __A, __m512i __I, __mmask8 __U,
+                             __m512d __B)
 {
-  return (__m512d) __builtin_ia32_vpermt2varpd512_mask ((__v8di) __I
-                    /* idx */ ,
-                    (__v8df) __A,
-                    (__v8df) __B,
-                    (__mmask8) __U);
+  return (__m512d)__builtin_ia32_selectpd_512(__U,
+                                  (__v8df)_mm512_permutex2var_pd(__A, __I, __B),
+                                  (__v8df)(__m512d)__I);
 }
 
 static __inline__ __m512d __DEFAULT_FN_ATTRS
-_mm512_maskz_permutex2var_pd (__mmask8 __U, __m512d __A, __m512i __I,
-            __m512d __B)
+_mm512_maskz_permutex2var_pd(__mmask8 __U, __m512d __A, __m512i __I,
+                             __m512d __B)
 {
-  return (__m512d) __builtin_ia32_vpermt2varpd512_maskz ((__v8di) __I
-                                                         /* idx */ ,
-                                                         (__v8df) __A,
-                                                         (__v8df) __B,
-                                                         (__mmask8) __U);
+  return (__m512d)__builtin_ia32_selectpd_512(__U,
+                                  (__v8df)_mm512_permutex2var_pd(__A, __I, __B),
+                                  (__v8df)_mm512_setzero_pd());
 }
 
 static __inline __m512 __DEFAULT_FN_ATTRS
 _mm512_permutex2var_ps(__m512 __A, __m512i __I, __m512 __B)
 {
-  return (__m512) __builtin_ia32_vpermt2varps512_mask ((__v16si) __I
-                                                         /* idx */ ,
-                                                         (__v16sf) __A,
-                                                         (__v16sf) __B,
-                                                         (__mmask16) -1);
+  return (__m512)__builtin_ia32_vpermi2varps512((__v16sf)__A, (__v16si)__I,
+                                                (__v16sf) __B);
 }
 
 static __inline__ __m512 __DEFAULT_FN_ATTRS
-_mm512_mask_permutex2var_ps (__m512 __A, __mmask16 __U, __m512i __I, __m512 __B)
+_mm512_mask_permutex2var_ps(__m512 __A, __mmask16 __U, __m512i __I, __m512 __B)
 {
-  return (__m512) __builtin_ia32_vpermt2varps512_mask ((__v16si) __I
-                                                         /* idx */ ,
-                                                         (__v16sf) __A,
-                                                         (__v16sf) __B,
-                                                         (__mmask16) __U);
+  return (__m512)__builtin_ia32_selectps_512(__U,
+                                 (__v16sf)_mm512_permutex2var_ps(__A, __I, __B),
+                                 (__v16sf)__A);
 }
 
 static __inline__ __m512 __DEFAULT_FN_ATTRS
-_mm512_maskz_permutex2var_ps (__mmask16 __U, __m512 __A, __m512i __I,
-            __m512 __B)
-{
-  return (__m512) __builtin_ia32_vpermt2varps512_maskz ((__v16si) __I
-                                                        /* idx */ ,
-                                                        (__v16sf) __A,
-                                                        (__v16sf) __B,
-                                                        (__mmask16) __U);
+_mm512_mask2_permutex2var_ps(__m512 __A, __m512i __I, __mmask16 __U, __m512 __B)
+{
+  return (__m512)__builtin_ia32_selectps_512(__U,
+                                 (__v16sf)_mm512_permutex2var_ps(__A, __I, __B),
+                                 (__v16sf)(__m512)__I);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_maskz_permutex2var_ps(__mmask16 __U, __m512 __A, __m512i __I, __m512 __B)
+{
+  return (__m512)__builtin_ia32_selectps_512(__U,
+                                 (__v16sf)_mm512_permutex2var_ps(__A, __I, __B),
+                                 (__v16sf)_mm512_setzero_ps());
 }
 
 
diff --git a/clang/lib/Headers/avx512vbmiintrin.h b/clang/lib/Headers/avx512vbmiintrin.h
index a8ae95c7f29..dd6593e7b41 100644
--- a/clang/lib/Headers/avx512vbmiintrin.h
+++ b/clang/lib/Headers/avx512vbmiintrin.h
@@ -33,46 +33,37 @@
 
