[Clang][AVX512][intrinsics] Fix vperm intrinsics.

[igorb]

Summary

[Clang][AVX512][intrinsics] Fix vperm{w|d|q|ps|pd} intrinsics. Index is first argument to buildin function.

[m_zuckerman]

lgtm

[craig.topper]

Don't the type casts need to be changed too? For example permutexvar_pd has the index as __v8df.

[m_zuckerman]

Thanks, you are right !

[igorb]

Update path according to comments.
Thanks for review, Craig !

[craig.topper]

Looking at this again. This doesn't match the gcc implementation of the builtins. Unless their header file is also wrong. Can you clarify?

[igorb]

In D20321#436494, @craig.topper wrote:

Looking at this again. This doesn't match the gcc implementation of the builtins. Unless their header file is also wrong. Can you clarify?

Thanks,
You are correct. I implemented all changes in code-gen ( http://reviews.llvm.org/D20515 ).
If there's no objection, I abandon this review.

include/clang/Basic/BuiltinsX86.def

 TARGET_BUILTIN(__builtin_ia32_vfmaddsd3_mask,  "V2dV2dV2dV2dUcIi", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_vfmaddsd3_maskz, "V2dV2dV2dV2dUcIi", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_vfmaddsd3_mask3, "V2dV2dV2dV2dUcIi", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_permdf512_mask, "V8dV8dIiV8dUc","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_permdi512_mask, "V8LLiV8LLiIiV8LLiUc","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_permdf256_mask, "V4dV4dIiV4dUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_permdi256_mask, "V4LLiV4LLiIiV4LLiUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_permvarhi512_mask, "V32sV32sV32sV32sUi","","avx512bw")
-TARGET_BUILTIN(__builtin_ia32_permvardf512_mask, "V8dV8dV8LLiV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_permvardf512_mask, "V8dV8LLiV8dV8dUc","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_permvardi512_mask, "V8LLiV8LLiV8LLiV8LLiUc","","avx512f")
-TARGET_BUILTIN(__builtin_ia32_permvarsf512_mask, "V16fV16fV16iV16fUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_permvarsf512_mask, "V16fV16iV16fV16fUs","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_permvarsi512_mask, "V16iV16iV16iV16iUs","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_permvarqi512_mask, "V64cV64cV64cV64cULLi","","avx512vbmi")
 TARGET_BUILTIN(__builtin_ia32_permvarqi128_mask, "V16cV16cV16cV16cUs","","avx512vbmi,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_permvarqi256_mask, "V32cV32cV32cV32cUi","","avx512vbmi,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_permvarhi128_mask, "V8sV8sV8sV8sUc","","avx512bw,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_permvarhi256_mask, "V16sV16sV16sV16sUs","","avx512bw,avx512vl")
-TARGET_BUILTIN(__builtin_ia32_permvardf256_mask, "V4dV4dV4LLiV4dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_permvardf256_mask, "V4dV4LLiV4dV4dUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_permvardi256_mask, "V4LLiV4LLiV4LLiV4LLiUc","","avx512vl")
-TARGET_BUILTIN(__builtin_ia32_permvarsf256_mask, "V8fV8fV8iV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_permvarsf256_mask, "V8fV8iV8fV8fUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_permvarsi256_mask, "V8iV8iV8iV8iUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_fpclasspd128_mask, "UcV2dIiUc","","avx512dq,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_fpclasspd256_mask, "UcV4dIiUc","","avx512dq,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_fpclassps128_mask, "UcV4fIiUc","","avx512dq,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_fpclassps256_mask, "UcV8fIiUc","","avx512dq,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_fpclassps512_mask, "UsV16fIiUs","","avx512dq")
 TARGET_BUILTIN(__builtin_ia32_fpclasspd512_mask, "UcV8dIiUc","","avx512dq")
 TARGET_BUILTIN(__builtin_ia32_fpclasssd_mask, "UcV2dIiUc","","avx512dq")

