diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def
index 8404f2e1c241ac067b01263e49d49bafdde24265..44bb15359d87964c455347a649d60378e0481a28 100644
--- a/include/clang/Basic/BuiltinsX86.def
+++ b/include/clang/Basic/BuiltinsX86.def
@@ -2151,6 +2151,24 @@ TARGET_BUILTIN(__builtin_ia32_vfmaddss3_mask3, "V4fV4fV4fV4fUcIi", "", "avx512f"
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, "V8dV8dUcV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_permdi512_mask, "V8LLiV8LLiUcV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_permdf256_mask, "V4dV4dUcV4dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_permdi256_mask, "V4LLiV4LLiUcV4LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_permvarhi512_mask, "V32sV32sV32sV32sUi","","avx512bw")
+TARGET_BUILTIN(__builtin_ia32_permvardf512_mask, "V8dV8dV8LLiV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_permvardi512_mask, "V8LLiV8LLiV8LLiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_permvarsf512_mask, "V16fV16fV16iV16fUs","","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_permvardi256_mask, "V4LLiV4LLiV4LLiV4LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_permvarsf256_mask, "V8fV8fV8iV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_permvarsi256_mask, "V8iV8iV8iV8iUc","","avx512vl")
#undef BUILTIN
#undef TARGET_BUILTIN
diff --git a/lib/Headers/avx512bwintrin.h b/lib/Headers/avx512bwintrin.h
index ffaf496a1f0f695bde96b2ae64b745bf449a9f6b..e0307cdcf8678dd2ff706c2ced47389e2a39e679 100644
--- a/lib/Headers/avx512bwintrin.h
+++ b/lib/Headers/avx512bwintrin.h
@@ -2139,6 +2139,35 @@ _mm512_maskz_broadcastw_epi16 (__mmask32 __M, __m128i __A)
__M);
}
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_permutexvar_epi16 (__m512i __A, __m512i __B)
+{
+ return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B,
+ (__v32hi) __A,
+ (__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,
+ (__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,
+ (__v32hi) __W,
+ (__mmask32) __M);
+}
+
#undef __DEFAULT_FN_ATTRS
#endif
diff --git a/lib/Headers/avx512fintrin.h b/lib/Headers/avx512fintrin.h
index 171fdc91357aa44eda54a71cda15f2db67fa4da9..558eabf510a4465a9ac1bedb7efb6b7930b9d3b7 100644
--- a/lib/Headers/avx512fintrin.h
+++ b/lib/Headers/avx512fintrin.h
@@ -7048,6 +7048,154 @@ __builtin_ia32_vfmaddsd3_mask3 ((__v2df) -(__W),\
__R);\
})
+#define _mm512_permutex_pd( __X, __M) __extension__ ({ \
+__builtin_ia32_permdf512_mask ((__v8df)( __X),( __M),\
+ (__v8df) _mm512_undefined_pd (),\
+ (__mmask8) -1);\
+})
+
+#define _mm512_mask_permutex_pd( __W, __U, __X, __M) __extension__ ({ \
+__builtin_ia32_permdf512_mask ((__v8df)( __X),( __M),\
+ (__v8df)( __W),\
+ (__mmask8)( __U));\
+})
+
+#define _mm512_maskz_permutex_pd( __U, __X, __M) __extension__ ({ \
+__builtin_ia32_permdf512_mask ((__v8df)( __X),( __M),\
+ (__v8df) _mm512_setzero_pd (),\
+ (__mmask8)( __U));\
+})
+
+#define _mm512_permutex_epi64( __X, __I) __extension__ ({ \
+__builtin_ia32_permdi512_mask ((__v8di)( __X),( __I),\
+ (__v8di) _mm512_undefined_epi32 (),\
+ (__mmask8) (-1));\
+})
+
+#define _mm512_mask_permutex_epi64( __W, __M, __X, __I) __extension__ ({ \