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask2_permutex2var_epi8 (__m512i __A, __m512i __I,
-         __mmask64 __U, __m512i __B)
+_mm512_permutex2var_epi8(__m512i __A, __m512i __I, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermi2varqi512_mask ((__v64qi) __A,
-              (__v64qi) __I
-              /* idx */ ,
-              (__v64qi) __B,
-              (__mmask64) __U);
+  return (__m512i)__builtin_ia32_vpermi2varqi512((__v64qi)__A, (__v64qi)__I,
+                                                 (__v64qi) __B);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_permutex2var_epi8 (__m512i __A, __m512i __I, __m512i __B)
+_mm512_mask_permutex2var_epi8(__m512i __A, __mmask64 __U, __m512i __I,
+                              __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermt2varqi512_mask ((__v64qi) __I
-              /* idx */ ,
-              (__v64qi) __A,
-              (__v64qi) __B,
-              (__mmask64) -1);
+  return (__m512i)__builtin_ia32_selectb_512(__U,
+                               (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B),
+                               (__v64qi)__A);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask_permutex2var_epi8 (__m512i __A, __mmask64 __U,
-        __m512i __I, __m512i __B)
+_mm512_mask2_permutex2var_epi8(__m512i __A, __m512i __I, __mmask64 __U,
+                               __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermt2varqi512_mask ((__v64qi) __I
-              /* idx */ ,
-              (__v64qi) __A,
-              (__v64qi) __B,
-              (__mmask64) __U);
+  return (__m512i)__builtin_ia32_selectb_512(__U,
+                               (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B),
+                               (__v64qi)__I);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_maskz_permutex2var_epi8 (__mmask64 __U, __m512i __A,
-         __m512i __I, __m512i __B)
+_mm512_maskz_permutex2var_epi8(__mmask64 __U, __m512i __A, __m512i __I,
+                               __m512i __B)
 {
-  return (__m512i) __builtin_ia32_vpermt2varqi512_maskz ((__v64qi) __I
-               /* idx */ ,
-               (__v64qi) __A,
-               (__v64qi) __B,
-               (__mmask64) __U);
+  return (__m512i)__builtin_ia32_selectb_512(__U,
+                               (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B),
+                               (__v64qi)_mm512_setzero_si512());
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
diff --git a/clang/lib/Headers/avx512vbmivlintrin.h b/clang/lib/Headers/avx512vbmivlintrin.h
index b2336554eff..34b8d763659 100644
--- a/clang/lib/Headers/avx512vbmivlintrin.h
+++ b/clang/lib/Headers/avx512vbmivlintrin.h
@@ -33,97 +33,72 @@
 
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask2_permutex2var_epi8 (__m128i __A, __m128i __I, __mmask16 __U,
-            __m128i __B)
+_mm_permutex2var_epi8(__m128i __A, __m128i __I, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_vpermi2varqi128_mask ((__v16qi) __A,
-              (__v16qi) __I
-              /* idx */ ,
-              (__v16qi) __B,
-              (__mmask16)
-              __U);
+  return (__m128i)__builtin_ia32_vpermi2varqi128((__v16qi)__A,
+                                                 (__v16qi)__I,
+                                                 (__v16qi)__B);
 }
 
-static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask2_permutex2var_epi8 (__m256i __A, __m256i __I,
-         __mmask32 __U, __m256i __B)
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_permutex2var_epi8(__m128i __A, __mmask16 __U, __m128i __I,
+                           __m128i __B)
 {
-  return (__m256i) __builtin_ia32_vpermi2varqi256_mask ((__v32qi) __A,
-              (__v32qi) __I
-              /* idx */ ,
-              (__v32qi) __B,
-              (__mmask32)
-              __U);
+  return (__m128i)__builtin_ia32_selectb_128(__U,
+                                  (__v16qi)_mm_permutex2var_epi8(__A, __I, __B),
+                                  (__v16qi)__A);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_permutex2var_epi8 (__m128i __A, __m128i __I, __m128i __B)
+_mm_mask2_permutex2var_epi8(__m128i __A, __m128i __I, __mmask16 __U,
+                            __m128i __B)
 {
-  return (__m128i) __builtin_ia32_vpermt2varqi128_mask ((__v16qi) __I
-              /* idx */ ,
-              (__v16qi) __A,
-              (__v16qi) __B,
-              (__mmask16) -
-              1);
+  return (__m128i)__builtin_ia32_selectb_128(__U,
+                                  (__v16qi)_mm_permutex2var_epi8(__A, __I, __B),
+                                  (__v16qi)__I);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_permutex2var_epi8 (__m128i __A, __mmask16 __U, __m128i __I,
-           __m128i __B)
+_mm_maskz_permutex2var_epi8(__mmask16 __U, __m128i __A, __m128i __I,
+                            __m128i __B)
 {
-  return (__m128i) __builtin_ia32_vpermt2varqi128_mask ((__v16qi) __I
-              /* idx */ ,
-              (__v16qi) __A,
-              (__v16qi) __B,
-              (__mmask16)
-              __U);
+  return (__m128i)__builtin_ia32_selectb_128(__U,
+                                  (__v16qi)_mm_permutex2var_epi8(__A, __I, __B),
+                                  (__v16qi)_mm_setzero_si128());
 }
 