lib/Headers/avx512bwintrin.h

_mm512_maskz_broadcastw_epi16 (__mmask32 __M, __m128i __A)
   return (__m512i) __builtin_ia32_pbroadcastw512_mask ((__v8hi) __A,
                    (__v32hi) _mm512_setzero_hi(),
                    __M);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_permutexvar_epi16 (__m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B,
-                 (__v32hi) __A,
+  return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __A /* idx */,
+                 (__v32hi) __B,
                  (__v32hi) _mm512_undefined_epi32 (),
                  (__mmask32) -1);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_maskz_permutexvar_epi16 (__mmask32 __M, __m512i __A,
         __m512i __B)
 {
-  return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B,
-                 (__v32hi) __A,
+  return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __A /* idx */,
+                 (__v32hi) __B,
                  (__v32hi) _mm512_setzero_hi(),
                  (__mmask32) __M);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_mask_permutexvar_epi16 (__m512i __W, __mmask32 __M, __m512i __A,
              __m512i __B)
 {
-  return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B,
-                 (__v32hi) __A,
+  return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __A /* idx */,
+                 (__v32hi) __B,
                  (__v32hi) __W,
                  (__mmask32) __M);
 }
 
 #define _mm512_alignr_epi8(A, B, N) __extension__ ({\
   (__m512i)__builtin_ia32_palignr512_mask((__v64qi)(__m512i)(A), \
                                           (__v64qi)(__m512i)(B), (int)(N) * 8, \
                                           (__v64qi)_mm512_undefined_pd(), \

lib/Headers/avx512fintrin.h

 #define _mm512_maskz_permutex_epi64(M, X, I) __extension__ ({ \
   (__m512i)__builtin_ia32_permdi512_mask((__v8di)(__m512i)(X), (int)(I), \
                                          (__v8di)_mm512_setzero_si512(), \
                                          (__mmask8)(M)); })
 
 static __inline__ __m512d __DEFAULT_FN_ATTRS
 _mm512_permutexvar_pd (__m512i __X, __m512d __Y)
 {
-  return (__m512d) __builtin_ia32_permvardf512_mask ((__v8df) __Y,
-                 (__v8di) __X,
+  return (__m512d) __builtin_ia32_permvardf512_mask ((__v8di) __X /* idx */,
+                 (__v8df) __Y,
                  (__v8df) _mm512_undefined_pd (),
                  (__mmask8) -1);
 }
 
 static __inline__ __m512d __DEFAULT_FN_ATTRS
 _mm512_mask_permutexvar_pd (__m512d __W, __mmask8 __U, __m512i __X, __m512d __Y)
 {
-  return (__m512d) __builtin_ia32_permvardf512_mask ((__v8df) __Y,
-                 (__v8di) __X,
+  return (__m512d) __builtin_ia32_permvardf512_mask ((__v8di) __X /* idx */,
+                 (__v8df) __Y,
                  (__v8df) __W,
                  (__mmask8) __U);
 }
 
 static __inline__ __m512d __DEFAULT_FN_ATTRS
 _mm512_maskz_permutexvar_pd (__mmask8 __U, __m512i __X, __m512d __Y)
 {
-  return (__m512d) __builtin_ia32_permvardf512_mask ((__v8df) __Y,
-                 (__v8di) __X,
+  return (__m512d) __builtin_ia32_permvardf512_mask ((__v8di) __X /* idx */,
+                 (__v8df) __Y,
                  (__v8df) _mm512_setzero_pd (),
                  (__mmask8) __U);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_maskz_permutexvar_epi64 (__mmask8 __M, __m512i __X, __m512i __Y)
 {
-  return (__m512i) __builtin_ia32_permvardi512_mask ((__v8di) __Y,
-                 (__v8di) __X,
+  return (__m512i) __builtin_ia32_permvardi512_mask ((__v8di) __X /* idx */,
+                 (__v8di) __Y,
                  (__v8di) _mm512_setzero_si512 (),
                  __M);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_permutexvar_epi64 (__m512i __X, __m512i __Y)
 {
-  return (__m512i) __builtin_ia32_permvardi512_mask ((__v8di) __Y,
-                 (__v8di) __X,
+  return (__m512i) __builtin_ia32_permvardi512_mask ((__v8di) __X /* idx */,
+                 (__v8di) __Y,
                  (__v8di) _mm512_undefined_epi32 (),
                  (__mmask8) -1);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_mask_permutexvar_epi64 (__m512i __W, __mmask8 __M, __m512i __X,
              __m512i __Y)
 {
-  return (__m512i) __builtin_ia32_permvardi512_mask ((__v8di) __Y,
-                 (__v8di) __X,
+  return (__m512i) __builtin_ia32_permvardi512_mask ((__v8di) __X /* idx */,
+                 (__v8di) __Y,
                  (__v8di) __W,
                  __M);
 }
 