+__builtin_ia32_permdi512_mask ((__v8di)( __X),( __I),\
+ (__v8di)( __W),\
+ (__mmask8)( __M));\
+})
+
+#define _mm512_maskz_permutex_epi64( __M, __X, __I) __extension__ ({ \
+__builtin_ia32_permdi512_mask ((__v8di)( __X),( __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,
+ (__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,
+ (__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,
+ (__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,
+ (__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,
+ (__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,
+ (__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,
+ (__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,
+ (__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,
+ (__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,
+ (__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,
+ (__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,
+ (__v16si) __W,
+ __M);
+}
+
+
+
#undef __DEFAULT_FN_ATTRS
#endif // __AVX512FINTRIN_H
diff --git a/lib/Headers/avx512vbmiintrin.h b/lib/Headers/avx512vbmiintrin.h
index 7ac9fe70ef0050071b3a8bff33353874335ceeb8..a2c3b8f2bf92b83a0fa2c955d4fe7ae59a69463f 100644
--- a/lib/Headers/avx512vbmiintrin.h
+++ b/lib/Headers/avx512vbmiintrin.h
@@ -79,6 +79,35 @@ _mm512_maskz_permutex2var_epi8 (__mmask64 __U, __m512i __A,
__U);
}
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_permutexvar_epi8 (__m512i __A, __m512i __B)
+{
+ return (__m512i) __builtin_ia32_permvarqi512_mask ((__v64qi) __B,
+ (__v64qi) __A,
+ (__v64qi) _mm512_undefined_epi32 (),
+ (__mmask64) -1);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_permutexvar_epi8 (__mmask64 __M, __m512i __A,
+ __m512i __B)
+{
+ return (__m512i) __builtin_ia32_permvarqi512_mask ((__v64qi) __B,
+ (__v64qi) __A,
+ (__v64qi) _mm512_setzero_si512(),
+ (__mmask64) __M);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask_permutexvar_epi8 (__m512i __W, __mmask64 __M, __m512i __A,
+ __m512i __B)
+{
+ return (__m512i) __builtin_ia32_permvarqi512_mask ((__v64qi) __B,
+ (__v64qi) __A,
+ (__v64qi) __W,
+ (__mmask64) __M);
+}
+
#undef __DEFAULT_FN_ATTRS
#endif
diff --git a/lib/Headers/avx512vbmivlintrin.h b/lib/Headers/avx512vbmivlintrin.h
index 04b7a47dd5ea53491f64c428802d2a2e666c41ce..57989699a86ff9f896c002ff418c878b99c11eb9 100644
--- a/lib/Headers/avx512vbmivlintrin.h
+++ b/lib/Headers/avx512vbmivlintrin.h
@@ -126,6 +126,62 @@ _mm256_maskz_permutex2var_epi8 (__mmask32 __U, __m256i __A,
__U);
}
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_permutexvar_epi8 (__m128i __A, __m128i __B)
+{
+ return (__m128i) __builtin_ia32_permvarqi128_mask ((__v16qi) __B,
+ (__v16qi) __A,
+ (__v16qi) _mm_undefined_si128 (),
+ (__mmask16) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_permutexvar_epi8 (__mmask16 __M, __m128i __A, __m128i __B)
+{
+ return (__m128i) __builtin_ia32_permvarqi128_mask ((__v16qi) __B,
+ (__v16qi) __A,
+ (__v16qi) _mm_setzero_si128 (),
+ (__mmask16) __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_permutexvar_epi8 (__m128i __W, __mmask16 __M, __m128i __A,