-static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_maskz_permutex2var_epi8 (__mmask16 __U, __m128i __A, __m128i __I,
-            __m128i __B)
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_permutex2var_epi8(__m256i __A, __m256i __I, __m256i __B)
 {
-  return (__m128i) __builtin_ia32_vpermt2varqi128_maskz ((__v16qi) __I
-               /* idx */ ,
-               (__v16qi) __A,
-               (__v16qi) __B,
-               (__mmask16)
-               __U);
+  return (__m256i)__builtin_ia32_vpermi2varqi256((__v32qi)__A, (__v32qi)__I,
+                                                 (__v32qi)__B);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_permutex2var_epi8 (__m256i __A, __m256i __I, __m256i __B)
+_mm256_mask_permutex2var_epi8(__m256i __A, __mmask32 __U, __m256i __I,
+                              __m256i __B)
 {
-  return (__m256i) __builtin_ia32_vpermt2varqi256_mask ((__v32qi) __I
-              /* idx */ ,
-              (__v32qi) __A,
-              (__v32qi) __B,
-              (__mmask32) -
-              1);
+  return (__m256i)__builtin_ia32_selectb_256(__U,
+                               (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B),
+                               (__v32qi)__A);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_permutex2var_epi8 (__m256i __A, __mmask32 __U,
-        __m256i __I, __m256i __B)
+_mm256_mask2_permutex2var_epi8(__m256i __A, __m256i __I, __mmask32 __U,
+                               __m256i __B)
 {
-  return (__m256i) __builtin_ia32_vpermt2varqi256_mask ((__v32qi) __I
-              /* idx */ ,
-              (__v32qi) __A,
-              (__v32qi) __B,
-              (__mmask32)
-              __U);
+  return (__m256i)__builtin_ia32_selectb_256(__U,
+                               (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B),
+                               (__v32qi)__I);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_permutex2var_epi8 (__mmask32 __U, __m256i __A,
-         __m256i __I, __m256i __B)
-{
-  return (__m256i) __builtin_ia32_vpermt2varqi256_maskz ((__v32qi) __I
-               /* idx */ ,
-               (__v32qi) __A,
-               (__v32qi) __B,
-               (__mmask32)
-               __U);
+_mm256_maskz_permutex2var_epi8(__mmask32 __U, __m256i __A, __m256i __I,
+                               __m256i __B)
+{
+  return (__m256i)__builtin_ia32_selectb_256(__U,
+                               (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B),
+                               (__v32qi)_mm256_setzero_si256());
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
diff --git a/clang/lib/Headers/avx512vlbwintrin.h b/clang/lib/Headers/avx512vlbwintrin.h
index 6c2d3eadc9c..30a12b6072c 100644
--- a/clang/lib/Headers/avx512vlbwintrin.h
+++ b/clang/lib/Headers/avx512vlbwintrin.h
@@ -1290,81 +1290,71 @@ _mm256_maskz_subs_epu16(__mmask16 __U, __m256i __A, __m256i __B)
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask2_permutex2var_epi16 (__m128i __A, __m128i __I, __mmask8 __U,
-            __m128i __B)
-{
-  return (__m128i) __builtin_ia32_vpermi2varhi128_mask ((__v8hi) __A,
-               (__v8hi) __I /* idx */ ,
-               (__v8hi) __B,
-               (__mmask8) __U);
-}
-
-static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask2_permutex2var_epi16 (__m256i __A, __m256i __I,
-         __mmask16 __U, __m256i __B)
+_mm_permutex2var_epi16(__m128i __A, __m128i __I, __m128i __B)
 {
-  return (__m256i) __builtin_ia32_vpermi2varhi256_mask ((__v16hi) __A,
-               (__v16hi) __I /* idx */ ,
-               (__v16hi) __B,
-               (__mmask16) __U);
+  return (__m128i)__builtin_ia32_vpermi2varhi128((__v8hi)__A, (__v8hi)__I,
+                                                 (__v8hi) __B);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_permutex2var_epi16 (__m128i __A, __m128i __I, __m128i __B)
+_mm_mask_permutex2var_epi16(__m128i __A, __mmask8 __U, __m128i __I,
+                            __m128i __B)
 {
-  return (__m128i) __builtin_ia32_vpermt2varhi128_mask ((__v8hi) __I/* idx */,
-               (__v8hi) __A,
-               (__v8hi) __B,
-               (__mmask8) -1);
+  return (__m128i)__builtin_ia32_selectw_128(__U,
+                                  (__v8hi)_mm_permutex2var_epi16(__A, __I, __B),
+                                  (__v8hi)__A);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_permutex2var_epi16 (__m128i __A, __mmask8 __U, __m128i __I,
-           __m128i __B)
+_mm_mask2_permutex2var_epi16(__m128i __A, __m128i __I, __mmask8 __U,
+                             __m128i __B)
 {
-  return (__m128i) __builtin_ia32_vpermt2varhi128_mask ((__v8hi) __I/* idx */,
-               (__v8hi) __A,
-               (__v8hi) __B,
-               (__mmask8) __U);
+  return (__m128i)__builtin_ia32_selectw_128(__U,
+                                  (__v8hi)_mm_permutex2var_epi16(__A, __I, __B),
+                                  (__v8hi)__I);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_maskz_permutex2var_epi16 (__mmask8 __U, __m128i __A, __m128i __I,
             __m128i __B)
 {
-  return (__m128i) __builtin_ia32_vpermt2varhi128_maskz ((__v8hi) __I/* idx */,
-               (__v8hi) __A,
-               (__v8hi) __B,
-               (__mmask8) __U);
+  return (__m128i)__builtin_ia32_selectw_128(__U,
+                                  (__v8hi)_mm_permutex2var_epi16(__A, __I, __B),
+                                  (__v8hi)_mm_setzero_si128());
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_permutex2var_epi16 (__m256i __A, __m256i __I, __m256i __B)
+_mm256_permutex2var_epi16(__m256i __A, __m256i __I, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_vpermt2varhi256_mask ((__v16hi) __I/* idx */,
-               (__v16hi) __A,
-               (__v16hi) __B,
-               (__mmask16) -1);
+  return (__m256i)__builtin_ia32_vpermi2varhi256((__v16hi)__A, (__v16hi)__I,
+                                                 (__v16hi)__B);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_permutex2var_epi16(__m256i __A, __mmask16 __U, __m256i __I,
+                               __m256i __B)
+{
+  return (__m256i)__builtin_ia32_selectw_256(__U,
+                              (__v16hi)_mm256_permutex2var_epi16(__A, __I, __B),