 static __inline__ __m512 __DEFAULT_FN_ATTRS
 _mm512_permutexvar_ps (__m512i __X, __m512 __Y)
 {
-  return (__m512) __builtin_ia32_permvarsf512_mask ((__v16sf) __Y,
-                (__v16si) __X,
+  return (__m512) __builtin_ia32_permvarsf512_mask ((__v16si) __X /* idx */,
+                (__v16sf) __Y,
                 (__v16sf) _mm512_undefined_ps (),
                 (__mmask16) -1);
 }
 
 static __inline__ __m512 __DEFAULT_FN_ATTRS
 _mm512_mask_permutexvar_ps (__m512 __W, __mmask16 __U, __m512i __X, __m512 __Y)
 {
-  return (__m512) __builtin_ia32_permvarsf512_mask ((__v16sf) __Y,
-                (__v16si) __X,
+  return (__m512) __builtin_ia32_permvarsf512_mask ((__v16si) __X /* idx */,
+                (__v16sf) __Y,
                 (__v16sf) __W,
                 (__mmask16) __U);
 }
 
 static __inline__ __m512 __DEFAULT_FN_ATTRS
 _mm512_maskz_permutexvar_ps (__mmask16 __U, __m512i __X, __m512 __Y)
 {
-  return (__m512) __builtin_ia32_permvarsf512_mask ((__v16sf) __Y,
-                (__v16si) __X,
+  return (__m512) __builtin_ia32_permvarsf512_mask ((__v16si) __X /* idx */,
+                (__v16sf) __Y,
                 (__v16sf) _mm512_setzero_ps (),
                 (__mmask16) __U);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_maskz_permutexvar_epi32 (__mmask16 __M, __m512i __X, __m512i __Y)
 {
-  return (__m512i) __builtin_ia32_permvarsi512_mask ((__v16si) __Y,
-                 (__v16si) __X,
+  return (__m512i) __builtin_ia32_permvarsi512_mask ((__v16si) __X /* idx */,
+                 (__v16si) __Y,
                  (__v16si) _mm512_setzero_si512 (),
                  __M);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_permutexvar_epi32 (__m512i __X, __m512i __Y)
 {
-  return (__m512i) __builtin_ia32_permvarsi512_mask ((__v16si) __Y,
-                 (__v16si) __X,
+  return (__m512i) __builtin_ia32_permvarsi512_mask ((__v16si) __X /* idx */,
+                 (__v16si) __Y,
                  (__v16si) _mm512_undefined_epi32 (),
                  (__mmask16) -1);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_mask_permutexvar_epi32 (__m512i __W, __mmask16 __M, __m512i __X,
              __m512i __Y)
 {
-  return (__m512i) __builtin_ia32_permvarsi512_mask ((__v16si) __Y,
-                 (__v16si) __X,
+  return (__m512i) __builtin_ia32_permvarsi512_mask ((__v16si) __X /* idx */,
+                 (__v16si) __Y,
                  (__v16si) __W,
                  __M);
 }
 
 static __inline__ __mmask16 __DEFAULT_FN_ATTRS
 _mm512_kand (__mmask16 __A, __mmask16 __B)
 {
   return (__mmask16) __builtin_ia32_kandhi ((__mmask16) __A, (__mmask16) __B);

