10b57cec5SDimitry Andric /*===------------- avx512vbmiintrin.h - VBMI intrinsics ------------------===
20b57cec5SDimitry Andric *
30b57cec5SDimitry Andric *
40b57cec5SDimitry Andric * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
50b57cec5SDimitry Andric * See https://llvm.org/LICENSE.txt for license information.
60b57cec5SDimitry Andric * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
70b57cec5SDimitry Andric *
80b57cec5SDimitry Andric *===-----------------------------------------------------------------------===
90b57cec5SDimitry Andric */
100b57cec5SDimitry Andric #ifndef __IMMINTRIN_H
110b57cec5SDimitry Andric #error "Never use <avx512vbmiintrin.h> directly; include <immintrin.h> instead."
120b57cec5SDimitry Andric #endif
130b57cec5SDimitry Andric
140b57cec5SDimitry Andric #ifndef __VBMIINTRIN_H
150b57cec5SDimitry Andric #define __VBMIINTRIN_H
160b57cec5SDimitry Andric
170b57cec5SDimitry Andric /* Define the default attributes for the functions in this file. */
18*5f757f3fSDimitry Andric #define __DEFAULT_FN_ATTRS \
19*5f757f3fSDimitry Andric __attribute__((__always_inline__, __nodebug__, \
20*5f757f3fSDimitry Andric __target__("avx512vbmi,evex512"), __min_vector_width__(512)))
210b57cec5SDimitry Andric
220b57cec5SDimitry Andric static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_permutex2var_epi8(__m512i __A,__m512i __I,__m512i __B)230b57cec5SDimitry Andric _mm512_permutex2var_epi8(__m512i __A, __m512i __I, __m512i __B)
240b57cec5SDimitry Andric {
250b57cec5SDimitry Andric return (__m512i)__builtin_ia32_vpermi2varqi512((__v64qi)__A, (__v64qi)__I,
260b57cec5SDimitry Andric (__v64qi) __B);
270b57cec5SDimitry Andric }
280b57cec5SDimitry Andric
290b57cec5SDimitry Andric static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_mask_permutex2var_epi8(__m512i __A,__mmask64 __U,__m512i __I,__m512i __B)300b57cec5SDimitry Andric _mm512_mask_permutex2var_epi8(__m512i __A, __mmask64 __U, __m512i __I,
310b57cec5SDimitry Andric __m512i __B)
320b57cec5SDimitry Andric {
330b57cec5SDimitry Andric return (__m512i)__builtin_ia32_selectb_512(__U,
340b57cec5SDimitry Andric (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B),
350b57cec5SDimitry Andric (__v64qi)__A);
360b57cec5SDimitry Andric }
370b57cec5SDimitry Andric
380b57cec5SDimitry Andric static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_mask2_permutex2var_epi8(__m512i __A,__m512i __I,__mmask64 __U,__m512i __B)390b57cec5SDimitry Andric _mm512_mask2_permutex2var_epi8(__m512i __A, __m512i __I, __mmask64 __U,
400b57cec5SDimitry Andric __m512i __B)
410b57cec5SDimitry Andric {
420b57cec5SDimitry Andric return (__m512i)__builtin_ia32_selectb_512(__U,
430b57cec5SDimitry Andric (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B),
440b57cec5SDimitry Andric (__v64qi)__I);
450b57cec5SDimitry Andric }
460b57cec5SDimitry Andric
470b57cec5SDimitry Andric static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_maskz_permutex2var_epi8(__mmask64 __U,__m512i __A,__m512i __I,__m512i __B)480b57cec5SDimitry Andric _mm512_maskz_permutex2var_epi8(__mmask64 __U, __m512i __A, __m512i __I,
490b57cec5SDimitry Andric __m512i __B)
500b57cec5SDimitry Andric {
510b57cec5SDimitry Andric return (__m512i)__builtin_ia32_selectb_512(__U,
520b57cec5SDimitry Andric (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B),