+                              (__v16hi)__A);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_permutex2var_epi16 (__m256i __A, __mmask16 __U,
-        __m256i __I, __m256i __B)
+_mm256_mask2_permutex2var_epi16(__m256i __A, __m256i __I, __mmask16 __U,
+                                __m256i __B)
 {
-  return (__m256i) __builtin_ia32_vpermt2varhi256_mask ((__v16hi) __I/* idx */,
-               (__v16hi) __A,
-               (__v16hi) __B,
-               (__mmask16) __U);
+  return (__m256i)__builtin_ia32_selectw_256(__U,
+                              (__v16hi)_mm256_permutex2var_epi16(__A, __I, __B),
+                              (__v16hi)__I);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_permutex2var_epi16 (__mmask16 __U, __m256i __A,
-         __m256i __I, __m256i __B)
+_mm256_maskz_permutex2var_epi16 (__mmask16 __U, __m256i __A, __m256i __I,
+                                 __m256i __B)
 {
-  return (__m256i) __builtin_ia32_vpermt2varhi256_maskz ((__v16hi) __I/* idx */,
-               (__v16hi) __A,
-               (__v16hi) __B,
-               (__mmask16) __U);
+  return (__m256i)__builtin_ia32_selectw_256(__U,
+                              (__v16hi)_mm256_permutex2var_epi16(__A, __I, __B),
+                              (__v16hi)_mm256_setzero_si256());
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
diff --git a/clang/lib/Headers/avx512vlintrin.h b/clang/lib/Headers/avx512vlintrin.h
index 3fc2b80294c..833e70a9471 100644
--- a/clang/lib/Headers/avx512vlintrin.h
+++ b/clang/lib/Headers/avx512vlintrin.h
@@ -3598,330 +3598,237 @@ _mm256_maskz_sub_ps(__mmask8 __U, __m256 __A, __m256 __B) {
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask2_permutex2var_epi32 (__m128i __A, __m128i __I, __mmask8 __U,
-            __m128i __B) {
-  return (__m128i) __builtin_ia32_vpermi2vard128_mask ((__v4si) __A,
-                   (__v4si) __I
-                   /* idx */ ,
-                   (__v4si) __B,
-                   (__mmask8) __U);
-}
-
-static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask2_permutex2var_epi32 (__m256i __A, __m256i __I,
-         __mmask8 __U, __m256i __B) {
-  return (__m256i) __builtin_ia32_vpermi2vard256_mask ((__v8si) __A,
-                   (__v8si) __I
-                   /* idx */ ,
-                   (__v8si) __B,
-                   (__mmask8) __U);
-}
-
-static __inline__ __m128d __DEFAULT_FN_ATTRS
-_mm_mask2_permutex2var_pd (__m128d __A, __m128i __I, __mmask8 __U,
-         __m128d __B) {
-  return (__m128d) __builtin_ia32_vpermi2varpd128_mask ((__v2df) __A,
-              (__v2di) __I
-              /* idx */ ,
-              (__v2df) __B,
-              (__mmask8)
-              __U);
-}
-
-static __inline__ __m256d __DEFAULT_FN_ATTRS
-_mm256_mask2_permutex2var_pd (__m256d __A, __m256i __I, __mmask8 __U,
-            __m256d __B) {
-  return (__m256d) __builtin_ia32_vpermi2varpd256_mask ((__v4df) __A,
-              (__v4di) __I
-              /* idx */ ,
-              (__v4df) __B,
-              (__mmask8)
-              __U);
-}
-
-static __inline__ __m128 __DEFAULT_FN_ATTRS
-_mm_mask2_permutex2var_ps (__m128 __A, __m128i __I, __mmask8 __U,
-         __m128 __B) {
-  return (__m128) __builtin_ia32_vpermi2varps128_mask ((__v4sf) __A,
-                   (__v4si) __I
-                   /* idx */ ,
-                   (__v4sf) __B,
-                   (__mmask8) __U);
-}
-
-static __inline__ __m256 __DEFAULT_FN_ATTRS
-_mm256_mask2_permutex2var_ps (__m256 __A, __m256i __I, __mmask8 __U,
-            __m256 __B) {
-  return (__m256) __builtin_ia32_vpermi2varps256_mask ((__v8sf) __A,
-                   (__v8si) __I
-                   /* idx */ ,
-                   (__v8sf) __B,
-                   (__mmask8) __U);
+_mm_permutex2var_epi32(__m128i __A, __m128i __I, __m128i __B) {
+  return (__m128i)__builtin_ia32_vpermi2vard128((__v4si) __A, (__v4si)__I,
+                                                (__v4si)__B);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask2_permutex2var_epi64 (__m128i __A, __m128i __I, __mmask8 __U,
-            __m128i __B) {
-  return (__m128i) __builtin_ia32_vpermi2varq128_mask ((__v2di) __A,
-                   (__v2di) __I
-                   /* idx */ ,
-                   (__v2di) __B,
-                   (__mmask8) __U);
-}
-
-static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask2_permutex2var_epi64 (__m256i __A, __m256i __I,
-         __mmask8 __U, __m256i __B) {
-  return (__m256i) __builtin_ia32_vpermi2varq256_mask ((__v4di) __A,
-                   (__v4di) __I
-                   /* idx */ ,
-                   (__v4di) __B,
-                   (__mmask8) __U);
+_mm_mask_permutex2var_epi32(__m128i __A, __mmask8 __U, __m128i __I,
+                            __m128i __B) {
+  return (__m128i)__builtin_ia32_selectd_128(__U,
+                                  (__v4si)_mm_permutex2var_epi32(__A, __I, __B),
+                                  (__v4si)__A);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_permutex2var_epi32 (__m128i __A, __m128i __I, __m128i __B) {
-  return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I
-                   /* idx */ ,
-                   (__v4si) __A,
-                   (__v4si) __B,
-                   (__mmask8) -1);
+_mm_mask2_permutex2var_epi32(__m128i __A, __m128i __I, __mmask8 __U,
+                             __m128i __B) {
+  return (__m128i)__builtin_ia32_selectd_128(__U,
+                                  (__v4si)_mm_permutex2var_epi32(__A, __I, __B),
+                                  (__v4si)__I);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_permutex2var_epi32 (__m128i __A, __mmask8 __U, __m128i __I,
-           __m128i __B) {
-  return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I
-                   /* idx */ ,
-                   (__v4si) __A,
-                   (__v4si) __B,
-                   (__mmask8) __U);
+_mm_maskz_permutex2var_epi32(__mmask8 __U, __m128i __A, __m128i __I,
+                             __m128i __B) {
+  return (__m128i)__builtin_ia32_selectd_128(__U,
+                                  (__v4si)_mm_permutex2var_epi32(__A, __I, __B),
+                                  (__v4si)_mm_setzero_si128());
 }
 