lib/Headers/avx512vlbwintrin.h

_mm_maskz_set1_epi16 (__mmask8 __M, short __A)
   return (__m128i) __builtin_ia32_pbroadcastw128_gpr_mask (__A,
                  (__v8hi) _mm_setzero_si128 (),
                  __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_permutexvar_epi16 (__m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_permvarhi128_mask ((__v8hi) __B,
-                 (__v8hi) __A,
+  return (__m128i) __builtin_ia32_permvarhi128_mask ((__v8hi) __A /* idx */,
+                 (__v8hi) __B,
                  (__v8hi) _mm_undefined_si128 (),
                  (__mmask8) -1);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_maskz_permutexvar_epi16 (__mmask8 __M, __m128i __A, __m128i __B)
 {
-  return (__m128i) __builtin_ia32_permvarhi128_mask ((__v8hi) __B,
-                 (__v8hi) __A,
+  return (__m128i) __builtin_ia32_permvarhi128_mask ((__v8hi) __A /* idx */,
+                 (__v8hi) __B,
                  (__v8hi) _mm_setzero_si128 (),
                  (__mmask8) __M);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_mask_permutexvar_epi16 (__m128i __W, __mmask8 __M, __m128i __A,
           __m128i __B)
 {
-  return (__m128i) __builtin_ia32_permvarhi128_mask ((__v8hi) __B,
-                 (__v8hi) __A,
+  return (__m128i) __builtin_ia32_permvarhi128_mask ((__v8hi) __A /* idx */,
+                 (__v8hi) __B,
                  (__v8hi) __W,
                  (__mmask8) __M);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
 _mm256_permutexvar_epi16 (__m256i __A, __m256i __B)
 {
-  return (__m256i) __builtin_ia32_permvarhi256_mask ((__v16hi) __B,
-                 (__v16hi) __A,
+  return (__m256i) __builtin_ia32_permvarhi256_mask ((__v16hi) __A /* idx */,
+                 (__v16hi) __B,
                  (__v16hi) _mm256_undefined_si256 (),
                  (__mmask16) -1);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
 _mm256_maskz_permutexvar_epi16 (__mmask16 __M, __m256i __A,
         __m256i __B)
 {
-  return (__m256i) __builtin_ia32_permvarhi256_mask ((__v16hi) __B,
-                 (__v16hi) __A,
+  return (__m256i) __builtin_ia32_permvarhi256_mask ((__v16hi) __A/* idx */,
+                 (__v16hi) __B,
                  (__v16hi) _mm256_setzero_si256 (),
                  (__mmask16) __M);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
 _mm256_mask_permutexvar_epi16 (__m256i __W, __mmask16 __M, __m256i __A,
              __m256i __B)
 {
-  return (__m256i) __builtin_ia32_permvarhi256_mask ((__v16hi) __B,
-                 (__v16hi) __A,
+  return (__m256i) __builtin_ia32_permvarhi256_mask ((__v16hi) __A /* idx */,
+                 (__v16hi) __B,
                  (__v16hi) __W,
                  (__mmask16) __M);
 }
 
 #define _mm_mask_alignr_epi8(W, U, A, B, N) __extension__ ({ \
   (__m128i)__builtin_ia32_palignr128_mask((__v16qi)(__m128i)(A), \
                                           (__v16qi)(__m128i)(B), (int)(N) * 8, \
                                           (__v16qi)(__m128i)(W), \

lib/Headers/avx512vlintrin.h

 #define _mm256_permutex_epi64(X, I) __extension__ ({ \
   (__m256i)__builtin_ia32_permdi256_mask((__v4di)(__m256i)(X), (int)(I), \
                                          (__v4di)_mm256_undefined_si256(), \
                                          (__mmask8)-1); })
 
 static __inline__ __m256d __DEFAULT_FN_ATTRS
 _mm256_permutexvar_pd (__m256i __X, __m256d __Y)
 {
-  return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y,
-                 (__v4di) __X,
+  return (__m256d) __builtin_ia32_permvardf256_mask ((__v4di) __X /* idx */,
+                 (__v4df) __Y,
                  (__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,
+  return (__m256d) __builtin_ia32_permvardf256_mask ((__v4di) __X /* idx */,
+                 (__v4df) __Y,
                  (__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,
+  return (__m256d) __builtin_ia32_permvardf256_mask ((__v4di) __X /* idx */,
+                 (__v4df) __Y,
                  (__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,
+  return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __X /* idx */,
+                 (__v4di) __Y,
                  (__v4di) _mm256_setzero_si256 (),
                  (__mmask8) __M);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
 _mm256_permutexvar_epi64 (__mmask8 __M, __m256i __X, __m256i __Y)
 {
-  return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y,
-                 (__v4di) __X,
+  return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __X /* idx */,
+                 (__v4di) __Y,
                  (__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,
+  return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __X /* idx */,
+                 (__v4di) __Y,
                  (__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,
+  return (__m256) __builtin_ia32_permvarsf256_mask ((__v8si) __X /* idx */,
+                (__v8sf) __Y,
                 (__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,
+  return (__m256) __builtin_ia32_permvarsf256_mask ((__v8si) __X /* idx */,
+                (__v8sf) __Y,
                 (__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,
+  return (__m256) __builtin_ia32_permvarsf256_mask ((__v8si) __X /* idx */,
+                (__v8sf) __Y,
                 (__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,
+  return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __X /* idx */,
+                 (__v8si) __Y,
                  (__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,
+  return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __X /* idx */,
+                 (__v8si) __Y,
                  (__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,
+  return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __X /* idx */,
+                 (__v8si) __Y,
                  (__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(), \