530b57cec5SDimitry Andric (__v64qi)_mm512_setzero_si512());
540b57cec5SDimitry Andric }
550b57cec5SDimitry Andric
560b57cec5SDimitry Andric static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_permutexvar_epi8(__m512i __A,__m512i __B)570b57cec5SDimitry Andric _mm512_permutexvar_epi8 (__m512i __A, __m512i __B)
580b57cec5SDimitry Andric {
590b57cec5SDimitry Andric return (__m512i)__builtin_ia32_permvarqi512((__v64qi) __B, (__v64qi) __A);
600b57cec5SDimitry Andric }
610b57cec5SDimitry Andric
620b57cec5SDimitry Andric static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_maskz_permutexvar_epi8(__mmask64 __M,__m512i __A,__m512i __B)630b57cec5SDimitry Andric _mm512_maskz_permutexvar_epi8 (__mmask64 __M, __m512i __A,
640b57cec5SDimitry Andric __m512i __B)
650b57cec5SDimitry Andric {
660b57cec5SDimitry Andric return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M,
670b57cec5SDimitry Andric (__v64qi)_mm512_permutexvar_epi8(__A, __B),
680b57cec5SDimitry Andric (__v64qi)_mm512_setzero_si512());
690b57cec5SDimitry Andric }
700b57cec5SDimitry Andric
710b57cec5SDimitry Andric static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_mask_permutexvar_epi8(__m512i __W,__mmask64 __M,__m512i __A,__m512i __B)720b57cec5SDimitry Andric _mm512_mask_permutexvar_epi8 (__m512i __W, __mmask64 __M, __m512i __A,
730b57cec5SDimitry Andric __m512i __B)
740b57cec5SDimitry Andric {
750b57cec5SDimitry Andric return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M,
760b57cec5SDimitry Andric (__v64qi)_mm512_permutexvar_epi8(__A, __B),
770b57cec5SDimitry Andric (__v64qi)__W);
780b57cec5SDimitry Andric }
790b57cec5SDimitry Andric
800b57cec5SDimitry Andric static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_multishift_epi64_epi8(__m512i __X,__m512i __Y)810b57cec5SDimitry Andric _mm512_multishift_epi64_epi8(__m512i __X, __m512i __Y)
820b57cec5SDimitry Andric {
830b57cec5SDimitry Andric return (__m512i)__builtin_ia32_vpmultishiftqb512((__v64qi)__X, (__v64qi) __Y);
840b57cec5SDimitry Andric }
850b57cec5SDimitry Andric
860b57cec5SDimitry Andric static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_mask_multishift_epi64_epi8(__m512i __W,__mmask64 __M,__m512i __X,__m512i __Y)870b57cec5SDimitry Andric _mm512_mask_multishift_epi64_epi8(__m512i __W, __mmask64 __M, __m512i __X,
880b57cec5SDimitry Andric __m512i __Y)
890b57cec5SDimitry Andric {
900b57cec5SDimitry Andric return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M,
910b57cec5SDimitry Andric (__v64qi)_mm512_multishift_epi64_epi8(__X, __Y),
920b57cec5SDimitry Andric (__v64qi)__W);
930b57cec5SDimitry Andric }
940b57cec5SDimitry Andric
950b57cec5SDimitry Andric static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_maskz_multishift_epi64_epi8(__mmask64 __M,__m512i __X,__m512i __Y)960b57cec5SDimitry Andric _mm512_maskz_multishift_epi64_epi8(__mmask64 __M, __m512i __X, __m512i __Y)
970b57cec5SDimitry Andric {
980b57cec5SDimitry Andric return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M,
990b57cec5SDimitry Andric (__v64qi)_mm512_multishift_epi64_epi8(__X, __Y),
1000b57cec5SDimitry Andric (__v64qi)_mm512_setzero_si512());
1010b57cec5SDimitry Andric }
1020b57cec5SDimitry Andric
1030b57cec5SDimitry Andric
1040b57cec5SDimitry Andric #undef __DEFAULT_FN_ATTRS
1050b57cec5SDimitry Andric
1060b57cec5SDimitry Andric #endif
107