-static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_maskz_permutex2var_epi32 (__mmask8 __U, __m128i __A, __m128i __I,
-            __m128i __B) {
-  return (__m128i) __builtin_ia32_vpermt2vard128_maskz ((__v4si) __I
-              /* idx */ ,
-              (__v4si) __A,
-              (__v4si) __B,
-              (__mmask8)
-              __U);
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_permutex2var_epi32(__m256i __A, __m256i __I, __m256i __B) {
+  return (__m256i)__builtin_ia32_vpermi2vard256((__v8si)__A, (__v8si) __I,
+                                                (__v8si) __B);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_permutex2var_epi32 (__m256i __A, __m256i __I, __m256i __B) {
-  return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I
-                   /* idx */ ,
-                   (__v8si) __A,
-                   (__v8si) __B,
-                   (__mmask8) -1);
+_mm256_mask_permutex2var_epi32(__m256i __A, __mmask8 __U, __m256i __I,
+                               __m256i __B) {
+  return (__m256i)__builtin_ia32_selectd_256(__U,
+                               (__v8si)_mm256_permutex2var_epi32(__A, __I, __B),
+                               (__v8si)__A);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_permutex2var_epi32 (__m256i __A, __mmask8 __U, __m256i __I,
-        __m256i __B) {
-  return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I
-                   /* idx */ ,
-                   (__v8si) __A,
-                   (__v8si) __B,
-                   (__mmask8) __U);
+_mm256_mask2_permutex2var_epi32(__m256i __A, __m256i __I, __mmask8 __U,
+                                __m256i __B) {
+  return (__m256i)__builtin_ia32_selectd_256(__U,
+                               (__v8si)_mm256_permutex2var_epi32(__A, __I, __B),
+                               (__v8si)__I);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_permutex2var_epi32 (__mmask8 __U, __m256i __A,
-         __m256i __I, __m256i __B) {
-  return (__m256i) __builtin_ia32_vpermt2vard256_maskz ((__v8si) __I
-              /* idx */ ,
-              (__v8si) __A,
-              (__v8si) __B,
-              (__mmask8)
-              __U);
+_mm256_maskz_permutex2var_epi32(__mmask8 __U, __m256i __A, __m256i __I,
+                                __m256i __B) {
+  return (__m256i)__builtin_ia32_selectd_256(__U,
+                               (__v8si)_mm256_permutex2var_epi32(__A, __I, __B),
+                               (__v8si)_mm256_setzero_si256());
 }
 