+ __m128i __B)
+{
+ return (__m128i) __builtin_ia32_permvarqi128_mask ((__v16qi) __B,
+ (__v16qi) __A,
+ (__v16qi) __W,
+ (__mmask16) __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_permutexvar_epi8 (__m256i __A, __m256i __B)
+{
+ return (__m256i) __builtin_ia32_permvarqi256_mask ((__v32qi) __B,
+ (__v32qi) __A,
+ (__v32qi) _mm256_undefined_si256 (),
+ (__mmask32) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_maskz_permutexvar_epi8 (__mmask32 __M, __m256i __A,
+ __m256i __B)
+{
+ return (__m256i) __builtin_ia32_permvarqi256_mask ((__v32qi) __B,
+ (__v32qi) __A,
+ (__v32qi) _mm256_setzero_si256 (),
+ (__mmask32) __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_mask_permutexvar_epi8 (__m256i __W, __mmask32 __M, __m256i __A,
+ __m256i __B)
+{
+ return (__m256i) __builtin_ia32_permvarqi256_mask ((__v32qi) __B,
+ (__v32qi) __A,
+ (__v32qi) __W,
+ (__mmask32) __M);
+}
#undef __DEFAULT_FN_ATTRS
diff --git a/lib/Headers/avx512vlbwintrin.h b/lib/Headers/avx512vlbwintrin.h
index 2c11af48c6dbee76d134b4b3234321a441211aa7..361df9391163f5cebb73c2b68b8407b989ade5aa 100644
--- a/lib/Headers/avx512vlbwintrin.h
+++ b/lib/Headers/avx512vlbwintrin.h
@@ -3301,7 +3301,62 @@ _mm_maskz_set1_epi16 (__mmask8 __M, short __A)
__M);
}
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_permutexvar_epi16 (__m128i __A, __m128i __B)
+{
+ return (__m128i) __builtin_ia32_permvarhi128_mask ((__v8hi) __B,
+ (__v8hi) __A,
+ (__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,
+ (__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,
+ (__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,
+ (__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,
+ (__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,
+ (__v16hi) __W,
+ (__mmask16) __M);
+}
#undef __DEFAULT_FN_ATTRS
diff --git a/lib/Headers/avx512vlintrin.h b/lib/Headers/avx512vlintrin.h
index a88fa996b302abd4634a1c1e46a13930f3306e1a..77d98b887ff573ebbe88e80b5ff282684340f231 100644
--- a/lib/Headers/avx512vlintrin.h
+++ b/lib/Headers/avx512vlintrin.h
@@ -9058,6 +9058,157 @@ __builtin_ia32_gather3siv8si ((__v8si) __v1_old, __addr, (__v8si) __index,\
__mask, __scale);\
})
+#define _mm256_mask_permutex_pd( __W, __U, __X, __imm) __extension__ ({ \
+__builtin_ia32_permdf256_mask ((__v4df)( __X),( __imm),\
+ (__v4df)( __W),\
+ (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_permutex_pd( __U, __X, __imm) __extension__ ({ \
+__builtin_ia32_permdf256_mask ((__v4df)( __X),( __imm),\
+ (__v4df) _mm256_setzero_pd (),\
+ (__mmask8)( __U));\
+})
+
+#define _mm256_permutex_pd( __X, __M) __extension__ ({ \
+__builtin_ia32_permdf256_mask ((__v4df)( __X),( __M),\
+ (__v4df) _mm256_undefined_pd (),\
+ (__mmask8) -1);\
+})
+
+#define _mm256_mask_permutex_epi64( __W, __M, __X, __I) __extension__ ({ \
+__builtin_ia32_permdi256_mask ((__v4di)( __X),\
+ ( __I),\
+ (__v4di)( __W),\
+ (__mmask8)( __M));\
+})
+
+#define _mm256_maskz_permutex_epi64( __M, __X, __I) __extension__ ({ \
+__builtin_ia32_permdi256_mask ((__v4di)( __X),\
+ ( __I),\
+ (__v4di) _mm256_setzero_si256 (),\
+ (__mmask8)( __M));\
+})
+
