diff options
Diffstat (limited to 'lib/Headers')
-rw-r--r-- | lib/Headers/avx512bwintrin.h | 205 | ||||
-rw-r--r-- | lib/Headers/avx512fintrin.h | 196 | ||||
-rw-r--r-- | lib/Headers/avx512vlintrin.h | 236 |
3 files changed, 233 insertions, 404 deletions
diff --git a/lib/Headers/avx512bwintrin.h b/lib/Headers/avx512bwintrin.h index fc0faa0d62..ed04a55fd6 100644 --- a/lib/Headers/avx512bwintrin.h +++ b/lib/Headers/avx512bwintrin.h @@ -310,49 +310,45 @@ _mm512_mask_blend_epi16 (__mmask32 __U, __m512i __A, __m512i __W) static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_abs_epi8 (__m512i __A) { - return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A, - (__v64qi) _mm512_setzero_qi(), - (__mmask64) -1); + return (__m512i)__builtin_ia32_pabsb512((__v64qi)__A); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_abs_epi8 (__m512i __W, __mmask64 __U, __m512i __A) { - return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A, - (__v64qi) __W, - (__mmask64) __U); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, + (__v64qi)_mm512_abs_epi8(__A), + (__v64qi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_abs_epi8 (__mmask64 __U, __m512i __A) { - return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A, - (__v64qi) _mm512_setzero_qi(), - (__mmask64) __U); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, + (__v64qi)_mm512_abs_epi8(__A), + (__v64qi)_mm512_setzero_qi()); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_abs_epi16 (__m512i __A) { - return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A, - (__v32hi) _mm512_setzero_hi(), - (__mmask32) -1); + return (__m512i)__builtin_ia32_pabsw512((__v32hi)__A); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_abs_epi16 (__m512i __W, __mmask32 __U, __m512i __A) { - return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A, - (__v32hi) __W, - (__mmask32) __U); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, + (__v32hi)_mm512_abs_epi16(__A), + (__v32hi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_abs_epi16 (__mmask32 __U, __m512i __A) { - return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A, - (__v32hi) _mm512_setzero_hi(), - (__mmask32) __U); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, + (__v32hi)_mm512_abs_epi16(__A), + (__v32hi)_mm512_setzero_hi()); } static __inline__ __m512i __DEFAULT_FN_ATTRS @@ -612,225 +608,178 @@ _mm512_maskz_avg_epu16 (__mmask32 __U, __m512i __A, __m512i __B) static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_max_epi8 (__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) _mm512_setzero_qi(), - (__mmask64) -1); + return (__m512i)__builtin_ia32_pmaxsb512((__v64qi) __A, (__v64qi) __B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_max_epi8 (__mmask64 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) _mm512_setzero_qi(), - (__mmask64) __M); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, + (__v64qi)_mm512_max_epi8(__A, __B), + (__v64qi)_mm512_setzero_qi()); } static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_max_epi8 (__m512i __W, __mmask64 __M, __m512i __A, - __m512i __B) +_mm512_mask_max_epi8 (__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) __W, - (__mmask64) __M); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, + (__v64qi)_mm512_max_epi8(__A, __B), + (__v64qi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_max_epi16 (__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) _mm512_setzero_hi(), - (__mmask32) -1); + return (__m512i)__builtin_ia32_pmaxsw512((__v32hi) __A, (__v32hi) __B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_max_epi16 (__mmask32 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) _mm512_setzero_hi(), - (__mmask32) __M); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, + (__v32hi)_mm512_max_epi16(__A, __B), + (__v32hi)_mm512_setzero_hi()); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_max_epi16 (__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) __W, - (__mmask32) __M); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, + (__v32hi)_mm512_max_epi16(__A, __B), + (__v32hi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_max_epu8 (__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) _mm512_setzero_qi(), - (__mmask64) -1); + return (__m512i)__builtin_ia32_pmaxub512((__v64qi)__A, (__v64qi)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_max_epu8 (__mmask64 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) _mm512_setzero_qi(), - (__mmask64) __M); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, + (__v64qi)_mm512_max_epu8(__A, __B), + (__v64qi)_mm512_setzero_qi()); } static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_max_epu8 (__m512i __W, __mmask64 __M, __m512i __A, - __m512i __B) +_mm512_mask_max_epu8 (__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) __W, - (__mmask64) __M); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, + (__v64qi)_mm512_max_epu8(__A, __B), + (__v64qi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_max_epu16 (__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) _mm512_setzero_hi(), - (__mmask32) -1); + return (__m512i)__builtin_ia32_pmaxuw512((__v32hi)__A, (__v32hi)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_max_epu16 (__mmask32 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) _mm512_setzero_hi(), - (__mmask32) __M); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, + (__v32hi)_mm512_max_epu16(__A, __B), + (__v32hi)_mm512_setzero_hi()); } static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_max_epu16 (__m512i __W, __mmask32 __M, __m512i __A, - __m512i __B) +_mm512_mask_max_epu16 (__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) __W, - (__mmask32) __M); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, + (__v32hi)_mm512_max_epu16(__A, __B), + (__v32hi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_min_epi8 (__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) _mm512_setzero_qi(), - (__mmask64) -1); + return (__m512i)__builtin_ia32_pminsb512((__v64qi) __A, (__v64qi) __B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_min_epi8 (__mmask64 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) _mm512_setzero_qi(), - (__mmask64) __M); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, + (__v64qi)_mm512_min_epi8(__A, __B), + (__v64qi)_mm512_setzero_qi()); } static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_min_epi8 (__m512i __W, __mmask64 __M, __m512i __A, - __m512i __B) +_mm512_mask_min_epi8 (__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) __W, - (__mmask64) __M); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, + (__v64qi)_mm512_min_epi8(__A, __B), + (__v64qi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_min_epi16 (__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) _mm512_setzero_hi(), - (__mmask32) -1); + return (__m512i)__builtin_ia32_pminsw512((__v32hi) __A, (__v32hi) __B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_min_epi16 (__mmask32 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) _mm512_setzero_hi(), - (__mmask32) __M); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, + (__v32hi)_mm512_min_epi16(__A, __B), + (__v32hi)_mm512_setzero_hi()); } static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_min_epi16 (__m512i __W, __mmask32 __M, __m512i __A, - __m512i __B) +_mm512_mask_min_epi16 (__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) __W, - (__mmask32) __M); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, + (__v32hi)_mm512_min_epi16(__A, __B), + (__v32hi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_min_epu8 (__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) _mm512_setzero_qi(), - (__mmask64) -1); + return (__m512i)__builtin_ia32_pminub512((__v64qi)__A, (__v64qi)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_min_epu8 (__mmask64 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) _mm512_setzero_qi(), - (__mmask64) __M); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, + (__v64qi)_mm512_min_epu8(__A, __B), + (__v64qi)_mm512_setzero_qi()); } static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_min_epu8 (__m512i __W, __mmask64 __M, __m512i __A, - __m512i __B) +_mm512_mask_min_epu8 (__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) __W, - (__mmask64) __M); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, + (__v64qi)_mm512_min_epu8(__A, __B), + (__v64qi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_min_epu16 (__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) _mm512_setzero_hi(), - (__mmask32) -1); + return (__m512i)__builtin_ia32_pminuw512((__v32hi)__A, (__v32hi)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_min_epu16 (__mmask32 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) _mm512_setzero_hi(), - (__mmask32) __M); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, + (__v32hi)_mm512_min_epu16(__A, __B), + (__v32hi)_mm512_setzero_hi()); } static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_min_epu16 (__m512i __W, __mmask32 __M, __m512i __A, - __m512i __B) +_mm512_mask_min_epu16 (__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) __W, - (__mmask32) __M); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, + (__v32hi)_mm512_min_epu16(__A, __B), + (__v32hi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS diff --git a/lib/Headers/avx512fintrin.h b/lib/Headers/avx512fintrin.h index b4cfe09a42..2273f97326 100644 --- a/lib/Headers/avx512fintrin.h +++ b/lib/Headers/avx512fintrin.h @@ -1121,113 +1121,89 @@ static __inline __m512i __DEFAULT_FN_ATTRS _mm512_max_epi32(__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxsd512_mask ((__v16si) __A, - (__v16si) __B, - (__v16si) - _mm512_setzero_si512 (), - (__mmask16) -1); + return (__m512i)__builtin_ia32_pmaxsd512((__v16si)__A, (__v16si)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_max_epi32 (__m512i __W, __mmask16 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxsd512_mask ((__v16si) __A, - (__v16si) __B, - (__v16si) __W, __M); + return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, + (__v16si)_mm512_max_epi32(__A, __B), + (__v16si)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_max_epi32 (__mmask16 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxsd512_mask ((__v16si) __A, - (__v16si) __B, - (__v16si) - _mm512_setzero_si512 (), - __M); + return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, + (__v16si)_mm512_max_epi32(__A, __B), + (__v16si)_mm512_setzero_si512()); } static __inline __m512i __DEFAULT_FN_ATTRS _mm512_max_epu32(__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxud512_mask ((__v16si) __A, - (__v16si) __B, - (__v16si) - _mm512_setzero_si512 (), - (__mmask16) -1); + return (__m512i)__builtin_ia32_pmaxud512((__v16si)__A, (__v16si)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_max_epu32 (__m512i __W, __mmask16 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxud512_mask ((__v16si) __A, - (__v16si) __B, - (__v16si) __W, __M); + return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, + (__v16si)_mm512_max_epu32(__A, __B), + (__v16si)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_max_epu32 (__mmask16 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxud512_mask ((__v16si) __A, - (__v16si) __B, - (__v16si) - _mm512_setzero_si512 (), - __M); + return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, + (__v16si)_mm512_max_epu32(__A, __B), + (__v16si)_mm512_setzero_si512()); } static __inline __m512i __DEFAULT_FN_ATTRS _mm512_max_epi64(__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxsq512_mask ((__v8di) __A, - (__v8di) __B, - (__v8di) - _mm512_setzero_si512 (), - (__mmask8) -1); + return (__m512i)__builtin_ia32_pmaxsq512((__v8di)__A, (__v8di)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_max_epi64 (__m512i __W, __mmask8 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxsq512_mask ((__v8di) __A, - (__v8di) __B, - (__v8di) __W, __M); + return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M, + (__v8di)_mm512_max_epi64(__A, __B), + (__v8di)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_max_epi64 (__mmask8 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxsq512_mask ((__v8di) __A, - (__v8di) __B, - (__v8di) - _mm512_setzero_si512 (), - __M); + return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M, + (__v8di)_mm512_max_epi64(__A, __B), + (__v8di)_mm512_setzero_si512()); } static __inline __m512i __DEFAULT_FN_ATTRS _mm512_max_epu64(__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxuq512_mask ((__v8di) __A, - (__v8di) __B, - (__v8di) - _mm512_setzero_si512 (), - (__mmask8) -1); + return (__m512i)__builtin_ia32_pmaxuq512((__v8di)__A, (__v8di)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_max_epu64 (__m512i __W, __mmask8 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxuq512_mask ((__v8di) __A, - (__v8di) __B, - (__v8di) __W, __M); + return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M, + (__v8di)_mm512_max_epu64(__A, __B), + (__v8di)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_max_epu64 (__mmask8 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pmaxuq512_mask ((__v8di) __A, - (__v8di) __B, - (__v8di) - _mm512_setzero_si512 (), - __M); + return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M, + (__v8di)_mm512_max_epu64(__A, __B), + (__v8di)_mm512_setzero_si512()); } #define _mm512_mask_min_round_pd(W, U, A, B, R) __extension__ ({ \ @@ -1406,113 +1382,89 @@ static __inline __m512i __DEFAULT_FN_ATTRS _mm512_min_epi32(__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminsd512_mask ((__v16si) __A, - (__v16si) __B, - (__v16si) - _mm512_setzero_si512 (), - (__mmask16) -1); + return (__m512i)__builtin_ia32_pminsd512((__v16si)__A, (__v16si)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_min_epi32 (__m512i __W, __mmask16 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminsd512_mask ((__v16si) __A, - (__v16si) __B, - (__v16si) __W, __M); + return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, + (__v16si)_mm512_min_epi32(__A, __B), + (__v16si)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_min_epi32 (__mmask16 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminsd512_mask ((__v16si) __A, - (__v16si) __B, - (__v16si) - _mm512_setzero_si512 (), - __M); + return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, + (__v16si)_mm512_min_epi32(__A, __B), + (__v16si)_mm512_setzero_si512()); } static __inline __m512i __DEFAULT_FN_ATTRS _mm512_min_epu32(__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminud512_mask ((__v16si) __A, - (__v16si) __B, - (__v16si) - _mm512_setzero_si512 (), - (__mmask16) -1); + return (__m512i)__builtin_ia32_pminud512((__v16si)__A, (__v16si)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_min_epu32 (__m512i __W, __mmask16 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminud512_mask ((__v16si) __A, - (__v16si) __B, - (__v16si) __W, __M); + return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, + (__v16si)_mm512_min_epu32(__A, __B), + (__v16si)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_min_epu32 (__mmask16 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminud512_mask ((__v16si) __A, - (__v16si) __B, - (__v16si) - _mm512_setzero_si512 (), - __M); + return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, + (__v16si)_mm512_min_epu32(__A, __B), + (__v16si)_mm512_setzero_si512()); } static __inline __m512i __DEFAULT_FN_ATTRS _mm512_min_epi64(__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminsq512_mask ((__v8di) __A, - (__v8di) __B, - (__v8di) - _mm512_setzero_si512 (), - (__mmask8) -1); + return (__m512i)__builtin_ia32_pminsq512((__v8di)__A, (__v8di)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_min_epi64 (__m512i __W, __mmask8 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminsq512_mask ((__v8di) __A, - (__v8di) __B, - (__v8di) __W, __M); + return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M, + (__v8di)_mm512_min_epi64(__A, __B), + (__v8di)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_min_epi64 (__mmask8 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminsq512_mask ((__v8di) __A, - (__v8di) __B, - (__v8di) - _mm512_setzero_si512 (), - __M); + return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M, + (__v8di)_mm512_min_epi64(__A, __B), + (__v8di)_mm512_setzero_si512()); } static __inline __m512i __DEFAULT_FN_ATTRS _mm512_min_epu64(__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminuq512_mask ((__v8di) __A, - (__v8di) __B, - (__v8di) - _mm512_setzero_si512 (), - (__mmask8) -1); + return (__m512i)__builtin_ia32_pminuq512((__v8di)__A, (__v8di)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_min_epu64 (__m512i __W, __mmask8 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminuq512_mask ((__v8di) __A, - (__v8di) __B, - (__v8di) __W, __M); + return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M, + (__v8di)_mm512_min_epu64(__A, __B), + (__v8di)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_min_epu64 (__mmask8 __M, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_pminuq512_mask ((__v8di) __A, - (__v8di) __B, - (__v8di) - _mm512_setzero_si512 (), - __M); + return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M, + (__v8di)_mm512_min_epu64(__A, __B), + (__v8di)_mm512_setzero_si512()); } static __inline __m512i __DEFAULT_FN_ATTRS @@ -1968,53 +1920,45 @@ _mm512_mask_ceil_pd (__m512d __W, __mmask8 __U, __m512d __A) static __inline __m512i __DEFAULT_FN_ATTRS _mm512_abs_epi64(__m512i __A) { - return (__m512i) __builtin_ia32_pabsq512_mask ((__v8di) __A, - (__v8di) - _mm512_setzero_si512 (), - (__mmask8) -1); + return (__m512i)__builtin_ia32_pabsq512((__v8di)__A); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_abs_epi64 (__m512i __W, __mmask8 __U, __m512i __A) { - return (__m512i) __builtin_ia32_pabsq512_mask ((__v8di) __A, - (__v8di) __W, - (__mmask8) __U); + return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U, + (__v8di)_mm512_abs_epi64(__A), + (__v8di)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_abs_epi64 (__mmask8 __U, __m512i __A) { - return (__m512i) __builtin_ia32_pabsq512_mask ((__v8di) __A, - (__v8di) - _mm512_setzero_si512 (), - (__mmask8) __U); + return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U, + (__v8di)_mm512_abs_epi64(__A), + (__v8di)_mm512_setzero_si512()); } static __inline __m512i __DEFAULT_FN_ATTRS _mm512_abs_epi32(__m512i __A) { - return (__m512i) __builtin_ia32_pabsd512_mask ((__v16si) __A, - (__v16si) - _mm512_setzero_si512 (), - (__mmask16) -1); + return (__m512i)__builtin_ia32_pabsd512((__v16si) __A); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_abs_epi32 (__m512i __W, __mmask16 __U, __m512i __A) { - return (__m512i) __builtin_ia32_pabsd512_mask ((__v16si) __A, - (__v16si) __W, - (__mmask16) __U); + return (__m512i)__builtin_ia32_selectd_512((__mmask8)__U, + (__v16si)_mm512_abs_epi32(__A), + (__v16si)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_abs_epi32 (__mmask16 __U, __m512i __A) { - return (__m512i) __builtin_ia32_pabsd512_mask ((__v16si) __A, - (__v16si) - _mm512_setzero_si512 (), - (__mmask16) __U); + return (__m512i)__builtin_ia32_selectd_512((__mmask8)__U, + (__v16si)_mm512_abs_epi32(__A), + (__v16si)_mm512_setzero_si512()); } static __inline__ __m128 __DEFAULT_FN_ATTRS diff --git a/lib/Headers/avx512vlintrin.h b/lib/Headers/avx512vlintrin.h index 9e03b20a3a..3fc2b80294 100644 --- a/lib/Headers/avx512vlintrin.h +++ b/lib/Headers/avx512vlintrin.h @@ -2844,48 +2844,40 @@ _mm256_maskz_abs_epi32(__mmask8 __U, __m256i __A) { static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_abs_epi64 (__m128i __A) { - return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, - (__v2di) - _mm_setzero_si128 (), - (__mmask8) -1); + return (__m128i)__builtin_ia32_pabsq128((__v2di)__A); } static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_mask_abs_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { - return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, - (__v2di) __W, - (__mmask8) __U); + return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, + (__v2di)_mm_abs_epi64(__A), + (__v2di)__W); } static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_maskz_abs_epi64 (__mmask8 __U, __m128i __A) { - return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, - (__v2di) - _mm_setzero_si128 (), - (__mmask8) __U); + return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, + (__v2di)_mm_abs_epi64(__A), + (__v2di)_mm_setzero_si128()); } static __inline__ __m256i __DEFAULT_FN_ATTRS _mm256_abs_epi64 (__m256i __A) { - return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, - (__v4di) - _mm256_setzero_si256 (), - (__mmask8) -1); + return (__m256i)__builtin_ia32_pabsq256 ((__v4di)__A); } static __inline__ __m256i __DEFAULT_FN_ATTRS _mm256_mask_abs_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { - return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, - (__v4di) __W, - (__mmask8) __U); + return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, + (__v4di)_mm256_abs_epi64(__A), + (__v4di)__W); } static __inline__ __m256i __DEFAULT_FN_ATTRS _mm256_maskz_abs_epi64 (__mmask8 __U, __m256i __A) { - return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, - (__v4di) - _mm256_setzero_si256 (), - (__mmask8) __U); + return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, + (__v4di)_mm256_abs_epi64(__A), + (__v4di)_mm256_setzero_si256()); } static __inline__ __m128i __DEFAULT_FN_ATTRS @@ -2917,55 +2909,41 @@ _mm256_mask_max_epi32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_maskz_max_epi64 (__mmask8 __M, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, - (__v2di) __B, - (__v2di) - _mm_setzero_si128 (), - __M); +_mm_max_epi64 (__m128i __A, __m128i __B) { + return (__m128i)__builtin_ia32_pmaxsq128((__v2di)__A, (__v2di)__B); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_mask_max_epi64 (__m128i __W, __mmask8 __M, __m128i __A, - __m128i __B) { - return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, - (__v2di) __B, - (__v2di) __W, __M); +_mm_maskz_max_epi64 (__mmask8 __M, __m128i __A, __m128i __B) { + return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, + (__v2di)_mm_max_epi64(__A, __B), + (__v2di)_mm_setzero_si128()); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_max_epi64 (__m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, - (__v2di) __B, - (__v2di) - _mm_setzero_si128 (), - (__mmask8) -1); +_mm_mask_max_epi64 (__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) { + return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, + (__v2di)_mm_max_epi64(__A, __B), + (__v2di)__W); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_maskz_max_epi64 (__mmask8 __M, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, - (__v4di) __B, - (__v4di) - _mm256_setzero_si256 (), - __M); +_mm256_max_epi64 (__m256i __A, __m256i __B) { + return (__m256i)__builtin_ia32_pmaxsq256((__v4di)__A, (__v4di)__B); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_mask_max_epi64 (__m256i __W, __mmask8 __M, __m256i __A, - __m256i __B) { - return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, - (__v4di) __B, - (__v4di) __W, __M); +_mm256_maskz_max_epi64 (__mmask8 __M, __m256i __A, __m256i __B) { + return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, + (__v4di)_mm256_max_epi64(__A, __B), + (__v4di)_mm256_setzero_si256()); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_max_epi64 (__m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, - (__v4di) __B, - (__v4di) - _mm256_setzero_si256 (), - (__mmask8) -1); +_mm256_mask_max_epi64 (__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { + return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, + (__v4di)_mm256_max_epi64(__A, __B), + (__v4di)__W); } static __inline__ __m128i __DEFAULT_FN_ATTRS @@ -2997,55 +2975,41 @@ _mm256_mask_max_epu32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_maskz_max_epu64 (__mmask8 __M, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, - (__v2di) __B, - (__v2di) - _mm_setzero_si128 (), - __M); +_mm_max_epu64 (__m128i __A, __m128i __B) { + return (__m128i)__builtin_ia32_pmaxuq128((__v2di)__A, (__v2di)__B); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_max_epu64 (__m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, - (__v2di) __B, - (__v2di) - _mm_setzero_si128 (), - (__mmask8) -1); +_mm_maskz_max_epu64 (__mmask8 __M, __m128i __A, __m128i __B) { + return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, + (__v2di)_mm_max_epu64(__A, __B), + (__v2di)_mm_setzero_si128()); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_mask_max_epu64 (__m128i __W, __mmask8 __M, __m128i __A, - __m128i __B) { - return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, - (__v2di) __B, - (__v2di) __W, __M); +_mm_mask_max_epu64 (__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) { + return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, + (__v2di)_mm_max_epu64(__A, __B), + (__v2di)__W); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_maskz_max_epu64 (__mmask8 __M, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, - (__v4di) __B, - (__v4di) - _mm256_setzero_si256 (), - __M); +_mm256_max_epu64 (__m256i __A, __m256i __B) { + return (__m256i)__builtin_ia32_pmaxuq256((__v4di)__A, (__v4di)__B); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_max_epu64 (__m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, - (__v4di) __B, - (__v4di) - _mm256_setzero_si256 (), - (__mmask8) -1); +_mm256_maskz_max_epu64 (__mmask8 __M, __m256i __A, __m256i __B) { + return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, + (__v4di)_mm256_max_epu64(__A, __B), + (__v4di)_mm256_setzero_si256()); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_mask_max_epu64 (__m256i __W, __mmask8 __M, __m256i __A, - __m256i __B) { - return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, - (__v4di) __B, - (__v4di) __W, __M); +_mm256_mask_max_epu64 (__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { + return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, + (__v4di)_mm256_max_epu64(__A, __B), + (__v4di)__W); } static __inline__ __m128i __DEFAULT_FN_ATTRS @@ -3078,54 +3042,40 @@ _mm256_mask_min_epi32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_min_epi64 (__m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, - (__v2di) __B, - (__v2di) - _mm_setzero_si128 (), - (__mmask8) -1); + return (__m128i)__builtin_ia32_pminsq128((__v2di)__A, (__v2di)__B); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_mask_min_epi64 (__m128i __W, __mmask8 __M, __m128i __A, - __m128i __B) { - return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, - (__v2di) __B, - (__v2di) __W, __M); +_mm_mask_min_epi64 (__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) { + return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, + (__v2di)_mm_min_epi64(__A, __B), + (__v2di)__W); } static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_maskz_min_epi64 (__mmask8 __M, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, - (__v2di) __B, - (__v2di) - _mm_setzero_si128 (), - __M); + return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, + (__v2di)_mm_min_epi64(__A, __B), + (__v2di)_mm_setzero_si128()); } static __inline__ __m256i __DEFAULT_FN_ATTRS _mm256_min_epi64 (__m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, - (__v4di) __B, - (__v4di) - _mm256_setzero_si256 (), - (__mmask8) -1); + return (__m256i)__builtin_ia32_pminsq256((__v4di)__A, (__v4di)__B); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_mask_min_epi64 (__m256i __W, __mmask8 __M, __m256i __A, - __m256i __B) { - return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, - (__v4di) __B, - (__v4di) __W, __M); +_mm256_mask_min_epi64 (__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { + return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, + (__v4di)_mm256_min_epi64(__A, __B), + (__v4di)__W); } static __inline__ __m256i __DEFAULT_FN_ATTRS _mm256_maskz_min_epi64 (__mmask8 __M, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, - (__v4di) __B, - (__v4di) - _mm256_setzero_si256 (), - __M); + return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, + (__v4di)_mm256_min_epi64(__A, __B), + (__v4di)_mm256_setzero_si256()); } static __inline__ __m128i __DEFAULT_FN_ATTRS @@ -3158,54 +3108,40 @@ _mm256_mask_min_epu32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_min_epu64 (__m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, - (__v2di) __B, - (__v2di) - _mm_setzero_si128 (), - (__mmask8) -1); + return (__m128i)__builtin_ia32_pminuq128((__v2di)__A, (__v2di)__B); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_mask_min_epu64 (__m128i __W, __mmask8 __M, __m128i __A, - __m128i __B) { - return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, - (__v2di) __B, - (__v2di) __W, __M); +_mm_mask_min_epu64 (__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) { + return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, + (__v2di)_mm_min_epu64(__A, __B), + (__v2di)__W); } static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_maskz_min_epu64 (__mmask8 __M, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, - (__v2di) __B, - (__v2di) - _mm_setzero_si128 (), - __M); + return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, + (__v2di)_mm_min_epu64(__A, __B), + (__v2di)_mm_setzero_si128()); } static __inline__ __m256i __DEFAULT_FN_ATTRS _mm256_min_epu64 (__m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, - (__v4di) __B, - (__v4di) - _mm256_setzero_si256 (), - (__mmask8) -1); + return (__m256i)__builtin_ia32_pminuq256((__v4di)__A, (__v4di)__B); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_mask_min_epu64 (__m256i __W, __mmask8 __M, __m256i __A, - __m256i __B) { - return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, - (__v4di) __B, - (__v4di) __W, __M); +_mm256_mask_min_epu64 (__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { + return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, + (__v4di)_mm256_min_epu64(__A, __B), + (__v4di)__W); } static __inline__ __m256i __DEFAULT_FN_ATTRS _mm256_maskz_min_epu64 (__mmask8 __M, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, - (__v4di) __B, - (__v4di) - _mm256_setzero_si256 (), - __M); + return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, + (__v4di)_mm256_min_epu64(__A, __B), + (__v4di)_mm256_setzero_si256()); } #define _mm_roundscale_pd(A, imm) __extension__ ({ \ |