 static __inline__ __m128d __DEFAULT_FN_ATTRS
-_mm_permutex2var_pd (__m128d __A, __m128i __I, __m128d __B) {
-  return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I
-              /* idx */ ,
-              (__v2df) __A,
-              (__v2df) __B,
-              (__mmask8) -
-              1);
+_mm_permutex2var_pd(__m128d __A, __m128i __I, __m128d __B) {
+  return (__m128d)__builtin_ia32_vpermi2varpd128((__v2df)__A, (__v2di)__I,
+                                                 (__v2df)__B);
 }
 
 static __inline__ __m128d __DEFAULT_FN_ATTRS
-_mm_mask_permutex2var_pd (__m128d __A, __mmask8 __U, __m128i __I,
-        __m128d __B) {
-  return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I
-              /* idx */ ,
-              (__v2df) __A,
-              (__v2df) __B,
-              (__mmask8)
-              __U);
+_mm_mask_permutex2var_pd(__m128d __A, __mmask8 __U, __m128i __I, __m128d __B) {
+  return (__m128d)__builtin_ia32_selectpd_128(__U,
+                                     (__v2df)_mm_permutex2var_pd(__A, __I, __B),
+                                     (__v2df)__A);
 }
 
 static __inline__ __m128d __DEFAULT_FN_ATTRS
-_mm_maskz_permutex2var_pd (__mmask8 __U, __m128d __A, __m128i __I,
-         __m128d __B) {
-  return (__m128d) __builtin_ia32_vpermt2varpd128_maskz ((__v2di) __I
-               /* idx */ ,
-               (__v2df) __A,
-               (__v2df) __B,
-               (__mmask8)
-               __U);
+_mm_mask2_permutex2var_pd(__m128d __A, __m128i __I, __mmask8 __U, __m128d __B) {
+  return (__m128d)__builtin_ia32_selectpd_128(__U,
+                                     (__v2df)_mm_permutex2var_pd(__A, __I, __B),
+                                     (__v2df)(__m128d)__I);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_permutex2var_pd(__mmask8 __U, __m128d __A, __m128i __I, __m128d __B) {
+  return (__m128d)__builtin_ia32_selectpd_128(__U,
+                                     (__v2df)_mm_permutex2var_pd(__A, __I, __B),
+                                     (__v2df)_mm_setzero_pd());
 }
 
 static __inline__ __m256d __DEFAULT_FN_ATTRS
-_mm256_permutex2var_pd (__m256d __A, __m256i __I, __m256d __B) {
-  return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I
-              /* idx */ ,
-              (__v4df) __A,
-              (__v4df) __B,
-              (__mmask8) -
-              1);
+_mm256_permutex2var_pd(__m256d __A, __m256i __I, __m256d __B) {
+  return (__m256d)__builtin_ia32_vpermi2varpd256((__v4df)__A, (__v4di)__I,
+                                                 (__v4df)__B);
 }
 
 static __inline__ __m256d __DEFAULT_FN_ATTRS
-_mm256_mask_permutex2var_pd (__m256d __A, __mmask8 __U, __m256i __I,
-           __m256d __B) {
-  return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I
-              /* idx */ ,
-              (__v4df) __A,
-              (__v4df) __B,
-              (__mmask8)
-              __U);
+_mm256_mask_permutex2var_pd(__m256d __A, __mmask8 __U, __m256i __I,
+                            __m256d __B) {
+  return (__m256d)__builtin_ia32_selectpd_256(__U,
+                                  (__v4df)_mm256_permutex2var_pd(__A, __I, __B),
+                                  (__v4df)__A);
 }
 