+#define _mm256_permutex_epi64( __X, __I) __extension__ ({ \
+__builtin_ia32_permdi256_mask ((__v4di)( __X),\
+ ( __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,
+ (__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 (__mmask8 __M, __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);
+}
+
#undef __DEFAULT_FN_ATTRS
#undef __DEFAULT_FN_ATTRS_BOTH
diff --git a/test/CodeGen/avx512bw-builtins.c b/test/CodeGen/avx512bw-builtins.c
index 5de57929dc690f9c20764307bc9b55e4af09e65d..3023e6060044919f78c8c3ef1612686cab1e6e54 100644
--- a/test/CodeGen/avx512bw-builtins.c
+++ b/test/CodeGen/avx512bw-builtins.c
@@ -1470,3 +1470,20 @@ __m512i test_mm512_maskz_set1_epi16(__mmask32 __M, short __A) {
// CHECK: @llvm.x86.avx512.mask.pbroadcast.w.gpr.512
return _mm512_maskz_set1_epi16(__M, __A);
}
+__m512i test_mm512_permutexvar_epi16(__m512i __A, __m512i __B) {
+ // CHECK-LABEL: @test_mm512_permutexvar_epi16
+ // CHECK: @llvm.x86.avx512.mask.permvar.hi.512
+ return _mm512_permutexvar_epi16(__A, __B);
+}
+
+__m512i test_mm512_maskz_permutexvar_epi16(__mmask32 __M, __m512i __A, __m512i __B) {
+ // CHECK-LABEL: @test_mm512_maskz_permutexvar_epi16
+ // CHECK: @llvm.x86.avx512.mask.permvar.hi.512
+ return _mm512_maskz_permutexvar_epi16(__M, __A, __B);
+}
+
+__m512i test_mm512_mask_permutexvar_epi16(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) {
+ // CHECK-LABEL: @test_mm512_mask_permutexvar_epi16
+ // CHECK: @llvm.x86.avx512.mask.permvar.hi.512
+ return _mm512_mask_permutexvar_epi16(__W, __M, __A, __B);
+}
diff --git a/test/CodeGen/avx512f-builtins.c b/test/CodeGen/avx512f-builtins.c
index 21d116cfe1f579dbd9d9ecea86eb11574226e335..e869f51872f61e409e26ccc595b9d5dc23a48ee8 100644
--- a/test/CodeGen/avx512f-builtins.c
+++ b/test/CodeGen/avx512f-builtins.c
@@ -4801,3 +4801,111 @@ __m128d test_mm_mask3_fnmsub_round_sd(__m128d __W, __m128d __X, __m128d __Y, __m
// CHECK: @llvm.x86.avx512.mask3.vfmadd.sd
return _mm_mask3_fnmsub_round_sd(__W, __X, __Y, __U, _MM_FROUND_CUR_DIRECTION);
}
+
+__m512d test_mm512_permutex_pd(__m512d __X) {
+ // CHECK-LABEL: @test_mm512_permutex_pd
+ // CHECK: @llvm.x86.avx512.mask.perm.df.512
+ return _mm512_permutex_pd(__X, 0);
+}
+
+__m512d test_mm512_mask_permutex_pd(__m512d __W, __mmask8 __U, __m512d __X) {
+ // CHECK-LABEL: @test_mm512_mask_permutex_pd
+ // CHECK: @llvm.x86.avx512.mask.perm.df.512
+ return _mm512_mask_permutex_pd(__W, __U, __X, 0);
+}
+
+__m512d test_mm512_maskz_permutex_pd(__mmask8 __U, __m512d __X) {
+ // CHECK-LABEL: @test_mm512_maskz_permutex_pd
+ // CHECK: @llvm.x86.avx512.mask.perm.df.512
+ return _mm512_maskz_permutex_pd(__U, __X, 0);
+}
+
+__m512i test_mm512_permutex_epi64(__m512i __X) {
+ // CHECK-LABEL: @test_mm512_permutex_epi64
+ // CHECK: @llvm.x86.avx512.mask.perm.di.512
+ return _mm512_permutex_epi64(__X, 0);
+}
+
+__m512i test_mm512_mask_permutex_epi64(__m512i __W, __mmask8 __M, __m512i __X) {
+ // CHECK-LABEL: @test_mm512_mask_permutex_epi64
+ // CHECK: @llvm.x86.avx512.mask.perm.di.512
+ return _mm512_mask_permutex_epi64(__W, __M, __X, 0);
+}
+
+__m512i test_mm512_maskz_permutex_epi64(__mmask8 __M, __m512i __X) {
+ // CHECK-LABEL: @test_mm512_maskz_permutex_epi64
+ // CHECK: @llvm.x86.avx512.mask.perm.di.512
+ return _mm512_maskz_permutex_epi64(__M, __X, 0);
+}
+
+__m512d test_mm512_permutexvar_pd(__m512i __X, __m512d __Y) {
+ // CHECK-LABEL: @test_mm512_permutexvar_pd
+ // CHECK: @llvm.x86.avx512.mask.permvar.df.512
+ return _mm512_permutexvar_pd(__X, __Y);
+}
+
+__m512d test_mm512_mask_permutexvar_pd(__m512d __W, __mmask8 __U, __m512i __X, __m512d __Y) {
+ // CHECK-LABEL: @test_mm512_mask_permutexvar_pd
+ // CHECK: @llvm.x86.avx512.mask.permvar.df.512
+ return _mm512_mask_permutexvar_pd(__W, __U, __X, __Y);
+}
+
+__m512d test_mm512_maskz_permutexvar_pd(__mmask8 __U, __m512i __X, __m512d __Y) {
+ // CHECK-LABEL: @test_mm512_maskz_permutexvar_pd
+ // CHECK: @llvm.x86.avx512.mask.permvar.df.512
+ return _mm512_maskz_permutexvar_pd(__U, __X, __Y);
+}
+
+__m512i test_mm512_maskz_permutexvar_epi64(__mmask8 __M, __m512i __X, __m512i __Y) {
+ // CHECK-LABEL: @test_mm512_maskz_permutexvar_epi64
+ // CHECK: @llvm.x86.avx512.mask.permvar.di.512
+ return _mm512_maskz_permutexvar_epi64(__M, __X, __Y);
+}
+
+__m512i test_mm512_permutexvar_epi64(__m512i __X, __m512i __Y) {
+ // CHECK-LABEL: @test_mm512_permutexvar_epi64
+ // CHECK: @llvm.x86.avx512.mask.permvar.di.512
+ return _mm512_permutexvar_epi64(__X, __Y);
+}
+
+__m512i test_mm512_mask_permutexvar_epi64(__m512i __W, __mmask8 __M, __m512i __X, __m512i __Y) {
+ // CHECK-LABEL: @test_mm512_mask_permutexvar_epi64
+ // CHECK: @llvm.x86.avx512.mask.permvar.di.512
+ return _mm512_mask_permutexvar_epi64(__W, __M, __X, __Y);
+}
+
+__m512 test_mm512_permutexvar_ps(__m512i __X, __m512 __Y) {
+ // CHECK-LABEL: @test_mm512_permutexvar_ps
+ // CHECK: @llvm.x86.avx512.mask.permvar.sf.512
+ return _mm512_permutexvar_ps(__X, __Y);
+}
+
+__m512 test_mm512_mask_permutexvar_ps(__m512 __W, __mmask16 __U, __m512i __X, __m512 __Y) {
+ // CHECK-LABEL: @test_mm512_mask_permutexvar_ps
+ // CHECK: @llvm.x86.avx512.mask.permvar.sf.512
+ return _mm512_mask_permutexvar_ps(__W, __U, __X, __Y);
+}
+
+__m512 test_mm512_maskz_permutexvar_ps(__mmask16 __U, __m512i __X, __m512 __Y) {
+ // CHECK-LABEL: @test_mm512_maskz_permutexvar_ps
+ // CHECK: @llvm.x86.avx512.mask.permvar.sf.512
+ return _mm512_maskz_permutexvar_ps(__U, __X, __Y);
+}
+
+__m512i test_mm512_maskz_permutexvar_epi32(__mmask16 __M, __m512i __X, __m512i __Y) {
+ // CHECK-LABEL: @test_mm512_maskz_permutexvar_epi32
+ // CHECK: @llvm.x86.avx512.mask.permvar.si.512
+ return _mm512_maskz_permutexvar_epi32(__M, __X, __Y);
+}
+
+__m512i test_mm512_permutexvar_epi32(__m512i __X, __m512i __Y) {
+ // CHECK-LABEL: @test_mm512_permutexvar_epi32
+ // CHECK: @llvm.x86.avx512.mask.permvar.si.512
+ return _mm512_permutexvar_epi32(__X, __Y);
+}
+
+__m512i test_mm512_mask_permutexvar_epi32(__m512i __W, __mmask16 __M, __m512i __X, __m512i __Y) {
+ // CHECK-LABEL: @test_mm512_mask_permutexvar_epi32
+ // CHECK: @llvm.x86.avx512.mask.permvar.si.512
+ return _mm512_mask_permutexvar_epi32(__W, __M, __X, __Y);
+}
diff --git a/test/CodeGen/avx512vbmi-builtins.c b/test/CodeGen/avx512vbmi-builtins.c
index 73ed969e48e378190a61d12727b962b646488dfb..c1aa04cc191150be0459b8b45ed38b2f98a98686 100644
--- a/test/CodeGen/avx512vbmi-builtins.c
+++ b/test/CodeGen/avx512vbmi-builtins.c
@@ -25,7 +25,24 @@ __m512i test_mm512_mask_permutex2var_epi8(__m512i __A, __mmask64 __U, __m512i __
__m512i test_mm512_maskz_permutex2var_epi8(__mmask64 __U, __m512i __A, __m512i __I, __m512i __B) {
// CHECK-LABEL: @test_mm512_maskz_permutex2var_epi8
- // CHECK: @llvm.x86.avx512.mask.vpermt2var.qi.512
+ // CHECK: @llvm.x86.avx512.maskz.vpermt2var.qi.512
return _mm512_maskz_permutex2var_epi8(__U, __A, __I, __B);
}
+__m512i test_mm512_permutexvar_epi8(__m512i __A, __m512i __B) {
+ // CHECK-LABEL: @test_mm512_permutexvar_epi8
+ // CHECK: @llvm.x86.avx512.mask.permvar.qi.512
+ return _mm512_permutexvar_epi8(__A, __B);
+}
+
+__m512i test_mm512_maskz_permutexvar_epi8(__mmask64 __M, __m512i __A, __m512i __B) {
+ // CHECK-LABEL: @test_mm512_maskz_permutexvar_epi8
+ // CHECK: @llvm.x86.avx512.mask.permvar.qi.512
+ return _mm512_maskz_permutexvar_epi8(__M, __A, __B);
+}
+
+__m512i test_mm512_mask_permutexvar_epi8(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) {
+ // CHECK-LABEL: @test_mm512_mask_permutexvar_epi8
+ // CHECK: @llvm.x86.avx512.mask.permvar.qi.512
+ return _mm512_mask_permutexvar_epi8(__W, __M, __A, __B);
+}
diff --git a/test/CodeGen/avx512vbmivl-builtin.c b/test/CodeGen/avx512vbmivl-builtin.c
index 90c5b4cc74aabeeb045dcd403a27486593cc7785..29864eccdf257ec3e5f0f91dad1ea248df453ed7 100644
--- a/test/CodeGen/avx512vbmivl-builtin.c
+++ b/test/CodeGen/avx512vbmivl-builtin.c
@@ -5,6 +5,41 @@
#include <immintrin.h>
+__m128i test_mm_permutexvar_epi8(__m128i __A, __m128i __B) {
+ // CHECK-LABEL: @test_mm_permutexvar_epi8
+ // CHECK: @llvm.x86.avx512.mask.permvar.qi.128
+ return _mm_permutexvar_epi8(__A, __B);
+}
+
+__m128i test_mm_maskz_permutexvar_epi8(__mmask16 __M, __m128i __A, __m128i __B) {
+ // CHECK-LABEL: @test_mm_maskz_permutexvar_epi8
+ // CHECK: @llvm.x86.avx512.mask.permvar.qi.128
+ return _mm_maskz_permutexvar_epi8(__M, __A, __B);
+}
+
+__m128i test_mm_mask_permutexvar_epi8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) {
+ // CHECK-LABEL: @test_mm_mask_permutexvar_epi8
+ // CHECK: @llvm.x86.avx512.mask.permvar.qi.128
+ return _mm_mask_permutexvar_epi8(__W, __M, __A, __B);
+}
+
+__m256i test_mm256_permutexvar_epi8(__m256i __A, __m256i __B) {
+ // CHECK-LABEL: @test_mm256_permutexvar_epi8
+ // CHECK: @llvm.x86.avx512.mask.permvar.qi.256
+ return _mm256_permutexvar_epi8(__A, __B);
+}
+
+__m256i test_mm256_maskz_permutexvar_epi8(__mmask32 __M, __m256i __A, __m256i __B) {
+ // CHECK-LABEL: @test_mm256_maskz_permutexvar_epi8
+ // CHECK: @llvm.x86.avx512.mask.permvar.qi.256
+ return _mm256_maskz_permutexvar_epi8(__M, __A, __B);
+}
+
+__m256i test_mm256_mask_permutexvar_epi8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) {
+ // CHECK-LABEL: @test_mm256_mask_permutexvar_epi8
+ // CHECK: @llvm.x86.avx512.mask.permvar.qi.256
+ return _mm256_mask_permutexvar_epi8(__W, __M, __A, __B);
+}
__m128i test_mm_mask2_permutex2var_epi8(__m128i __A, __m128i __I, __mmask16 __U, __m128i __B) {
// CHECK-LABEL: @test_mm_mask2_permutex2var_epi8
diff --git a/test/CodeGen/avx512vl-builtins.c b/test/CodeGen/avx512vl-builtins.c
index 167d79bd2d515108078434d05d9f72a118d88b3a..9ba949e72963d6e4d11648e2c387afe77887bb69 100644
--- a/test/CodeGen/avx512vl-builtins.c
+++ b/test/CodeGen/avx512vl-builtins.c
@@ -6358,3 +6358,106 @@ __m256i test_mm256_mask_i32gather_epi32(__m256i __v1_old, __mmask8 __mask, __m25
// CHECK: @llvm.x86.avx512.gather3siv8.si
return _mm256_mmask_i32gather_epi32(__v1_old, __mask, __index, __addr, 2);
}
+
+__m256d test_mm256_mask_permutex_pd(__m256d __W, __mmask8 __U, __m256d __X) {
+ // CHECK-LABEL: @test_mm256_mask_permutex_pd
+ // CHECK: @llvm.x86.avx512.mask.perm.df.256
+ return _mm256_mask_permutex_pd(__W, __U, __X, 1);
+}
+
+__m256d test_mm256_maskz_permutex_pd(__mmask8 __U, __m256d __X) {
+ // CHECK-LABEL: @test_mm256_maskz_permutex_pd
+ // CHECK: @llvm.x86.avx512.mask.perm.df.256
+ return _mm256_maskz_permutex_pd(__U, __X, 1);
+}
+
+__m256d test_mm256_permutex_pd(__m256d __X) {
+ // CHECK-LABEL: @test_mm256_permutex_pd
+ // CHECK: @llvm.x86.avx512.mask.perm.df.256
+ return _mm256_permutex_pd(__X, 3);
+}
+
+
+__m256i test_mm256_mask_permutex_epi64(__m256i __W, __mmask8 __M, __m256i __X) {
+ // CHECK-LABEL: @test_mm256_mask_permutex_epi64
+ // CHECK: @llvm.x86.avx512.mask.perm.di.256
+ return _mm256_mask_permutex_epi64(__W, __M, __X, 3);
+}
+
+__m256i test_mm256_maskz_permutex_epi64(__mmask8 __M, __m256i __X) {
+ // CHECK-LABEL: @test_mm256_maskz_permutex_epi64
+ // CHECK: @llvm.x86.avx512.mask.perm.di.256
+ return _mm256_maskz_permutex_epi64(__M, __X, 3);
+}
+
+__m256i test_mm256_permutex_epi64( __m256i __X) {
+ // CHECK-LABEL: @test_mm256_permutex_epi64
+ // CHECK: @llvm.x86.avx512.mask.perm.di.256
+ return _mm256_permutex_epi64(__X, 3);
+}
+
+__m256d test_mm256_permutexvar_pd(__m256i __X, __m256d __Y) {
+ // CHECK-LABEL: @test_mm256_permutexvar_pd
+ // CHECK: @llvm.x86.avx512.mask.permvar.df.256
+ return _mm256_permutexvar_pd(__X, __Y);
+}
+
+__m256d test_mm256_mask_permutexvar_pd(__m256d __W, __mmask8 __U, __m256i __X, __m256d __Y) {
+ // CHECK-LABEL: @test_mm256_mask_permutexvar_pd
+ // CHECK: @llvm.x86.avx512.mask.permvar.df.256
+ return _mm256_mask_permutexvar_pd(__W, __U, __X, __Y);
+}
+
+__m256d test_mm256_maskz_permutexvar_pd(__mmask8 __U, __m256i __X, __m256d __Y) {
+ // CHECK-LABEL: @test_mm256_maskz_permutexvar_pd
+ // CHECK: @llvm.x86.avx512.mask.permvar.df.256
+ return _mm256_maskz_permutexvar_pd(__U, __X, __Y);
+}
+
+__m256i test_mm256_maskz_permutexvar_epi64(__mmask8 __M, __m256i __X, __m256i __Y) {
+ // CHECK-LABEL: @test_mm256_maskz_permutexvar_epi64
+ // CHECK: @llvm.x86.avx512.mask.permvar.di.256
+ return _mm256_maskz_permutexvar_epi64(__M, __X, __Y);
+}
+
+__m256i test_mm256_mask_permutexvar_epi64(__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y) {
+ // CHECK-LABEL: @test_mm256_mask_permutexvar_epi64
+ // CHECK: @llvm.x86.avx512.mask.permvar.di.256
+ return _mm256_mask_permutexvar_epi64(__W, __M, __X, __Y);
+}
+
+__m256 test_mm256_mask_permutexvar_ps(__m256 __W, __mmask8 __U, __m256i __X, __m256 __Y) {
+ // CHECK-LABEL: @test_mm256_mask_permutexvar_ps
+ // CHECK: @llvm.x86.avx512.mask.permvar.sf.256
+ return _mm256_mask_permutexvar_ps(__W, __U, __X, __Y);
+}
+
+__m256 test_mm256_maskz_permutexvar_ps(__mmask8 __U, __m256i __X, __m256 __Y) {
+ // CHECK-LABEL: @test_mm256_maskz_permutexvar_ps
+ // CHECK: @llvm.x86.avx512.mask.permvar.sf.256
+ return _mm256_maskz_permutexvar_ps(__U, __X, __Y);
+}
+
+__m256 test_mm256_permutexvar_ps(__m256i __X, __m256 __Y) {
+ // CHECK-LABEL: @test_mm256_permutexvar_ps
+ // CHECK: @llvm.x86.avx512.mask.permvar.sf.256
+ return _mm256_permutexvar_ps( __X, __Y);
+}
+
+__m256i test_mm256_maskz_permutexvar_epi32(__mmask8 __M, __m256i __X, __m256i __Y) {
+ // CHECK-LABEL: @test_mm256_maskz_permutexvar_epi32
+ // CHECK: @llvm.x86.avx512.mask.permvar.si.256
+ return _mm256_maskz_permutexvar_epi32(__M, __X, __Y);
+}
+
+__m256i test_mm256_permutexvar_epi32(__m256i __X, __m256i __Y) {
+ // CHECK-LABEL: @test_mm256_permutexvar_epi32
+ // CHECK: @llvm.x86.avx512.mask.permvar.si.256
+ return _mm256_permutexvar_epi32(__X, __Y);
+}
+
+__m256i test_mm256_mask_permutexvar_epi32(__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y) {
+ // CHECK-LABEL: @test_mm256_mask_permutexvar_epi32
+ // CHECK: @llvm.x86.avx512.mask.permvar.si.256
+ return _mm256_mask_permutexvar_epi32(__W, __M, __X, __Y);
+}
diff --git a/test/CodeGen/avx512vlbw-builtins.c b/test/CodeGen/avx512vlbw-builtins.c
index b063badf2b50eb59e012843a959db2b9d29bc794..f05b32d2fe6f0740888b1cace6fe65772fa1b0dc 100644
--- a/test/CodeGen/avx512vlbw-builtins.c
+++ b/test/CodeGen/avx512vlbw-builtins.c
@@ -2281,3 +2281,38 @@ __m128i test_mm_maskz_set1_epi16(__mmask8 __M, short __A) {
// CHECK: @llvm.x86.avx512.mask.pbroadcast.w.gpr.128
return _mm_maskz_set1_epi16(__M, __A);
}
+__m128i test_mm_permutexvar_epi16(__m128i __A, __m128i __B) {
+ // CHECK-LABEL: @test_mm_permutexvar_epi16
+ // CHECK: @llvm.x86.avx512.mask.permvar.hi.128
+ return _mm_permutexvar_epi16(__A, __B);
+}
+
+__m128i test_mm_maskz_permutexvar_epi16(__mmask8 __M, __m128i __A, __m128i __B) {
+ // CHECK-LABEL: @test_mm_maskz_permutexvar_epi16
+ // CHECK: @llvm.x86.avx512.mask.permvar.hi.128
+ return _mm_maskz_permutexvar_epi16(__M, __A, __B);
+}
+
+__m128i test_mm_mask_permutexvar_epi16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
+ // CHECK-LABEL: @test_mm_mask_permutexvar_epi16
+ // CHECK: @llvm.x86.avx512.mask.permvar.hi.128
+ return _mm_mask_permutexvar_epi16(__W, __M, __A, __B);
+}
+
+__m256i test_mm256_permutexvar_epi16(__m256i __A, __m256i __B) {
+ // CHECK-LABEL: @test_mm256_permutexvar_epi16
+ // CHECK: @llvm.x86.avx512.mask.permvar.hi.256
+ return _mm256_permutexvar_epi16(__A, __B);
+}
+
+__m256i test_mm256_maskz_permutexvar_epi16(__mmask16 __M, __m256i __A, __m256i __B) {
+ // CHECK-LABEL: @test_mm256_maskz_permutexvar_epi16
+ // CHECK: @llvm.x86.avx512.mask.permvar.hi.256
+ return _mm256_maskz_permutexvar_epi16(__M, __A, __B);
+}
+
+__m256i test_mm256_mask_permutexvar_epi16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) {
+ // CHECK-LABEL: @test_mm256_mask_permutexvar_epi16
+ // CHECK: @llvm.x86.avx512.mask.permvar.hi.256
+ return _mm256_mask_permutexvar_epi16(__W, __M, __A, __B);
+}