 static __inline__ __m256d __DEFAULT_FN_ATTRS
-_mm256_maskz_permutex2var_pd (__mmask8 __U, __m256d __A, __m256i __I,
-            __m256d __B) {
-  return (__m256d) __builtin_ia32_vpermt2varpd256_maskz ((__v4di) __I
-               /* idx */ ,
-               (__v4df) __A,
-               (__v4df) __B,
-               (__mmask8)
-               __U);
+_mm256_mask2_permutex2var_pd(__m256d __A, __m256i __I, __mmask8 __U,
+                             __m256d __B) {
+  return (__m256d)__builtin_ia32_selectpd_256(__U,
+                                  (__v4df)_mm256_permutex2var_pd(__A, __I, __B),
+                                  (__v4df)(__m256d)__I);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_permutex2var_pd(__mmask8 __U, __m256d __A, __m256i __I,
+                             __m256d __B) {
+  return (__m256d)__builtin_ia32_selectpd_256(__U,
+                                  (__v4df)_mm256_permutex2var_pd(__A, __I, __B),
+                                  (__v4df)_mm256_setzero_pd());
 }
 
 static __inline__ __m128 __DEFAULT_FN_ATTRS
-_mm_permutex2var_ps (__m128 __A, __m128i __I, __m128 __B) {
-  return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I
-                   /* idx */ ,
-                   (__v4sf) __A,
-                   (__v4sf) __B,
-                   (__mmask8) -1);
+_mm_permutex2var_ps(__m128 __A, __m128i __I, __m128 __B) {
+  return (__m128)__builtin_ia32_vpermi2varps128((__v4sf)__A, (__v4si)__I,
+                                                (__v4sf)__B);
 }
 
 static __inline__ __m128 __DEFAULT_FN_ATTRS
-_mm_mask_permutex2var_ps (__m128 __A, __mmask8 __U, __m128i __I,
-        __m128 __B) {
-  return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I
-                   /* idx */ ,
-                   (__v4sf) __A,
-                   (__v4sf) __B,
-                   (__mmask8) __U);
+_mm_mask_permutex2var_ps(__m128 __A, __mmask8 __U, __m128i __I, __m128 __B) {
+  return (__m128)__builtin_ia32_selectps_128(__U,
+                                     (__v4sf)_mm_permutex2var_ps(__A, __I, __B),
+                                     (__v4sf)__A);
 }
 
 static __inline__ __m128 __DEFAULT_FN_ATTRS
-_mm_maskz_permutex2var_ps (__mmask8 __U, __m128 __A, __m128i __I,
-         __m128 __B) {
-  return (__m128) __builtin_ia32_vpermt2varps128_maskz ((__v4si) __I
-              /* idx */ ,
-              (__v4sf) __A,
-              (__v4sf) __B,
-              (__mmask8)
-              __U);
+_mm_mask2_permutex2var_ps(__m128 __A, __m128i __I, __mmask8 __U, __m128 __B) {
+  return (__m128)__builtin_ia32_selectps_128(__U,
+                                     (__v4sf)_mm_permutex2var_ps(__A, __I, __B),
+                                     (__v4sf)(__m128)__I);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_permutex2var_ps(__mmask8 __U, __m128 __A, __m128i __I, __m128 __B) {
+  return (__m128)__builtin_ia32_selectps_128(__U,
+                                     (__v4sf)_mm_permutex2var_ps(__A, __I, __B),
+                                     (__v4sf)_mm_setzero_ps());
 }
 
 static __inline__ __m256 __DEFAULT_FN_ATTRS
-_mm256_permutex2var_ps (__m256 __A, __m256i __I, __m256 __B) {
-  return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I
-                   /* idx */ ,
-                   (__v8sf) __A,
-                   (__v8sf) __B,
-                   (__mmask8) -1);
+_mm256_permutex2var_ps(__m256 __A, __m256i __I, __m256 __B) {
+  return (__m256)__builtin_ia32_vpermi2varps256((__v8sf)__A, (__v8si)__I,
+                                                (__v8sf) __B);
 }
 
 static __inline__ __m256 __DEFAULT_FN_ATTRS
-_mm256_mask_permutex2var_ps (__m256 __A, __mmask8 __U, __m256i __I,
-           __m256 __B) {
-  return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I
-                   /* idx */ ,
-                   (__v8sf) __A,
-                   (__v8sf) __B,
-                   (__mmask8) __U);
+_mm256_mask_permutex2var_ps(__m256 __A, __mmask8 __U, __m256i __I, __m256 __B) {
+  return (__m256)__builtin_ia32_selectps_256(__U,
+                                  (__v8sf)_mm256_permutex2var_ps(__A, __I, __B),
+                                  (__v8sf)__A);
 }
 
 static __inline__ __m256 __DEFAULT_FN_ATTRS
-_mm256_maskz_permutex2var_ps (__mmask8 __U, __m256 __A, __m256i __I,
-            __m256 __B) {
-  return (__m256) __builtin_ia32_vpermt2varps256_maskz ((__v8si) __I
-              /* idx */ ,
-              (__v8sf) __A,
-              (__v8sf) __B,
-              (__mmask8)
-              __U);
+_mm256_mask2_permutex2var_ps(__m256 __A, __m256i __I, __mmask8 __U,
+                             __m256 __B) {
+  return (__m256)__builtin_ia32_selectps_256(__U,
+                                  (__v8sf)_mm256_permutex2var_ps(__A, __I, __B),
+                                  (__v8sf)(__m256)__I);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_permutex2var_ps(__mmask8 __U, __m256 __A, __m256i __I,
+                             __m256 __B) {
+  return (__m256)__builtin_ia32_selectps_256(__U,
+                                  (__v8sf)_mm256_permutex2var_ps(__A, __I, __B),
+                                  (__v8sf)_mm256_setzero_ps());
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_permutex2var_epi64 (__m128i __A, __m128i __I, __m128i __B) {
-  return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I
-                   /* idx */ ,
-                   (__v2di) __A,
-                   (__v2di) __B,
-                   (__mmask8) -1);
+_mm_permutex2var_epi64(__m128i __A, __m128i __I, __m128i __B) {
+  return (__m128i)__builtin_ia32_vpermi2varq128((__v2di)__A, (__v2di)__I,
+                                                (__v2di)__B);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_mask_permutex2var_epi64 (__m128i __A, __mmask8 __U, __m128i __I,
-           __m128i __B) {
-  return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I
-                   /* idx */ ,
-                   (__v2di) __A,
-                   (__v2di) __B,
-                   (__mmask8) __U);
+_mm_mask_permutex2var_epi64(__m128i __A, __mmask8 __U, __m128i __I,
+                            __m128i __B) {
+  return (__m128i)__builtin_ia32_selectq_128(__U,
+                                  (__v2di)_mm_permutex2var_epi64(__A, __I, __B),
+                                  (__v2di)__A);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
-_mm_maskz_permutex2var_epi64 (__mmask8 __U, __m128i __A, __m128i __I,
-            __m128i __B) {
-  return (__m128i) __builtin_ia32_vpermt2varq128_maskz ((__v2di) __I
-              /* idx */ ,
-              (__v2di) __A,
-              (__v2di) __B,
-              (__mmask8)
-              __U);
+_mm_mask2_permutex2var_epi64(__m128i __A, __m128i __I, __mmask8 __U,
+                             __m128i __B) {
+  return (__m128i)__builtin_ia32_selectq_128(__U,
+                                  (__v2di)_mm_permutex2var_epi64(__A, __I, __B),
+                                  (__v2di)__I);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_permutex2var_epi64(__mmask8 __U, __m128i __A, __m128i __I,
+                             __m128i __B) {
+  return (__m128i)__builtin_ia32_selectq_128(__U,
+                                  (__v2di)_mm_permutex2var_epi64(__A, __I, __B),
+                                  (__v2di)_mm_setzero_si128());
 }
 
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_permutex2var_epi64 (__m256i __A, __m256i __I, __m256i __B) {
-  return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I
-                   /* idx */ ,
-                   (__v4di) __A,
-                   (__v4di) __B,
-                   (__mmask8) -1);
+_mm256_permutex2var_epi64(__m256i __A, __m256i __I, __m256i __B) {
+  return (__m256i)__builtin_ia32_vpermi2varq256((__v4di)__A, (__v4di) __I,
+                                                (__v4di) __B);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_mask_permutex2var_epi64 (__m256i __A, __mmask8 __U, __m256i __I,
-        __m256i __B) {
-  return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I
-                   /* idx */ ,
-                   (__v4di) __A,
-                   (__v4di) __B,
-                   (__mmask8) __U);
+_mm256_mask_permutex2var_epi64(__m256i __A, __mmask8 __U, __m256i __I,
+                               __m256i __B) {
+  return (__m256i)__builtin_ia32_selectq_256(__U,
+                               (__v4di)_mm256_permutex2var_epi64(__A, __I, __B),
+                               (__v4di)__A);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
-_mm256_maskz_permutex2var_epi64 (__mmask8 __U, __m256i __A,
-         __m256i __I, __m256i __B) {
-  return (__m256i) __builtin_ia32_vpermt2varq256_maskz ((__v4di) __I
-              /* idx */ ,
-              (__v4di) __A,
-              (__v4di) __B,
-              (__mmask8)
-              __U);
+_mm256_mask2_permutex2var_epi64(__m256i __A, __m256i __I, __mmask8 __U,
+                                __m256i __B) {
+  return (__m256i)__builtin_ia32_selectq_256(__U,
+                               (__v4di)_mm256_permutex2var_epi64(__A, __I, __B),
+                               (__v4di)__I);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_permutex2var_epi64(__mmask8 __U, __m256i __A, __m256i __I,
+                                __m256i __B) {
+  return (__m256i)__builtin_ia32_selectq_256(__U,
+                               (__v4di)_mm256_permutex2var_epi64(__A, __I, __B),
+                               (__v4di)_mm256_setzero_si256());
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS