;
; Динамически подключаемая библиотека avo. Версия 1.0.
; Реализация векторных операций.
; Поддержка комплексных чисел, кватернионов и октонионов.
;
; Copyright © 2026 Малик Разработчик
;
; Это свободная программа: вы можете перераспространять ее и/или изменять
; ее на условиях Меньшей Стандартной общественной лицензии GNU в том виде,
; в каком она была опубликована Фондом свободного программного обеспечения;
; либо версии 3 лицензии, либо (по вашему выбору) любой более поздней версии.
;
; Эта программа распространяется в надежде, что она будет полезной,
; но БЕЗО ВСЯКИХ ГАРАНТИЙ; даже без неявной гарантии ТОВАРНОГО ВИДА
; или ПРИГОДНОСТИ ДЛЯ ОПРЕДЕЛЕННЫХ ЦЕЛЕЙ. Подробнее см. в Меньшей Стандартной
; общественной лицензии GNU.
;
; Вы должны были получить копию Меньшей Стандартной общественной лицензии GNU
; вместе с этой программой. Если это не так, см.
; <https://www.gnu.org/licenses/>.
;
; <fold проверка поддержки AVX-512>
function IsAVX512SupportedByCPU
push rbx
mov eax, $00000001
cpuid
and ecx, $10000001 ; SSE3 [бит 0], AVX [бит 28]
cmp ecx, $10000001
jne @F
mov eax, $00000007
mov ecx, $00000000
cpuid
and ebx, $80030020 ; AVX2 [бит 5], AVX512F [бит 16], AVX512DQ [бит 17], AVX512VL [бит 31]
cmp ebx, $80030020
jne @F
mov eax, $00000001
jmp .exit
@@: xor eax, eax
.exit: pop rbx
ret
function IsAVX512SupportedByOS
push rbx
mov eax, $00000001
cpuid
test ecx, $08000000 ; инструкция XGETBV включена операционной системой?
jz @F
xor ecx, ecx
xgetbv
and eax, $000000e7 ; состояние регистров FPU, SSE, AVX и AVX-512 сохраняется операционной системой при переключении контекстов?
cmp eax, $000000e7
jne @F
mov eax, $00000001
jmp .exit
@@: xor eax, eax
.exit: pop rbx
ret
; </fold>
; <fold побитовое «и»>
function BitAndByte2
movzx eax, word [val1]
movzx iVal2, word [val2]
and eax, iVal2
mov word [res], ax
ret
function BitAndByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpand xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function BitAndByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpand xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function BitAndShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpand xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function BitAndInt8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpand ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function BitAndLong8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vpandq zmm0, zmm0, zmm1
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold побитовое «не»>
function BitNotByte2
movzx eax, word [val]
not eax
mov word [res], ax
ret
function BitNotByte4
vmovd xmm0, [val]
vpxor xmm0, xmm0, [byte8.mask4]
vmovd dword [res], xmm0
ret
function BitNotByte8
vmovq xmm0, [val]
vpxor xmm0, xmm0, [byte8.mask8]
vmovq qword [res], xmm0
ret
function BitNotShort8
vmovdqu xmm0, [val]
vpxor xmm0, xmm0, [short8.mask8]
vmovdqu xword [res], xmm0
ret
function BitNotInt8
vmovdqu ymm0, [val]
vpxor ymm0, ymm0, [int8.m1]
vmovdqu yword [res], ymm0
ret
function BitNotLong8
vmovdqu64 zmm0, [val]
vpxorq zmm0, zmm0, [long8.m1]
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold побитовое «или»>
function BitOrByte2
movzx eax, word [val1]
movzx iVal2, word [val2]
or eax, iVal2
mov word [res], ax
ret
function BitOrByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpor xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function BitOrByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpor xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function BitOrShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpor xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function BitOrInt8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpor ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function BitOrLong8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vporq zmm0, zmm0, zmm1
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold побитовое «исключающее или»>
function BitXorByte2
movzx eax, word [val1]
movzx iVal2, word [val2]
xor eax, iVal2
mov word [res], ax
ret
function BitXorByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpxor xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function BitXorByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpxor xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function BitXorShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpxor xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function BitXorInt8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpxor ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function BitXorLong8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vpxorq zmm0, zmm0, zmm1
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold приведение значения типа double2 к другим типам>
function CastDouble2ToFloat2
vmovdqu xmm0, [val]
vcvtpd2ps xmm0, xmm0
vmovq qword [res], xmm0
ret
function CastDouble2ToByte2
vmovdqu xmm0, [val]
vcvttpd2dq xmm1, xmm0
vmovdqa xmm2, [double8.imax]
vmovdqa xmm3, [int8.m1_0]
vcmppd msk1, xmm2, xmm0, $02
vcmppd msk2, xmm0, xmm0, $07
vmovdqa32 xmm2{k1}{z}, xmm3
vmovdqa32 xmm3{k2}{z}, xmm3
vpaddd xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vpand xmm0, xmm0, [int8.maskb]
vpackusdw xmm0, xmm0, [zero]
vpackuswb xmm0, xmm0, [zero]
vmovd eax, xmm0
mov word [res], ax
ret
function CastDouble2ToShort2
vmovdqu xmm0, [val]
vcvttpd2dq xmm1, xmm0
vmovdqa xmm2, [double8.imax]
vmovdqa xmm3, [int8.m1_0]
vcmppd msk1, xmm2, xmm0, $02
vcmppd msk2, xmm0, xmm0, $07
vmovdqa32 xmm2{k1}{z}, xmm3
vmovdqa32 xmm3{k2}{z}, xmm3
vpaddd xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vpand xmm0, xmm0, [int8.maskw]
vpackusdw xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function CastDouble2ToInt2
vmovdqu xmm0, [val]
vcvttpd2dq xmm1, xmm0
vmovdqa xmm2, [double8.imax]
vmovdqa xmm3, [int8.m1_0]
vcmppd msk1, xmm2, xmm0, $02
vcmppd msk2, xmm0, xmm0, $07
vmovdqa32 xmm2{k1}{z}, xmm3
vmovdqa32 xmm3{k2}{z}, xmm3
vpaddd xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vmovq qword [res], xmm0
ret
function CastDouble2ToLong2
vmovdqu xmm0, [val]
vcvttpd2qq xmm1, xmm0
vmovdqa xmm2, [double8.lmax]
vcmppd xmm2, xmm2, xmm0, $02
vcmppd xmm3, xmm0, xmm0, $07
vpaddq xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold приведение значения типа double4 к другим типам>
function CastDouble4ToFloat4
vmovdqu ymm0, [val]
vcvtpd2ps xmm0, ymm0
vmovdqu xword [res], xmm0
ret
function CastDouble4ToByte4
vmovdqu ymm0, [val]
vcvttpd2dq xmm1, ymm0
vmovdqa ymm2, [double8.imax]
vmovdqa xmm3, [int8.m1]
vcmppd msk1, ymm2, ymm0, $02
vcmppd msk2, ymm0, ymm0, $07
vmovdqa32 xmm2{k1}{z}, xmm3
vmovdqa32 xmm3{k2}{z}, xmm3
vpaddd xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vpand xmm0, xmm0, [int8.maskb]
vpackusdw xmm0, xmm0, [zero]
vpackuswb xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function CastDouble4ToShort4
vmovdqu ymm0, [val]
vcvttpd2dq xmm1, ymm0
vmovdqa ymm2, [double8.imax]
vmovdqa xmm3, [int8.m1]
vcmppd msk1, ymm2, ymm0, $02
vcmppd msk2, ymm0, ymm0, $07
vmovdqa32 xmm2{k1}{z}, xmm3
vmovdqa32 xmm3{k2}{z}, xmm3
vpaddd xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vpand xmm0, xmm0, [int8.maskw]
vpackusdw xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function CastDouble4ToInt4
vmovdqu ymm0, [val]
vcvttpd2dq xmm1, ymm0
vmovdqa ymm2, [double8.imax]
vmovdqa xmm3, [int8.m1]
vcmppd msk1, ymm2, ymm0, $02
vcmppd msk2, ymm0, ymm0, $07
vmovdqa32 xmm2{k1}{z}, xmm3
vmovdqa32 xmm3{k2}{z}, xmm3
vpaddd xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vmovdqu xword [res], xmm0
ret
function CastDouble4ToLong4
vmovdqu ymm0, [val]
vcvttpd2qq ymm1, ymm0
vmovdqa ymm2, [double8.lmax]
vcmppd ymm2, ymm2, ymm0, $02
vcmppd ymm3, ymm0, ymm0, $07
vpaddq ymm0, ymm1, ymm2
vpand ymm0, ymm0, ymm3
vmovdqu yword [res], ymm0
ret
; </fold>
; <fold приведение значения типа double8 к другим типам>
function CastDouble8ToFloat8
vmovdqu64 zmm0, [val]
vcvtpd2ps ymm0, zmm0
vmovdqu yword [res], ymm0
ret
function CastDouble8ToByte8
vmovdqu64 zmm0, [val]
vcvttpd2dq ymm1, zmm0
vmovdqa64 zmm2, [double8.imax]
vmovdqa ymm3, [int8.m1]
vcmppd msk1, zmm2, zmm0, $02
vcmppd msk2, zmm0, zmm0, $07
vmovdqa32 ymm2{k1}{z}, ymm3
vmovdqa32 ymm3{k2}{z}, ymm3
vpaddd ymm0, ymm1, ymm2
vpand ymm0, ymm0, ymm3
vpand ymm0, ymm0, [int8.maskb]
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vpackuswb xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function CastDouble8ToShort8
vmovdqu64 zmm0, [val]
vcvttpd2dq ymm1, zmm0
vmovdqa64 zmm2, [double8.imax]
vmovdqa ymm3, [int8.m1]
vcmppd msk1, zmm2, zmm0, $02
vcmppd msk2, zmm0, zmm0, $07
vmovdqa32 ymm2{k1}{z}, ymm3
vmovdqa32 ymm3{k2}{z}, ymm3
vpaddd ymm0, ymm1, ymm2
vpand ymm0, ymm0, ymm3
vpand ymm0, ymm0, [int8.maskw]
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function CastDouble8ToInt8
vmovdqu64 zmm0, [val]
vcvttpd2dq ymm1, zmm0
vmovdqa64 zmm2, [double8.imax]
vmovdqa ymm3, [int8.m1]
vcmppd msk1, zmm2, zmm0, $02
vcmppd msk2, zmm0, zmm0, $07
vmovdqa32 ymm2{k1}{z}, ymm3
vmovdqa32 ymm3{k2}{z}, ymm3
vpaddd ymm0, ymm1, ymm2
vpand ymm0, ymm0, ymm3
vmovdqu yword [res], ymm0
ret
function CastDouble8ToLong8
vmovdqu64 zmm0, [val]
vcvttpd2qq zmm1, zmm0
vmovdqa64 zmm2, [double8.lmax]
vmovdqa64 zmm3, [long8.m1]
vcmppd msk1, zmm2, zmm0, $02
vcmppd msk2, zmm0, zmm0, $07
vmovdqa64 zmm2{k1}{z}, zmm3
vmovdqa64 zmm3{k2}{z}, zmm3
vpaddq zmm0, zmm1, zmm2
vpandq zmm0, zmm0, zmm3
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold приведение значения типа float2 к другим типам>
function CastFloat2ToDouble2
vmovq xmm0, [val]
vcvtps2pd xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function CastFloat2ToByte2
vmovq xmm0, [val]
vcvttps2dq xmm1, xmm0
vmovdqa xmm2, [float8.imax]
vcmpps xmm2, xmm2, xmm0, $02
vcmpps xmm3, xmm0, xmm0, $07
vpaddd xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vpand xmm0, xmm0, [int8.maskb]
vpackusdw xmm0, xmm0, [zero]
vpackuswb xmm0, xmm0, [zero]
vmovd eax, xmm0
mov word [res], ax
ret
function CastFloat2ToShort2
vmovq xmm0, [val]
vcvttps2dq xmm1, xmm0
vmovdqa xmm2, [float8.imax]
vcmpps xmm2, xmm2, xmm0, $02
vcmpps xmm3, xmm0, xmm0, $07
vpaddd xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vpand xmm0, xmm0, [int8.maskw]
vpackusdw xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function CastFloat2ToInt2
vmovq xmm0, [val]
vcvttps2dq xmm1, xmm0
vmovdqa xmm2, [float8.imax]
vcmpps xmm2, xmm2, xmm0, $02
vcmpps xmm3, xmm0, xmm0, $07
vpaddd xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vmovq qword [res], xmm0
ret
function CastFloat2ToLong2
vmovq xmm0, [val]
vcvtps2pd xmm0, xmm0
vcvttpd2qq xmm1, xmm0
vmovdqa xmm2, [double8.lmax]
vcmppd xmm2, xmm2, xmm0, $02
vcmppd xmm3, xmm0, xmm0, $07
vpaddq xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold приведение значения типа float4 к другим типам>
function CastFloat4ToDouble4
vmovdqu xmm0, [val]
vcvtps2pd ymm0, xmm0
vmovdqu yword [res], ymm0
ret
function CastFloat4ToByte4
vmovdqu xmm0, [val]
vcvttps2dq xmm1, xmm0
vmovdqa xmm2, [float8.imax]
vcmpps xmm2, xmm2, xmm0, $02
vcmpps xmm3, xmm0, xmm0, $07
vpaddd xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vpand xmm0, xmm0, [int8.maskb]
vpackusdw xmm0, xmm0, [zero]
vpackuswb xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function CastFloat4ToShort4
vmovdqu xmm0, [val]
vcvttps2dq xmm1, xmm0
vmovdqa xmm2, [float8.imax]
vcmpps xmm2, xmm2, xmm0, $02
vcmpps xmm3, xmm0, xmm0, $07
vpaddd xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vpand xmm0, xmm0, [int8.maskw]
vpackusdw xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function CastFloat4ToInt4
vmovdqu xmm0, [val]
vcvttps2dq xmm1, xmm0
vmovdqa xmm2, [float8.imax]
vcmpps xmm2, xmm2, xmm0, $02
vcmpps xmm3, xmm0, xmm0, $07
vpaddd xmm0, xmm1, xmm2
vpand xmm0, xmm0, xmm3
vmovdqu xword [res], xmm0
ret
function CastFloat4ToLong4
vmovdqu xmm0, [val]
vcvtps2pd ymm0, xmm0
vcvttpd2qq ymm1, ymm0
vmovdqa ymm2, [double8.lmax]
vcmppd ymm2, ymm2, ymm0, $02
vcmppd ymm3, ymm0, ymm0, $07
vpaddq ymm0, ymm1, ymm2
vpand ymm0, ymm0, ymm3
vmovdqu yword [res], ymm0
ret
; </fold>
; <fold приведение значения типа float8 к другим типам>
function CastFloat8ToDouble8
vmovdqu ymm0, [val]
vcvtps2pd zmm0, ymm0
vmovdqu64 zword [res], zmm0
ret
function CastFloat8ToByte8
vmovdqu ymm0, [val]
vcvttps2dq ymm1, ymm0
vmovdqa ymm2, [float8.imax]
vcmpps ymm2, ymm2, ymm0, $02
vcmpps ymm3, ymm0, ymm0, $07
vpaddd ymm0, ymm1, ymm2
vpand ymm0, ymm0, ymm3
vpand ymm0, ymm0, [int8.maskb]
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vpackuswb xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function CastFloat8ToShort8
vmovdqu ymm0, [val]
vcvttps2dq ymm1, ymm0
vmovdqa ymm2, [float8.imax]
vcmpps ymm2, ymm2, ymm0, $02
vcmpps ymm3, ymm0, ymm0, $07
vpaddd ymm0, ymm1, ymm2
vpand ymm0, ymm0, ymm3
vpand ymm0, ymm0, [int8.maskw]
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function CastFloat8ToInt8
vmovdqu ymm0, [val]
vcvttps2dq ymm1, ymm0
vmovdqa ymm2, [float8.imax]
vcmpps ymm2, ymm2, ymm0, $02
vcmpps ymm3, ymm0, ymm0, $07
vpaddd ymm0, ymm1, ymm2
vpand ymm0, ymm0, ymm3
vmovdqu yword [res], ymm0
ret
function CastFloat8ToLong8
vmovdqu ymm0, [val]
vcvtps2pd zmm0, ymm0
vcvttpd2qq zmm1, zmm0
vmovdqa64 zmm2, [double8.lmax]
vmovdqa64 zmm3, [long8.m1]
vcmppd msk1, zmm2, zmm0, $02
vcmppd msk2, zmm0, zmm0, $07
vmovdqa64 zmm2{k1}{z}, zmm3
vmovdqa64 zmm3{k2}{z}, zmm3
vpaddq zmm0, zmm1, zmm2
vpandq zmm0, zmm0, zmm3
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold приведение значения типа byte2 к другим типам>
function CastByte2ToDouble2
movzx iVal, word [val]
vmovd xmm0, iVal
vpmovsxbd xmm0, xmm0
vcvtdq2pd xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function CastByte2ToFloat2
movzx iVal, word [val]
vmovd xmm0, iVal
vpmovsxbd xmm0, xmm0
vcvtdq2ps xmm0, xmm0
vmovq qword [res], xmm0
ret
function CastByte2ToShort2
movzx iVal, word [val]
vmovd xmm0, iVal
vpmovsxbw xmm0, xmm0
vmovd dword [res], xmm0
ret
function CastByte2ToInt2
movzx iVal, word [val]
vmovd xmm0, iVal
vpmovsxbd xmm0, xmm0
vmovq qword [res], xmm0
ret
function CastByte2ToLong2
movzx iVal, word [val]
vmovd xmm0, iVal
vpmovsxbq xmm0, xmm0
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold приведение значения типа byte4 к другим типам>
function CastByte4ToDouble4
vmovd xmm0, [val]
vpmovsxbd xmm0, xmm0
vcvtdq2pd ymm0, xmm0
vmovdqu yword [res], ymm0
ret
function CastByte4ToFloat4
vmovd xmm0, [val]
vpmovsxbd xmm0, xmm0
vcvtdq2ps xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function CastByte4ToShort4
vmovd xmm0, [val]
vpmovsxbw xmm0, xmm0
vmovq qword [res], xmm0
ret
function CastByte4ToInt4
vmovd xmm0, [val]
vpmovsxbd xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function CastByte4ToLong4
vmovd xmm0, [val]
vpmovsxbq ymm0, xmm0
vmovdqu yword [res], ymm0
ret
; </fold>
; <fold приведение значения типа byte8 к другим типам>
function CastByte8ToDouble8
vmovq xmm0, [val]
vpmovsxbd ymm0, xmm0
vcvtdq2pd zmm0, ymm0
vmovdqu64 zword [res], zmm0
ret
function CastByte8ToFloat8
vmovq xmm0, [val]
vpmovsxbd ymm0, xmm0
vcvtdq2ps ymm0, ymm0
vmovdqu yword [res], ymm0
ret
function CastByte8ToShort8
vmovq xmm0, [val]
vpmovsxbw xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function CastByte8ToInt8
vmovq xmm0, [val]
vpmovsxbd ymm0, xmm0
vmovdqu yword [res], ymm0
ret
function CastByte8ToLong8
vmovq xmm0, [val]
vpmovsxbq zmm0, xmm0
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold приведение значения типа short2 к другим типам>
function CastShort2ToDouble2
vmovd xmm0, [val]
vpmovsxwd xmm0, xmm0
vcvtdq2pd xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function CastShort2ToFloat2
vmovd xmm0, [val]
vpmovsxwd xmm0, xmm0
vcvtdq2ps xmm0, xmm0
vmovq qword [res], xmm0
ret
function CastShort2ToByte2
vmovd xmm0, [val]
vpand xmm0, xmm0, [short8.$00ff]
vpackuswb xmm0, xmm0, [zero]
vmovd eax, xmm0
mov word [res], ax
ret
function CastShort2ToInt2
vmovd xmm0, [val]
vpmovsxwd xmm0, xmm0
vmovq qword [res], xmm0
ret
function CastShort2ToLong2
vmovd xmm0, [val]
vpmovsxwq xmm0, xmm0
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold приведение значения типа short4 к другим типам>
function CastShort4ToDouble4
vmovq xmm0, [val]
vpmovsxwd xmm0, xmm0
vcvtdq2pd ymm0, xmm0
vmovdqu yword [res], ymm0
ret
function CastShort4ToFloat4
vmovq xmm0, [val]
vpmovsxwd xmm0, xmm0
vcvtdq2ps xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function CastShort4ToByte4
vmovq xmm0, [val]
vpand xmm0, xmm0, [short8.$00ff]
vpackuswb xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function CastShort4ToInt4
vmovq xmm0, [val]
vpmovsxwd xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function CastShort4ToLong4
vmovq xmm0, [val]
vpmovsxwq ymm0, xmm0
vmovdqu yword [res], ymm0
ret
; </fold>
; <fold приведение значения типа short8 к другим типам>
function CastShort8ToDouble8
vmovdqu xmm0, [val]
vpmovsxwd ymm0, xmm0
vcvtdq2pd zmm0, ymm0
vmovdqu64 zword [res], zmm0
ret
function CastShort8ToFloat8
vmovdqu xmm0, [val]
vpmovsxwd ymm0, xmm0
vcvtdq2ps ymm0, ymm0
vmovdqu yword [res], ymm0
ret
function CastShort8ToByte8
vmovdqu xmm0, [val]
vpand xmm0, xmm0, [short8.$00ff]
vpackuswb xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function CastShort8ToInt8
vmovdqu xmm0, [val]
vpmovsxwd ymm0, xmm0
vmovdqu yword [res], ymm0
ret
function CastShort8ToLong8
vmovdqu xmm0, [val]
vpmovsxwq zmm0, xmm0
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold приведение значения типа int2 к другим типам>
function CastInt2ToDouble2
vmovq xmm0, [val]
vcvtdq2pd xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function CastInt2ToFloat2
vmovq xmm0, [val]
vcvtdq2ps xmm0, xmm0
vmovq qword [res], xmm0
ret
function CastInt2ToByte2
vmovq xmm0, [val]
vpand xmm0, xmm0, [int8.maskb]
vpackusdw xmm0, xmm0, [zero]
vpackuswb xmm0, xmm0, [zero]
vmovd eax, xmm0
mov word [res], ax
ret
function CastInt2ToShort2
vmovq xmm0, [val]
vpand xmm0, xmm0, [int8.maskw]
vpackusdw xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function CastInt2ToLong2
vmovq xmm0, [val]
vpmovsxdq xmm0, xmm0
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold приведение значения типа int4 к другим типам>
function CastInt4ToDouble4
vmovdqu xmm0, [val]
vcvtdq2pd ymm0, xmm0
vmovdqu yword [res], ymm0
ret
function CastInt4ToFloat4
vmovdqu xmm0, [val]
vcvtdq2ps xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function CastInt4ToByte4
vmovdqu xmm0, [val]
vpand xmm0, xmm0, [int8.maskb]
vpackusdw xmm0, xmm0, [zero]
vpackuswb xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function CastInt4ToShort4
vmovdqu xmm0, [val]
vpand xmm0, xmm0, [int8.maskw]
vpackusdw xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function CastInt4ToLong4
vmovdqu xmm0, [val]
vpmovsxdq ymm0, xmm0
vmovdqu yword [res], ymm0
ret
; </fold>
; <fold приведение значения типа int8 к другим типам>
function CastInt8ToDouble8
vmovdqu ymm0, [val]
vcvtdq2pd zmm0, ymm0
vmovdqu64 zword [res], zmm0
ret
function CastInt8ToFloat8
vmovdqu ymm0, [val]
vcvtdq2ps ymm0, ymm0
vmovdqu yword [res], ymm0
ret
function CastInt8ToByte8
vmovdqu ymm0, [val]
vpand ymm0, ymm0, [int8.maskb]
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vpackuswb xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function CastInt8ToShort8
vmovdqu ymm0, [val]
vpand ymm0, ymm0, [int8.maskw]
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function CastInt8ToLong8
vmovdqu ymm0, [val]
vpmovsxdq zmm0, ymm0
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold приведение значения типа long2 к другим типам>
function CastLong2ToDouble2
vmovdqu xmm0, [val]
vcvtqq2pd xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function CastLong2ToFloat2
vmovdqu xmm0, [val]
vcvtqq2ps xmm0, xmm0
vmovq qword [res], xmm0
ret
function CastLong2ToByte2
vmovdqu xmm0, [val]
vpand xmm0, xmm0, [long8.maskb]
vshufps xmm0, xmm0, xmm0, $d8
vpackusdw xmm0, xmm0, [zero]
vpackuswb xmm0, xmm0, [zero]
vmovd eax, xmm0
mov word [res], ax
ret
function CastLong2ToShort2
vmovdqu xmm0, [val]
vpand xmm0, xmm0, [long8.maskw]
vshufps xmm0, xmm0, xmm0, $d8
vpackusdw xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function CastLong2ToInt2
vmovdqu xmm0, [val]
vpand xmm0, xmm0, [long8.maskd]
vshufps xmm0, xmm0, xmm0, $d8
vmovq qword [res], xmm0
ret
; </fold>
; <fold приведение значения типа long4 к другим типам>
function CastLong4ToDouble4
vmovdqu ymm0, [val]
vcvtqq2pd ymm0, ymm0
vmovdqu yword [res], ymm0
ret
function CastLong4ToFloat4
vmovdqu ymm0, [val]
vcvtqq2ps xmm0, ymm0
vmovdqu xword [res], xmm0
ret
function CastLong4ToByte4
vmovdqu ymm0, [val]
vpand ymm0, ymm0, [long8.maskb]
vshufps ymm0, ymm0, ymm0, $d8
vextracti128 xmm1, ymm0, $01
vpslldq xmm1, xmm1, $08
vpor xmm0, xmm0, xmm1
vpackusdw xmm0, xmm0, [zero]
vpackuswb xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function CastLong4ToShort4
vmovdqu ymm0, [val]
vpand ymm0, ymm0, [long8.maskw]
vshufps ymm0, ymm0, ymm0, $d8
vextracti128 xmm1, ymm0, $01
vpslldq xmm1, xmm1, $08
vpor xmm0, xmm0, xmm1
vpackusdw xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function CastLong4ToInt4
vmovdqu ymm0, [val]
vpand ymm0, ymm0, [long8.maskd]
vshufps ymm0, ymm0, ymm0, $d8
vextracti128 xmm1, ymm0, $01
vpslldq xmm1, xmm1, $08
vpor xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold приведение значения типа long8 к другим типам>
function CastLong8ToDouble8
vmovdqu64 zmm0, [val]
vcvtqq2pd zmm0, zmm0
vmovdqu64 zword [res], zmm0
ret
function CastLong8ToFloat8
vmovdqu64 zmm0, [val]
vcvtqq2ps ymm0, zmm0
vmovdqu yword [res], ymm0
ret
function CastLong8ToByte8
vmovdqu64 zmm0, [val]
vpandq zmm0, zmm0, [long8.maskb]
vshufps zmm0, zmm0, zmm0, $d8
vshufi64x2 zmm0, zmm0, zmm0, $d8
vextracti64x4 ymm1, zmm0, $01
vpslldq ymm1, ymm1, $08
vpor ymm0, ymm0, ymm1
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vpackuswb xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function CastLong8ToShort8
vmovdqu64 zmm0, [val]
vpandq zmm0, zmm0, [long8.maskw]
vshufps zmm0, zmm0, zmm0, $d8
vshufi64x2 zmm0, zmm0, zmm0, $d8
vextracti64x4 ymm1, zmm0, $01
vpslldq ymm1, ymm1, $08
vpor ymm0, ymm0, ymm1
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function CastLong8ToInt8
vmovdqu64 zmm0, [val]
vpandq zmm0, zmm0, [long8.maskd]
vshufps zmm0, zmm0, zmm0, $d8
vshufi64x2 zmm0, zmm0, zmm0, $d8
vextracti64x4 ymm1, zmm0, $01
vpslldq ymm1, ymm1, $08
vpor ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
; </fold>
; <fold скалярное сложение>
function ScalarAddDouble2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vaddpd xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarAddDouble4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vaddpd ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function ScalarAddDouble8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vaddpd zmm0, zmm0, zmm1
vmovdqu64 zword [res], zmm0
ret
function ScalarAddFloat2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vaddps xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function ScalarAddFloat4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vaddps xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarAddFloat8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vaddps ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function ScalarAddInt2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpaddd xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function ScalarAddInt4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpaddd xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarAddInt8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpaddd ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function ScalarAddLong2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpaddq xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarAddLong4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpaddq ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function ScalarAddLong8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vpaddq zmm0, zmm0, zmm1
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold скалярное сопряжение>
function ScalarConjDouble2
vmovdqu xmm0, [val]
vmulpd xmm0, xmm0, [smuld2.c0]
vmovdqu xword [res], xmm0
ret
function ScalarConjDouble4
vmovdqu ymm0, [val]
vmulpd ymm0, ymm0, [smuld4.q0]
vmovdqu yword [res], ymm0
ret
function ScalarConjDouble8
vmovdqu64 zmm0, [val]
vmulpd zmm0, zmm0, [smuld8.o0]
vmovdqu64 zword [res], zmm0
ret
function ScalarConjFloat2
vmovq xmm0, [val]
vmulps xmm0, xmm0, [smulf2.c0]
vmovq qword [res], xmm0
ret
function ScalarConjFloat4
vmovdqu xmm0, [val]
vmulps xmm0, xmm0, [smulf4.q0]
vmovdqu xword [res], xmm0
ret
function ScalarConjFloat8
vmovdqu ymm0, [val]
vmulps ymm0, ymm0, [smulf8.o0]
vmovdqu yword [res], ymm0
ret
function ScalarConjInt2
vmovq xmm0, [val]
vpmulld xmm0, xmm0, [smuli2.c0]
vmovq qword [res], xmm0
ret
function ScalarConjInt4
vmovdqu xmm0, [val]
vpmulld xmm0, xmm0, [smuli4.q0]
vmovdqu xword [res], xmm0
ret
function ScalarConjInt8
vmovdqu ymm0, [val]
vpmulld ymm0, ymm0, [smuli8.o0]
vmovdqu yword [res], ymm0
ret
function ScalarConjLong2
vmovdqu xmm0, [val]
vpmullq xmm0, xmm0, [smull2.c0]
vmovdqu xword [res], xmm0
ret
function ScalarConjLong4
vmovdqu ymm0, [val]
vpmullq ymm0, ymm0, [smull4.q0]
vmovdqu yword [res], ymm0
ret
function ScalarConjLong8
vmovdqu64 zmm0, [val]
vpmullq zmm0, zmm0, [smull8.o0]
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold скалярное деление>
function ScalarDivDouble2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vmulpd xmm1, xmm1, [smuld2.c0]
vmulpd xmm2, xmm1, xmm1
vhaddpd xmm2, xmm2, xmm2
vdivpd xmm1, xmm1, xmm2
vshufpd xmm2, xmm1, xmm1, $02
vshufpd xmm3, xmm2, xmm2, $01
vmulpd xmm2, xmm2, [smuld2.c0]
vmulpd xmm1, xmm0, xmm3
vmulpd xmm0, xmm0, xmm2
vhaddpd xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarDivDouble4
enterq
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vmulpd ymm1, ymm1, [smuld4.q0]
vmulpd ymm2, ymm1, ymm1
vextractf128 xmm3, ymm2, $01
vhaddpd xmm2, xmm2, xmm3
vhaddpd xmm2, xmm2, xmm2
vinsertf128 ymm2, ymm2, xmm2, $01
vdivpd ymm1, ymm1, ymm2
vshuff64x2 ymm4, ymm1, ymm1, $02
vshuff64x2 ymm6, ymm4, ymm4, $01
vshufpd ymm5, ymm4, ymm4, $05
vshufpd ymm7, ymm6, ymm6, $05
vmulpd ymm4, ymm4, [smuld4.q0]
vmulpd ymm5, ymm5, [smuld4.q1]
vmulpd ymm6, ymm6, [smuld4.q2]
vmulpd ymm7, ymm7, [smuld4.q3]
vmulpd ymm3, ymm0, ymm7
vmulpd ymm2, ymm0, ymm6
vmulpd ymm1, ymm0, ymm5
vmulpd ymm0, ymm0, ymm4
vextractf128 xmm4, ymm0, $01
vextractf128 xmm5, ymm1, $01
vextractf128 xmm6, ymm2, $01
vextractf128 xmm7, ymm3, $01
vinsertf128 ymm0, ymm0, xmm2, $01
vinsertf128 ymm2, ymm1, xmm3, $01
vinsertf128 ymm1, ymm4, xmm6, $01
vinsertf128 ymm3, ymm5, xmm7, $01
vaddpd ymm0, ymm0, ymm1
vaddpd ymm1, ymm2, ymm3
vshufpd ymm2, ymm0, ymm1, $00
vshufpd ymm3, ymm0, ymm1, $0f
vaddpd ymm0, ymm2, ymm3
vmovdqu yword [res], ymm0
leaveq
ret
function ScalarDivDouble8
entero
vmovdqu64 zmm00, [val1]
vmovdqu64 zmm01, [val2]
vmulpd zmm01, zmm01, [smuld8.o0]
vmulpd zmm02, zmm01, zmm01, {rn-sae}
vextractf64x2 xmm03, zmm02, $01
vextractf64x2 xmm04, zmm02, $02
vextractf64x2 xmm05, zmm02, $03
vhaddpd xmm02, xmm02, xmm03
vhaddpd xmm03, xmm04, xmm05
vhaddpd xmm02, xmm02, xmm03
vhaddpd xmm02, xmm02, xmm02
vbroadcastsd zmm02, xmm02
vdivpd zmm01, zmm01, zmm02
vshuff64x2 zmm08, zmm01, zmm01, $e4
vshuff64x2 zmm10, zmm08, zmm08, $b1
vshuff64x2 zmm12, zmm08, zmm08, $4e
vshuff64x2 zmm14, zmm08, zmm08, $1b
vshufpd zmm09, zmm08, zmm08, $55
vshufpd zmm11, zmm10, zmm10, $55
vshufpd zmm13, zmm12, zmm12, $55
vshufpd zmm15, zmm14, zmm14, $55
vmulpd zmm08, zmm08, [smuld8.o0]
vmulpd zmm09, zmm09, [smuld8.o1]
vmulpd zmm10, zmm10, [smuld8.o2]
vmulpd zmm11, zmm11, [smuld8.o3]
vmulpd zmm12, zmm12, [smuld8.o4]
vmulpd zmm13, zmm13, [smuld8.o5]
vmulpd zmm14, zmm14, [smuld8.o6]
vmulpd zmm15, zmm15, [smuld8.o7]
vmulpd zmm07, zmm00, zmm15, {rn-sae}
vmulpd zmm06, zmm00, zmm14, {rn-sae}
vmulpd zmm05, zmm00, zmm13, {rn-sae}
vmulpd zmm04, zmm00, zmm12, {rn-sae}
vmulpd zmm03, zmm00, zmm11, {rn-sae}
vmulpd zmm02, zmm00, zmm10, {rn-sae}
vmulpd zmm01, zmm00, zmm09, {rn-sae}
vmulpd zmm00, zmm00, zmm08, {rn-sae}
vextractf64x4 ymm08, zmm00, $01
vextractf64x4 ymm09, zmm01, $01
vextractf64x4 ymm10, zmm02, $01
vextractf64x4 ymm11, zmm03, $01
vextractf64x4 ymm12, zmm04, $01
vextractf64x4 ymm13, zmm05, $01
vextractf64x4 ymm14, zmm06, $01
vextractf64x4 ymm15, zmm07, $01
vinsertf64x4 zmm00, zmm00, ymm02, $01
vinsertf64x4 zmm02, zmm04, ymm06, $01
vinsertf64x4 zmm04, zmm01, ymm03, $01
vinsertf64x4 zmm06, zmm05, ymm07, $01
vinsertf64x4 zmm01, zmm08, ymm10, $01
vinsertf64x4 zmm03, zmm12, ymm14, $01
vinsertf64x4 zmm05, zmm09, ymm11, $01
vinsertf64x4 zmm07, zmm13, ymm15, $01
vaddpd zmm00, zmm00, zmm01
vaddpd zmm01, zmm02, zmm03
vaddpd zmm02, zmm04, zmm05
vaddpd zmm03, zmm06, zmm07
vshuff64x2 zmm04, zmm00, zmm01, $88
vshuff64x2 zmm05, zmm00, zmm01, $dd
vshuff64x2 zmm06, zmm02, zmm03, $88
vshuff64x2 zmm07, zmm02, zmm03, $dd
vaddpd zmm00, zmm04, zmm05
vaddpd zmm01, zmm06, zmm07
vshufpd zmm02, zmm00, zmm01, $00
vshufpd zmm03, zmm00, zmm01, $ff
vaddpd zmm00, zmm02, zmm03
vmovdqu64 zword [res], zmm00
leaveo
ret
function ScalarDivFloat2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vmulps xmm1, xmm1, [smulf2.c0]
vmulps xmm2, xmm1, xmm1
vhaddps xmm2, xmm2, xmm2
vshufps xmm2, xmm2, xmm2, $a0
vdivps xmm1, xmm1, xmm2
vshufps xmm2, xmm1, xmm1, $e4
vshufps xmm3, xmm2, xmm2, $e1
vmulps xmm2, xmm2, [smulf2.c0]
vmulps xmm1, xmm0, xmm3
vmulps xmm0, xmm0, xmm2
vhaddps xmm0, xmm0, xmm1
vshufps xmm0, xmm0, xmm0, $d8
vmovq qword [res], xmm0
ret
function ScalarDivFloat4
enterq
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vmulps xmm1, xmm1, [smulf4.q0]
vmulps xmm2, xmm1, xmm1
vhaddps xmm2, xmm2, xmm2
vhaddps xmm2, xmm2, xmm2
vdivps xmm1, xmm1, xmm2
vshufps xmm4, xmm1, xmm1, $e4
vshufps xmm5, xmm4, xmm4, $b1
vshufps xmm6, xmm4, xmm4, $4e
vshufps xmm7, xmm4, xmm4, $1b
vmulps xmm4, xmm4, [smulf4.q0]
vmulps xmm5, xmm5, [smulf4.q1]
vmulps xmm6, xmm6, [smulf4.q2]
vmulps xmm7, xmm7, [smulf4.q3]
vmulps xmm3, xmm0, xmm7
vmulps xmm2, xmm0, xmm6
vmulps xmm1, xmm0, xmm5
vmulps xmm0, xmm0, xmm4
vhaddps xmm0, xmm0, xmm1
vhaddps xmm1, xmm2, xmm3
vhaddps xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
leaveq
ret
function ScalarDivFloat8
entero
vmovdqu ymm00, [val1]
vmovdqu ymm01, [val2]
vmulps ymm01, ymm01, [smulf8.o0]
vmulps ymm02, ymm01, ymm01
vextractf128 xmm03, ymm02, $01
vhaddps xmm02, xmm02, xmm03
vhaddps xmm02, xmm02, xmm02
vhaddps xmm02, xmm02, xmm02
vinsertf128 ymm02, ymm02, xmm02, $01
vdivps ymm01, ymm01, ymm02
vmovdqa ymm08, ymm01
vextractf128 xmm12, ymm08, $01
vinsertf128 ymm12, ymm12, xmm08, $01
vshufps ymm09, ymm08, ymm08, $b1
vshufps ymm10, ymm08, ymm08, $4e
vshufps ymm11, ymm08, ymm08, $1b
vshufps ymm13, ymm12, ymm12, $b1
vshufps ymm14, ymm12, ymm12, $4e
vshufps ymm15, ymm12, ymm12, $1b
vmulps ymm08, ymm08, [smulf8.o0]
vmulps ymm09, ymm09, [smulf8.o1]
vmulps ymm10, ymm10, [smulf8.o2]
vmulps ymm11, ymm11, [smulf8.o3]
vmulps ymm12, ymm12, [smulf8.o4]
vmulps ymm13, ymm13, [smulf8.o5]
vmulps ymm14, ymm14, [smulf8.o6]
vmulps ymm15, ymm15, [smulf8.o7]
vmulps ymm07, ymm00, ymm15
vmulps ymm06, ymm00, ymm14
vmulps ymm05, ymm00, ymm13
vmulps ymm04, ymm00, ymm12
vmulps ymm03, ymm00, ymm11
vmulps ymm02, ymm00, ymm10
vmulps ymm01, ymm00, ymm09
vmulps ymm00, ymm00, ymm08
vhaddps ymm00, ymm00, ymm01
vhaddps ymm01, ymm02, ymm03
vhaddps ymm02, ymm04, ymm05
vhaddps ymm03, ymm06, ymm07
vhaddps ymm00, ymm00, ymm01
vhaddps ymm01, ymm02, ymm03
vextractf128 xmm02, ymm00, $01
vextractf128 xmm03, ymm01, $01
vaddps xmm00, xmm00, xmm02
vaddps xmm01, xmm01, xmm03
vinsertf128 ymm00, ymm00, xmm01, $01
vmovdqu yword [res], ymm00
leaveo
ret
; </fold>
; <fold скалярное обращение>
function ScalarInvDouble2
vmovdqu xmm0, [val]
vmulpd xmm0, xmm0, [smuld2.c0]
vmulpd xmm1, xmm0, xmm0
vhaddpd xmm1, xmm1, xmm1
vdivpd xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarInvDouble4
vmovdqu ymm0, [val]
vmulpd ymm0, ymm0, [smuld4.q0]
vmulpd ymm1, ymm0, ymm0
vextractf128 xmm2, ymm1, $01
vhaddpd xmm1, xmm1, xmm2
vhaddpd xmm1, xmm1, xmm1
vinsertf128 ymm1, ymm1, xmm1, $01
vdivpd ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function ScalarInvDouble8
vmovdqu64 zmm0, [val]
vmulpd zmm0, zmm0, [smuld8.o0]
vmulpd zmm1, zmm0, zmm0, {rn-sae}
vextractf64x2 xmm2, zmm1, $01
vextractf64x2 xmm3, zmm1, $02
vextractf64x2 xmm4, zmm1, $03
vhaddpd xmm1, xmm1, xmm2
vhaddpd xmm2, xmm3, xmm4
vhaddpd xmm1, xmm1, xmm2
vhaddpd xmm1, xmm1, xmm1
vbroadcastsd zmm1, xmm1
vdivpd zmm0, zmm0, zmm1
vmovdqu64 zword [res], zmm0
ret
function ScalarInvFloat2
vmovq xmm0, [val]
vmulps xmm0, xmm0, [smulf2.c0]
vmulps xmm1, xmm0, xmm0
vhaddps xmm1, xmm1, xmm1
vshufps xmm1, xmm1, xmm1, $a0
vdivps xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function ScalarInvFloat4
vmovdqu xmm0, [val]
vmulps xmm0, xmm0, [smulf4.q0]
vmulps xmm1, xmm0, xmm0
vhaddps xmm1, xmm1, xmm1
vhaddps xmm1, xmm1, xmm1
vdivps xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarInvFloat8
vmovdqu ymm0, [val]
vmulps ymm0, ymm0, [smulf8.o0]
vmulps ymm1, ymm0, ymm0
vextractf128 xmm2, ymm1, $01
vhaddps xmm1, xmm1, xmm2
vhaddps xmm1, xmm1, xmm1
vhaddps xmm1, xmm1, xmm1
vinsertf128 ymm1, ymm1, xmm1, $01
vdivps ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
; </fold>
; <fold скалярное равно>
function ScalarEqualsDouble2
vmovdqu xmm0, [res]
vmovdqu xmm1, [val]
vcmppd xmm0, xmm0, xmm1, $00
vmovmskpd eax, xmm0
and eax, $03
cmp eax, $03
sete al
ret
function ScalarEqualsDouble4
vmovdqu ymm0, [res]
vmovdqu ymm1, [val]
vcmppd ymm0, ymm0, ymm1, $00
vmovmskpd eax, ymm0
and eax, $0f
cmp eax, $0f
sete al
ret
function ScalarEqualsDouble8
vmovdqu64 zmm0, [res]
vmovdqu64 zmm1, [val]
vcmppd msk1, zmm0, zmm1, $00
kmovw eax, msk1
and eax, $ff
cmp eax, $ff
sete al
ret
function ScalarEqualsFloat2
vmovq xmm0, [res]
vmovq xmm1, [val]
vcmpps xmm0, xmm0, xmm1, $00
vmovmskps eax, xmm0
and eax, $03
cmp eax, $03
sete al
ret
function ScalarEqualsFloat4
vmovdqu xmm0, [res]
vmovdqu xmm1, [val]
vcmpps xmm0, xmm0, xmm1, $00
vmovmskps eax, xmm0
and eax, $0f
cmp eax, $0f
sete al
ret
function ScalarEqualsFloat8
vmovdqu ymm0, [res]
vmovdqu ymm1, [val]
vcmpps ymm0, ymm0, ymm1, $00
vmovmskps eax, ymm0
and eax, $ff
cmp eax, $ff
sete al
ret
function ScalarEqualsInt2
vmovq xmm0, [res]
vmovq xmm1, [val]
vpcmpeqd xmm0, xmm0, xmm1
vmovmskps eax, xmm0
and eax, $03
cmp eax, $03
sete al
ret
function ScalarEqualsInt4
vmovdqu xmm0, [res]
vmovdqu xmm1, [val]
vpcmpeqd xmm0, xmm0, xmm1
vmovmskps eax, xmm0
and eax, $0f
cmp eax, $0f
sete al
ret
function ScalarEqualsInt8
vmovdqu ymm0, [res]
vmovdqu ymm1, [val]
vpcmpeqd ymm0, ymm0, ymm1
vmovmskps eax, ymm0
and eax, $ff
cmp eax, $ff
sete al
ret
function ScalarEqualsLong2
vmovdqu xmm0, [res]
vmovdqu xmm1, [val]
vpcmpeqq xmm0, xmm0, xmm1
vmovmskpd eax, xmm0
and eax, $03
cmp eax, $03
sete al
ret
function ScalarEqualsLong4
vmovdqu ymm0, [res]
vmovdqu ymm1, [val]
vpcmpeqq ymm0, ymm0, ymm1
vmovmskpd eax, ymm0
and eax, $0f
cmp eax, $0f
sete al
ret
function ScalarEqualsLong8
vmovdqu64 zmm0, [res]
vmovdqu64 zmm1, [val]
vpcmpeqq msk1, zmm0, zmm1
kmovw eax, msk1
and eax, $ff
cmp eax, $ff
sete al
ret
; </fold>
; <fold скалярное не равно>
function ScalarNotEqualsDouble2
vmovdqu xmm0, [res]
vmovdqu xmm1, [val]
vcmppd xmm0, xmm0, xmm1, $00
vmovmskpd eax, xmm0
and eax, $03
cmp eax, $03
setne al
ret
function ScalarNotEqualsDouble4
vmovdqu ymm0, [res]
vmovdqu ymm1, [val]
vcmppd ymm0, ymm0, ymm1, $00
vmovmskpd eax, ymm0
and eax, $0f
cmp eax, $0f
setne al
ret
function ScalarNotEqualsDouble8
vmovdqu64 zmm0, [res]
vmovdqu64 zmm1, [val]
vcmppd msk1, zmm0, zmm1, $00
kmovw eax, msk1
and eax, $ff
cmp eax, $ff
setne al
ret
function ScalarNotEqualsFloat2
vmovq xmm0, [res]
vmovq xmm1, [val]
vcmpps xmm0, xmm0, xmm1, $00
vmovmskps eax, xmm0
and eax, $03
cmp eax, $03
setne al
ret
function ScalarNotEqualsFloat4
vmovdqu xmm0, [res]
vmovdqu xmm1, [val]
vcmpps xmm0, xmm0, xmm1, $00
vmovmskps eax, xmm0
and eax, $0f
cmp eax, $0f
setne al
ret
function ScalarNotEqualsFloat8
vmovdqu ymm0, [res]
vmovdqu ymm1, [val]
vcmpps ymm0, ymm0, ymm1, $00
vmovmskps eax, ymm0
and eax, $ff
cmp eax, $ff
setne al
ret
function ScalarNotEqualsInt2
vmovq xmm0, [res]
vmovq xmm1, [val]
vpcmpeqd xmm0, xmm0, xmm1
vmovmskps eax, xmm0
and eax, $03
cmp eax, $03
setne al
ret
function ScalarNotEqualsInt4
vmovdqu xmm0, [res]
vmovdqu xmm1, [val]
vpcmpeqd xmm0, xmm0, xmm1
vmovmskps eax, xmm0
and eax, $0f
cmp eax, $0f
setne al
ret
function ScalarNotEqualsInt8
vmovdqu ymm0, [res]
vmovdqu ymm1, [val]
vpcmpeqd ymm0, ymm0, ymm1
vmovmskps eax, ymm0
and eax, $ff
cmp eax, $ff
setne al
ret
function ScalarNotEqualsLong2
vmovdqu xmm0, [res]
vmovdqu xmm1, [val]
vpcmpeqq xmm0, xmm0, xmm1
vmovmskpd eax, xmm0
and eax, $03
cmp eax, $03
setne al
ret
function ScalarNotEqualsLong4
vmovdqu ymm0, [res]
vmovdqu ymm1, [val]
vpcmpeqq ymm0, ymm0, ymm1
vmovmskpd eax, ymm0
and eax, $0f
cmp eax, $0f
setne al
ret
function ScalarNotEqualsLong8
vmovdqu64 zmm0, [res]
vmovdqu64 zmm1, [val]
vpcmpeqq msk1, zmm0, zmm1
kmovw eax, msk1
and eax, $ff
cmp eax, $ff
setne al
ret
; </fold>
; <fold скалярное умножение>
function ScalarMulDouble2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vshufpd xmm2, xmm1, xmm1, $02
vshufpd xmm3, xmm2, xmm2, $01
vmulpd xmm2, xmm2, [smuld2.c0]
vmulpd xmm1, xmm0, xmm3
vmulpd xmm0, xmm0, xmm2
vhaddpd xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarMulDouble4
enterq
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vshuff64x2 ymm4, ymm1, ymm1, $02
vshuff64x2 ymm6, ymm4, ymm4, $01
vshufpd ymm5, ymm4, ymm4, $05
vshufpd ymm7, ymm6, ymm6, $05
vmulpd ymm4, ymm4, [smuld4.q0]
vmulpd ymm5, ymm5, [smuld4.q1]
vmulpd ymm6, ymm6, [smuld4.q2]
vmulpd ymm7, ymm7, [smuld4.q3]
vmulpd ymm3, ymm0, ymm7
vmulpd ymm2, ymm0, ymm6
vmulpd ymm1, ymm0, ymm5
vmulpd ymm0, ymm0, ymm4
vextractf128 xmm4, ymm0, $01
vextractf128 xmm5, ymm1, $01
vextractf128 xmm6, ymm2, $01
vextractf128 xmm7, ymm3, $01
vinsertf128 ymm0, ymm0, xmm2, $01
vinsertf128 ymm2, ymm1, xmm3, $01
vinsertf128 ymm1, ymm4, xmm6, $01
vinsertf128 ymm3, ymm5, xmm7, $01
vaddpd ymm0, ymm0, ymm1
vaddpd ymm1, ymm2, ymm3
vshufpd ymm2, ymm0, ymm1, $00
vshufpd ymm3, ymm0, ymm1, $0f
vaddpd ymm0, ymm2, ymm3
vmovdqu yword [res], ymm0
leaveq
ret
function ScalarMulDouble8
entero
vmovdqu64 zmm00, [val1]
vmovdqu64 zmm01, [val2]
vshuff64x2 zmm08, zmm01, zmm01, $e4
vshuff64x2 zmm10, zmm08, zmm08, $b1
vshuff64x2 zmm12, zmm08, zmm08, $4e
vshuff64x2 zmm14, zmm08, zmm08, $1b
vshufpd zmm09, zmm08, zmm08, $55
vshufpd zmm11, zmm10, zmm10, $55
vshufpd zmm13, zmm12, zmm12, $55
vshufpd zmm15, zmm14, zmm14, $55
vmulpd zmm08, zmm08, [smuld8.o0]
vmulpd zmm09, zmm09, [smuld8.o1]
vmulpd zmm10, zmm10, [smuld8.o2]
vmulpd zmm11, zmm11, [smuld8.o3]
vmulpd zmm12, zmm12, [smuld8.o4]
vmulpd zmm13, zmm13, [smuld8.o5]
vmulpd zmm14, zmm14, [smuld8.o6]
vmulpd zmm15, zmm15, [smuld8.o7]
vmulpd zmm07, zmm00, zmm15, {rn-sae}
vmulpd zmm06, zmm00, zmm14, {rn-sae}
vmulpd zmm05, zmm00, zmm13, {rn-sae}
vmulpd zmm04, zmm00, zmm12, {rn-sae}
vmulpd zmm03, zmm00, zmm11, {rn-sae}
vmulpd zmm02, zmm00, zmm10, {rn-sae}
vmulpd zmm01, zmm00, zmm09, {rn-sae}
vmulpd zmm00, zmm00, zmm08, {rn-sae}
vextractf64x4 ymm08, zmm00, $01
vextractf64x4 ymm09, zmm01, $01
vextractf64x4 ymm10, zmm02, $01
vextractf64x4 ymm11, zmm03, $01
vextractf64x4 ymm12, zmm04, $01
vextractf64x4 ymm13, zmm05, $01
vextractf64x4 ymm14, zmm06, $01
vextractf64x4 ymm15, zmm07, $01
vinsertf64x4 zmm00, zmm00, ymm02, $01
vinsertf64x4 zmm02, zmm04, ymm06, $01
vinsertf64x4 zmm04, zmm01, ymm03, $01
vinsertf64x4 zmm06, zmm05, ymm07, $01
vinsertf64x4 zmm01, zmm08, ymm10, $01
vinsertf64x4 zmm03, zmm12, ymm14, $01
vinsertf64x4 zmm05, zmm09, ymm11, $01
vinsertf64x4 zmm07, zmm13, ymm15, $01
vaddpd zmm00, zmm00, zmm01
vaddpd zmm01, zmm02, zmm03
vaddpd zmm02, zmm04, zmm05
vaddpd zmm03, zmm06, zmm07
vshuff64x2 zmm04, zmm00, zmm01, $88
vshuff64x2 zmm05, zmm00, zmm01, $dd
vshuff64x2 zmm06, zmm02, zmm03, $88
vshuff64x2 zmm07, zmm02, zmm03, $dd
vaddpd zmm00, zmm04, zmm05
vaddpd zmm01, zmm06, zmm07
vshufpd zmm02, zmm00, zmm01, $00
vshufpd zmm03, zmm00, zmm01, $ff
vaddpd zmm00, zmm02, zmm03
vmovdqu64 zword [res], zmm00
leaveo
ret
function ScalarMulFloat2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vshufps xmm2, xmm1, xmm1, $e4
vshufps xmm3, xmm2, xmm2, $e1
vmulps xmm2, xmm2, [smulf2.c0]
vmulps xmm1, xmm0, xmm3
vmulps xmm0, xmm0, xmm2
vhaddps xmm0, xmm0, xmm1
vshufps xmm0, xmm0, xmm0, $d8
vmovq qword [res], xmm0
ret
function ScalarMulFloat4
enterq
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vshufps xmm4, xmm1, xmm1, $e4
vshufps xmm5, xmm4, xmm4, $b1
vshufps xmm6, xmm4, xmm4, $4e
vshufps xmm7, xmm4, xmm4, $1b
vmulps xmm4, xmm4, [smulf4.q0]
vmulps xmm5, xmm5, [smulf4.q1]
vmulps xmm6, xmm6, [smulf4.q2]
vmulps xmm7, xmm7, [smulf4.q3]
vmulps xmm3, xmm0, xmm7
vmulps xmm2, xmm0, xmm6
vmulps xmm1, xmm0, xmm5
vmulps xmm0, xmm0, xmm4
vhaddps xmm0, xmm0, xmm1
vhaddps xmm1, xmm2, xmm3
vhaddps xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
leaveq
ret
function ScalarMulFloat8
entero
vmovdqu ymm00, [val1]
vmovdqu ymm01, [val2]
vmovdqa ymm08, ymm01
vextractf128 xmm12, ymm08, $01
vinsertf128 ymm12, ymm12, xmm08, $01
vshufps ymm09, ymm08, ymm08, $b1
vshufps ymm10, ymm08, ymm08, $4e
vshufps ymm11, ymm08, ymm08, $1b
vshufps ymm13, ymm12, ymm12, $b1
vshufps ymm14, ymm12, ymm12, $4e
vshufps ymm15, ymm12, ymm12, $1b
vmulps ymm08, ymm08, [smulf8.o0]
vmulps ymm09, ymm09, [smulf8.o1]
vmulps ymm10, ymm10, [smulf8.o2]
vmulps ymm11, ymm11, [smulf8.o3]
vmulps ymm12, ymm12, [smulf8.o4]
vmulps ymm13, ymm13, [smulf8.o5]
vmulps ymm14, ymm14, [smulf8.o6]
vmulps ymm15, ymm15, [smulf8.o7]
vmulps ymm07, ymm00, ymm15
vmulps ymm06, ymm00, ymm14
vmulps ymm05, ymm00, ymm13
vmulps ymm04, ymm00, ymm12
vmulps ymm03, ymm00, ymm11
vmulps ymm02, ymm00, ymm10
vmulps ymm01, ymm00, ymm09
vmulps ymm00, ymm00, ymm08
vhaddps ymm00, ymm00, ymm01
vhaddps ymm01, ymm02, ymm03
vhaddps ymm02, ymm04, ymm05
vhaddps ymm03, ymm06, ymm07
vhaddps ymm00, ymm00, ymm01
vhaddps ymm01, ymm02, ymm03
vextractf128 xmm02, ymm00, $01
vextractf128 xmm03, ymm01, $01
vaddps xmm00, xmm00, xmm02
vaddps xmm01, xmm01, xmm03
vinsertf128 ymm00, ymm00, xmm01, $01
vmovdqu yword [res], ymm00
leaveo
ret
function ScalarMulInt2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vshufps xmm2, xmm1, xmm1, $e4
vshufps xmm3, xmm2, xmm2, $e1
vpmulld xmm2, xmm2, [smuli2.c0]
vpmulld xmm1, xmm0, xmm3
vpmulld xmm0, xmm0, xmm2
vphaddd xmm0, xmm0, xmm1
vshufps xmm0, xmm0, xmm0, $d8
vmovq qword [res], xmm0
ret
function ScalarMulInt4
enterq
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vshufps xmm4, xmm1, xmm1, $e4
vshufps xmm5, xmm4, xmm4, $b1
vshufps xmm6, xmm4, xmm4, $4e
vshufps xmm7, xmm4, xmm4, $1b
vpmulld xmm4, xmm4, [smuli4.q0]
vpmulld xmm5, xmm5, [smuli4.q1]
vpmulld xmm6, xmm6, [smuli4.q2]
vpmulld xmm7, xmm7, [smuli4.q3]
vpmulld xmm3, xmm0, xmm7
vpmulld xmm2, xmm0, xmm6
vpmulld xmm1, xmm0, xmm5
vpmulld xmm0, xmm0, xmm4
vphaddd xmm0, xmm0, xmm1
vphaddd xmm1, xmm2, xmm3
vphaddd xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
leaveq
ret
function ScalarMulInt8
entero
vmovdqu ymm00, [val1]
vmovdqu ymm01, [val2]
vmovdqa ymm08, ymm01
vextracti128 xmm12, ymm08, $01
vinserti128 ymm12, ymm12, xmm08, $01
vshufps ymm09, ymm08, ymm08, $b1
vshufps ymm10, ymm08, ymm08, $4e
vshufps ymm11, ymm08, ymm08, $1b
vshufps ymm13, ymm12, ymm12, $b1
vshufps ymm14, ymm12, ymm12, $4e
vshufps ymm15, ymm12, ymm12, $1b
vpmulld ymm08, ymm08, [smuli8.o0]
vpmulld ymm09, ymm09, [smuli8.o1]
vpmulld ymm10, ymm10, [smuli8.o2]
vpmulld ymm11, ymm11, [smuli8.o3]
vpmulld ymm12, ymm12, [smuli8.o4]
vpmulld ymm13, ymm13, [smuli8.o5]
vpmulld ymm14, ymm14, [smuli8.o6]
vpmulld ymm15, ymm15, [smuli8.o7]
vpmulld ymm07, ymm00, ymm15
vpmulld ymm06, ymm00, ymm14
vpmulld ymm05, ymm00, ymm13
vpmulld ymm04, ymm00, ymm12
vpmulld ymm03, ymm00, ymm11
vpmulld ymm02, ymm00, ymm10
vpmulld ymm01, ymm00, ymm09
vpmulld ymm00, ymm00, ymm08
vphaddd ymm00, ymm00, ymm01
vphaddd ymm01, ymm02, ymm03
vphaddd ymm02, ymm04, ymm05
vphaddd ymm03, ymm06, ymm07
vphaddd ymm00, ymm00, ymm01
vphaddd ymm01, ymm02, ymm03
vextracti128 xmm02, ymm00, $01
vextracti128 xmm03, ymm01, $01
vpaddd xmm00, xmm00, xmm02
vpaddd xmm01, xmm01, xmm03
vinserti128 ymm00, ymm00, xmm01, $01
vmovdqu yword [res], ymm00
leaveo
ret
function ScalarMulLong2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vshufpd xmm2, xmm1, xmm1, $02
vshufpd xmm3, xmm2, xmm2, $01
vpmullq xmm2, xmm2, [smull2.c0]
vpmullq xmm1, xmm0, xmm3
vpmullq xmm0, xmm0, xmm2
vshufpd xmm2, xmm0, xmm1, $00
vshufpd xmm3, xmm0, xmm1, $03
vpaddq xmm0, xmm2, xmm3
vmovdqu xword [res], xmm0
ret
function ScalarMulLong4
enterq
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vshufi64x2 ymm4, ymm1, ymm1, $02
vshufi64x2 ymm6, ymm4, ymm4, $01
vshufpd ymm5, ymm4, ymm4, $05
vshufpd ymm7, ymm6, ymm6, $05
vpmullq ymm4, ymm4, [smull4.q0]
vpmullq ymm5, ymm5, [smull4.q1]
vpmullq ymm6, ymm6, [smull4.q2]
vpmullq ymm7, ymm7, [smull4.q3]
vpmullq ymm3, ymm0, ymm7
vpmullq ymm2, ymm0, ymm6
vpmullq ymm1, ymm0, ymm5
vpmullq ymm0, ymm0, ymm4
vextracti128 xmm4, ymm0, $01
vextracti128 xmm5, ymm1, $01
vextracti128 xmm6, ymm2, $01
vextracti128 xmm7, ymm3, $01
vinserti128 ymm0, ymm0, xmm2, $01
vinserti128 ymm2, ymm1, xmm3, $01
vinserti128 ymm1, ymm4, xmm6, $01
vinserti128 ymm3, ymm5, xmm7, $01
vpaddq ymm0, ymm0, ymm1
vpaddq ymm1, ymm2, ymm3
vshufpd ymm2, ymm0, ymm1, $00
vshufpd ymm3, ymm0, ymm1, $0f
vpaddq ymm0, ymm2, ymm3
vmovdqu yword [res], ymm0
leaveq
ret
function ScalarMulLong8
entero
vmovdqu64 zmm00, [val1]
vmovdqu64 zmm01, [val2]
vshufi64x2 zmm08, zmm01, zmm01, $e4
vshufi64x2 zmm10, zmm08, zmm08, $b1
vshufi64x2 zmm12, zmm08, zmm08, $4e
vshufi64x2 zmm14, zmm08, zmm08, $1b
vshufpd zmm09, zmm08, zmm08, $55
vshufpd zmm11, zmm10, zmm10, $55
vshufpd zmm13, zmm12, zmm12, $55
vshufpd zmm15, zmm14, zmm14, $55
vpmullq zmm08, zmm08, [smull8.o0]
vpmullq zmm09, zmm09, [smull8.o1]
vpmullq zmm10, zmm10, [smull8.o2]
vpmullq zmm11, zmm11, [smull8.o3]
vpmullq zmm12, zmm12, [smull8.o4]
vpmullq zmm13, zmm13, [smull8.o5]
vpmullq zmm14, zmm14, [smull8.o6]
vpmullq zmm15, zmm15, [smull8.o7]
vpmullq zmm07, zmm00, zmm15
vpmullq zmm06, zmm00, zmm14
vpmullq zmm05, zmm00, zmm13
vpmullq zmm04, zmm00, zmm12
vpmullq zmm03, zmm00, zmm11
vpmullq zmm02, zmm00, zmm10
vpmullq zmm01, zmm00, zmm09
vpmullq zmm00, zmm00, zmm08
vextracti64x4 ymm08, zmm00, $01
vextracti64x4 ymm09, zmm01, $01
vextracti64x4 ymm10, zmm02, $01
vextracti64x4 ymm11, zmm03, $01
vextracti64x4 ymm12, zmm04, $01
vextracti64x4 ymm13, zmm05, $01
vextracti64x4 ymm14, zmm06, $01
vextracti64x4 ymm15, zmm07, $01
vinserti64x4 zmm00, zmm00, ymm02, $01
vinserti64x4 zmm02, zmm04, ymm06, $01
vinserti64x4 zmm04, zmm01, ymm03, $01
vinserti64x4 zmm06, zmm05, ymm07, $01
vinserti64x4 zmm01, zmm08, ymm10, $01
vinserti64x4 zmm03, zmm12, ymm14, $01
vinserti64x4 zmm05, zmm09, ymm11, $01
vinserti64x4 zmm07, zmm13, ymm15, $01
vpaddq zmm00, zmm00, zmm01
vpaddq zmm01, zmm02, zmm03
vpaddq zmm02, zmm04, zmm05
vpaddq zmm03, zmm06, zmm07
vshufi64x2 zmm04, zmm00, zmm01, $88
vshufi64x2 zmm05, zmm00, zmm01, $dd
vshufi64x2 zmm06, zmm02, zmm03, $88
vshufi64x2 zmm07, zmm02, zmm03, $dd
vpaddq zmm00, zmm04, zmm05
vpaddq zmm01, zmm06, zmm07
vshufpd zmm02, zmm00, zmm01, $00
vshufpd zmm03, zmm00, zmm01, $ff
vpaddq zmm00, zmm02, zmm03
vmovdqu64 zword [res], zmm00
leaveo
ret
; </fold>
; <fold скалярное отрицание>
function ScalarNegDouble2
vmovdqu xmm0, [val]
vpxor xmm1, xmm0, xmm0
vsubpd xmm0, xmm1, xmm0
vmovdqu xword [res], xmm0
ret
function ScalarNegDouble4
vmovdqu ymm0, [val]
vpxor ymm1, ymm0, ymm0
vsubpd ymm0, ymm1, ymm0
vmovdqu yword [res], ymm0
ret
function ScalarNegDouble8
vmovdqu64 zmm0, [val]
vpxorq zmm1, zmm0, zmm0
vsubpd zmm0, zmm1, zmm0
vmovdqu64 zword [res], zmm0
ret
function ScalarNegFloat2
vmovq xmm0, [val]
vpxor xmm1, xmm0, xmm0
vsubps xmm0, xmm1, xmm0
vmovq qword [res], xmm0
ret
function ScalarNegFloat4
vmovdqu xmm0, [val]
vpxor xmm1, xmm0, xmm0
vsubps xmm0, xmm1, xmm0
vmovdqu xword [res], xmm0
ret
function ScalarNegFloat8
vmovdqu ymm0, [val]
vpxor ymm1, ymm0, ymm0
vsubps ymm0, ymm1, ymm0
vmovdqu yword [res], ymm0
ret
function ScalarNegInt2
vmovq xmm0, [val]
vpxor xmm1, xmm0, xmm0
vpsubd xmm0, xmm1, xmm0
vmovq qword [res], xmm0
ret
function ScalarNegInt4
vmovdqu xmm0, [val]
vpxor xmm1, xmm0, xmm0
vpsubd xmm0, xmm1, xmm0
vmovdqu xword [res], xmm0
ret
function ScalarNegInt8
vmovdqu ymm0, [val]
vpxor ymm1, ymm0, ymm0
vpsubd ymm0, ymm1, ymm0
vmovdqu yword [res], ymm0
ret
function ScalarNegLong2
vmovdqu xmm0, [val]
vpxor xmm1, xmm0, xmm0
vpsubq xmm0, xmm1, xmm0
vmovdqu xword [res], xmm0
ret
function ScalarNegLong4
vmovdqu ymm0, [val]
vpxor ymm1, ymm0, ymm0
vpsubq ymm0, ymm1, ymm0
vmovdqu yword [res], ymm0
ret
function ScalarNegLong8
vmovdqu64 zmm0, [val]
vpxorq zmm1, zmm0, zmm0
vpsubq zmm0, zmm1, zmm0
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold скалярный сдвиг влево>
function ScalarShlInt2
and iVal2, $1f
vmovd xmm1, iVal2
vmovq xmm0, [val1]
vpslld xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function ScalarShlInt4
and iVal2, $1f
vmovd xmm1, iVal2
vmovdqu xmm0, [val1]
vpslld xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarShlInt8
and iVal2, $1f
vmovd xmm1, iVal2
vmovdqu ymm0, [val1]
vpslld ymm0, ymm0, xmm1
vmovdqu yword [res], ymm0
ret
function ScalarShlLong2
and iVal2, $3f
vmovd xmm1, iVal2
vmovdqu xmm0, [val1]
vpsllq xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarShlLong4
and iVal2, $3f
vmovd xmm1, iVal2
vmovdqu ymm0, [val1]
vpsllq ymm0, ymm0, xmm1
vmovdqu yword [res], ymm0
ret
function ScalarShlLong8
and iVal2, $3f
vmovd xmm1, iVal2
vmovdqu64 zmm0, [val1]
vpsllq zmm0, zmm0, xmm1
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold скалярный сдвиг вправо>
function ScalarShrInt2
and iVal2, $1f
vmovd xmm1, iVal2
vmovq xmm0, [val1]
vpsrad xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function ScalarShrInt4
and iVal2, $1f
vmovd xmm1, iVal2
vmovdqu xmm0, [val1]
vpsrad xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarShrInt8
and iVal2, $1f
vmovd xmm1, iVal2
vmovdqu ymm0, [val1]
vpsrad ymm0, ymm0, xmm1
vmovdqu yword [res], ymm0
ret
function ScalarShrLong2
and iVal2, $3f
vmovd xmm1, iVal2
vmovdqu xmm0, [val1]
vpsraq xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarShrLong4
and iVal2, $3f
vmovd xmm1, iVal2
vmovdqu ymm0, [val1]
vpsraq ymm0, ymm0, xmm1
vmovdqu yword [res], ymm0
ret
function ScalarShrLong8
and iVal2, $3f
vmovd xmm1, iVal2
vmovdqu64 zmm0, [val1]
vpsraq zmm0, zmm0, xmm1
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold скалярный беззнаковый сдвиг вправо>
function ScalarShruInt2
and iVal2, $1f
vmovd xmm1, iVal2
vmovq xmm0, [val1]
vpsrld xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function ScalarShruInt4
and iVal2, $1f
vmovd xmm1, iVal2
vmovdqu xmm0, [val1]
vpsrld xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarShruInt8
and iVal2, $1f
vmovd xmm1, iVal2
vmovdqu ymm0, [val1]
vpsrld ymm0, ymm0, xmm1
vmovdqu yword [res], ymm0
ret
function ScalarShruLong2
and iVal2, $3f
vmovd xmm1, iVal2
vmovdqu xmm0, [val1]
vpsrlq xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarShruLong4
and iVal2, $3f
vmovd xmm1, iVal2
vmovdqu ymm0, [val1]
vpsrlq ymm0, ymm0, xmm1
vmovdqu yword [res], ymm0
ret
function ScalarShruLong8
and iVal2, $3f
vmovd xmm1, iVal2
vmovdqu64 zmm0, [val1]
vpsrlq zmm0, zmm0, xmm1
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold скалярное вычитание>
function ScalarSubDouble2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vsubpd xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarSubDouble4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vsubpd ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function ScalarSubDouble8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vsubpd zmm0, zmm0, zmm1
vmovdqu64 zword [res], zmm0
ret
function ScalarSubFloat2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vsubps xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function ScalarSubFloat4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vsubps xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarSubFloat8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vsubps ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function ScalarSubInt2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpsubd xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function ScalarSubInt4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpsubd xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarSubInt8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpsubd ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function ScalarSubLong2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpsubq xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function ScalarSubLong4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpsubq ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function ScalarSubLong8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vpsubq zmm0, zmm0, zmm1
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold векторное сложение>
function VectorAddByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpaddb xmm0, xmm0, xmm1
vmovd eax, xmm0
mov word [res], ax
ret
function VectorAddByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpaddb xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorAddByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpaddb xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorAddShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpaddw xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorAddShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpaddw xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorAddShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpaddw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold векторное деление>
function VectorDivDouble2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vdivpd xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorDivDouble4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vdivpd ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function VectorDivDouble8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vdivpd zmm0, zmm0, zmm1
vmovdqu64 zword [res], zmm0
ret
function VectorDivFloat2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vdivps xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorDivFloat4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vdivps xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorDivFloat8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vdivps ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
; </fold>
; <fold векторное равно>
function VectorEqualsDouble
vcmpsd xmm0, xmm0, xmm1, $00
vmovq rax, xmm0
ret
function VectorEqualsDouble2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vcmppd xmm0, xmm0, xmm1, $00
vmovdqu xword [res], xmm0
ret
function VectorEqualsDouble4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vcmppd ymm0, ymm0, ymm1, $00
vmovdqu yword [res], ymm0
ret
function VectorEqualsDouble8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vcmppd msk1, zmm0, zmm1, $00
vmovdqa64 zmm0{k1}{z}, [long8.m1]
vmovdqu64 zword [res], zmm0
ret
function VectorEqualsFloat
vcmpss xmm0, xmm0, xmm1, $00
vmovd eax, xmm0
ret
function VectorEqualsFloat2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vcmpps xmm0, xmm0, xmm1, $00
vmovq qword [res], xmm0
ret
function VectorEqualsFloat4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vcmpps xmm0, xmm0, xmm1, $00
vmovdqu xword [res], xmm0
ret
function VectorEqualsFloat8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vcmpps ymm0, ymm0, ymm1, $00
vmovdqu yword [res], ymm0
ret
function VectorEqualsByte
cmp bRes, bVal
setne al
dec al
movsx eax, al
ret
function VectorEqualsByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpcmpeqb xmm0, xmm0, xmm1
vmovd eax, xmm0
mov word [res], ax
ret
function VectorEqualsByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpcmpeqb xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorEqualsByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpeqb xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorEqualsShort
cmp sRes, sVal
setne al
dec al
movsx eax, al
ret
function VectorEqualsShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpcmpeqw xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorEqualsShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpeqw xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorEqualsShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpeqw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorEqualsInt
cmp iRes, iVal
setne al
dec al
movsx eax, al
ret
function VectorEqualsInt2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpeqd xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorEqualsInt4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpeqd xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorEqualsInt8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpcmpeqd ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function VectorEqualsLong
cmp lRes, lVal
setne al
dec al
movsx rax, al
ret
function VectorEqualsLong2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpeqq xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorEqualsLong4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpcmpeqq ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function VectorEqualsLong8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vpcmpeqq msk1, zmm0, zmm1
vmovdqa64 zmm0{k1}{z}, [long8.m1]
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold векторное больше или равно>
function VectorGreaterOrEqualsDouble
vcmpsd xmm0, xmm1, xmm0, $02
vmovq rax, xmm0
ret
function VectorGreaterOrEqualsDouble2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vcmppd xmm0, xmm1, xmm0, $02
vmovdqu xword [res], xmm0
ret
function VectorGreaterOrEqualsDouble4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vcmppd ymm0, ymm1, ymm0, $02
vmovdqu yword [res], ymm0
ret
function VectorGreaterOrEqualsDouble8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vcmppd msk1, zmm1, zmm0, $02
vmovdqa64 zmm0{k1}{z}, [long8.m1]
vmovdqu64 zword [res], zmm0
ret
function VectorGreaterOrEqualsFloat
vcmpss xmm0, xmm1, xmm0, $02
vmovd eax, xmm0
ret
function VectorGreaterOrEqualsFloat2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vcmpps xmm0, xmm1, xmm0, $02
vmovq qword [res], xmm0
ret
function VectorGreaterOrEqualsFloat4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vcmpps xmm0, xmm1, xmm0, $02
vmovdqu xword [res], xmm0
ret
function VectorGreaterOrEqualsFloat8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vcmpps ymm0, ymm1, ymm0, $02
vmovdqu yword [res], ymm0
ret
function VectorGreaterOrEqualsByte
cmp bRes, bVal
setnge al
dec al
movsx eax, al
ret
function VectorGreaterOrEqualsByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpcmpgtb xmm0, xmm1, xmm0
vpxor xmm0, xmm0, [byte8.mask2]
vmovd eax, xmm0
mov word [res], ax
ret
function VectorGreaterOrEqualsByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpcmpgtb xmm0, xmm1, xmm0
vpxor xmm0, xmm0, [byte8.mask4]
vmovd dword [res], xmm0
ret
function VectorGreaterOrEqualsByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpgtb xmm0, xmm1, xmm0
vpxor xmm0, xmm0, [byte8.mask8]
vmovq qword [res], xmm0
ret
function VectorGreaterOrEqualsShort
cmp sRes, sVal
setnge al
dec al
movsx eax, al
ret
function VectorGreaterOrEqualsShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpcmpgtw xmm0, xmm1, xmm0
vpxor xmm0, xmm0, [short8.mask2]
vmovd dword [res], xmm0
ret
function VectorGreaterOrEqualsShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpgtw xmm0, xmm1, xmm0
vpxor xmm0, xmm0, [short8.mask4]
vmovq qword [res], xmm0
ret
function VectorGreaterOrEqualsShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpgtw xmm0, xmm1, xmm0
vpxor xmm0, xmm0, [short8.mask8]
vmovdqu xword [res], xmm0
ret
function VectorGreaterOrEqualsInt
cmp iRes, iVal
setnge al
dec al
movsx eax, al
ret
function VectorGreaterOrEqualsInt2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpgtd xmm0, xmm1, xmm0
vpxor xmm0, xmm0, [int8.mask2]
vmovq qword [res], xmm0
ret
function VectorGreaterOrEqualsInt4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpgtd xmm0, xmm1, xmm0
vpxor xmm0, xmm0, [int8.mask4]
vmovdqu xword [res], xmm0
ret
function VectorGreaterOrEqualsInt8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpcmpgtd ymm0, ymm1, ymm0
vpxor ymm0, ymm0, [int8.mask8]
vmovdqu yword [res], ymm0
ret
function VectorGreaterOrEqualsLong
cmp lRes, lVal
setnge al
dec al
movsx rax, al
ret
function VectorGreaterOrEqualsLong2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpgtq xmm0, xmm1, xmm0
vpxor xmm0, xmm0, [long8.mask2]
vmovdqu xword [res], xmm0
ret
function VectorGreaterOrEqualsLong4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpcmpgtq ymm0, ymm1, ymm0
vpxor ymm0, ymm0, [long8.mask4]
vmovdqu yword [res], ymm0
ret
function VectorGreaterOrEqualsLong8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vpcmpgtq msk1, zmm1, zmm0
knotw msk1, msk1
vmovdqa64 zmm0{k1}{z}, [long8.mask8]
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold векторное больше>
function VectorGreaterThanDouble
vcmpsd xmm0, xmm1, xmm0, $01
vmovq rax, xmm0
ret
function VectorGreaterThanDouble2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vcmppd xmm0, xmm1, xmm0, $01
vmovdqu xword [res], xmm0
ret
function VectorGreaterThanDouble4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vcmppd ymm0, ymm1, ymm0, $01
vmovdqu yword [res], ymm0
ret
function VectorGreaterThanDouble8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vcmppd msk1, zmm1, zmm0, $01
vmovdqa64 zmm0{k1}{z}, [long8.m1]
vmovdqu64 zword [res], zmm0
ret
function VectorGreaterThanFloat
vcmpss xmm0, xmm1, xmm0, $01
vmovd eax, xmm0
ret
function VectorGreaterThanFloat2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vcmpps xmm0, xmm1, xmm0, $01
vmovq qword [res], xmm0
ret
function VectorGreaterThanFloat4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vcmpps xmm0, xmm1, xmm0, $01
vmovdqu xword [res], xmm0
ret
function VectorGreaterThanFloat8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vcmpps ymm0, ymm1, ymm0, $01
vmovdqu yword [res], ymm0
ret
function VectorGreaterThanByte
cmp bRes, bVal
setng al
dec al
movsx eax, al
ret
function VectorGreaterThanByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpcmpgtb xmm0, xmm0, xmm1
vmovd eax, xmm0
mov word [res], ax
ret
function VectorGreaterThanByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpcmpgtb xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorGreaterThanByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpgtb xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorGreaterThanShort
cmp sRes, sVal
setng al
dec al
movsx eax, al
ret
function VectorGreaterThanShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpcmpgtw xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorGreaterThanShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpgtw xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorGreaterThanShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpgtw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorGreaterThanInt
cmp iRes, iVal
setng al
dec al
movsx eax, al
ret
function VectorGreaterThanInt2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpgtd xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorGreaterThanInt4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpgtd xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorGreaterThanInt8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpcmpgtd ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function VectorGreaterThanLong
cmp lRes, lVal
setng al
dec al
movsx rax, al
ret
function VectorGreaterThanLong2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpgtq xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorGreaterThanLong4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpcmpgtq ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function VectorGreaterThanLong8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vpcmpgtq msk1, zmm0, zmm1
vmovdqa64 zmm0{k1}{z}, [long8.m1]
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold векторное меньше или равно>
function VectorLessOrEqualsDouble
vcmpsd xmm0, xmm0, xmm1, $02
vmovq rax, xmm0
ret
function VectorLessOrEqualsDouble2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vcmppd xmm0, xmm0, xmm1, $02
vmovdqu xword [res], xmm0
ret
function VectorLessOrEqualsDouble4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vcmppd ymm0, ymm0, ymm1, $02
vmovdqu yword [res], ymm0
ret
function VectorLessOrEqualsDouble8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vcmppd msk1, zmm0, zmm1, $02
vmovdqa64 zmm0{k1}{z}, [long8.m1]
vmovdqu64 zword [res], zmm0
ret
function VectorLessOrEqualsFloat
vcmpss xmm0, xmm0, xmm1, $02
vmovd eax, xmm0
ret
function VectorLessOrEqualsFloat2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vcmpps xmm0, xmm0, xmm1, $02
vmovq qword [res], xmm0
ret
function VectorLessOrEqualsFloat4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vcmpps xmm0, xmm0, xmm1, $02
vmovdqu xword [res], xmm0
ret
function VectorLessOrEqualsFloat8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vcmpps ymm0, ymm0, ymm1, $02
vmovdqu yword [res], ymm0
ret
function VectorLessOrEqualsByte
cmp bRes, bVal
setnle al
dec al
movsx eax, al
ret
function VectorLessOrEqualsByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpcmpgtb xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [byte8.mask2]
vmovd eax, xmm0
mov word [res], ax
ret
function VectorLessOrEqualsByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpcmpgtb xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [byte8.mask4]
vmovd dword [res], xmm0
ret
function VectorLessOrEqualsByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpgtb xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [byte8.mask8]
vmovq qword [res], xmm0
ret
function VectorLessOrEqualsShort
cmp sRes, sVal
setnle al
dec al
movsx eax, al
ret
function VectorLessOrEqualsShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpcmpgtw xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [short8.mask2]
vmovd dword [res], xmm0
ret
function VectorLessOrEqualsShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpgtw xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [short8.mask4]
vmovq qword [res], xmm0
ret
function VectorLessOrEqualsShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpgtw xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [short8.mask8]
vmovdqu xword [res], xmm0
ret
function VectorLessOrEqualsInt
cmp iRes, iVal
setnle al
dec al
movsx eax, al
ret
function VectorLessOrEqualsInt2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpgtd xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [int8.mask2]
vmovq qword [res], xmm0
ret
function VectorLessOrEqualsInt4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpgtd xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [int8.mask4]
vmovdqu xword [res], xmm0
ret
function VectorLessOrEqualsInt8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpcmpgtd ymm0, ymm0, ymm1
vpxor ymm0, ymm0, [int8.mask8]
vmovdqu yword [res], ymm0
ret
function VectorLessOrEqualsLong
cmp lRes, lVal
setnle al
dec al
movsx rax, al
ret
function VectorLessOrEqualsLong2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpgtq xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [long8.mask2]
vmovdqu xword [res], xmm0
ret
function VectorLessOrEqualsLong4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpcmpgtq ymm0, ymm0, ymm1
vpxor ymm0, ymm0, [long8.mask4]
vmovdqu yword [res], ymm0
ret
function VectorLessOrEqualsLong8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vpcmpgtq msk1, zmm0, zmm1
knotw msk1, msk1
vmovdqa64 zmm0{k1}{z}, [long8.mask8]
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold векторное меньше>
function VectorLessThanDouble
vcmpsd xmm0, xmm0, xmm1, $01
vmovq rax, xmm0
ret
function VectorLessThanDouble2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vcmppd xmm0, xmm0, xmm1, $01
vmovdqu xword [res], xmm0
ret
function VectorLessThanDouble4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vcmppd ymm0, ymm0, ymm1, $01
vmovdqu yword [res], ymm0
ret
function VectorLessThanDouble8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vcmppd msk1, zmm0, zmm1, $01
vmovdqa64 zmm0{k1}{z}, [long8.m1]
vmovdqu64 zword [res], zmm0
ret
function VectorLessThanFloat
vcmpss xmm0, xmm0, xmm1, $01
vmovd eax, xmm0
ret
function VectorLessThanFloat2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vcmpps xmm0, xmm0, xmm1, $01
vmovq qword [res], xmm0
ret
function VectorLessThanFloat4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vcmpps xmm0, xmm0, xmm1, $01
vmovdqu xword [res], xmm0
ret
function VectorLessThanFloat8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vcmpps ymm0, ymm0, ymm1, $01
vmovdqu yword [res], ymm0
ret
function VectorLessThanByte
cmp bRes, bVal
setnl al
dec al
movsx eax, al
ret
function VectorLessThanByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpcmpgtb xmm0, xmm1, xmm0
vmovd eax, xmm0
mov word [res], ax
ret
function VectorLessThanByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpcmpgtb xmm0, xmm1, xmm0
vmovd dword [res], xmm0
ret
function VectorLessThanByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpgtb xmm0, xmm1, xmm0
vmovq qword [res], xmm0
ret
function VectorLessThanShort
cmp sRes, sVal
setnl al
dec al
movsx eax, al
ret
function VectorLessThanShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpcmpgtw xmm0, xmm1, xmm0
vmovd dword [res], xmm0
ret
function VectorLessThanShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpgtw xmm0, xmm1, xmm0
vmovq qword [res], xmm0
ret
function VectorLessThanShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpgtw xmm0, xmm1, xmm0
vmovdqu xword [res], xmm0
ret
function VectorLessThanInt
cmp iRes, iVal
setnl al
dec al
movsx eax, al
ret
function VectorLessThanInt2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpgtd xmm0, xmm1, xmm0
vmovq qword [res], xmm0
ret
function VectorLessThanInt4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpgtd xmm0, xmm1, xmm0
vmovdqu xword [res], xmm0
ret
function VectorLessThanInt8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpcmpgtd ymm0, ymm1, ymm0
vmovdqu yword [res], ymm0
ret
function VectorLessThanLong
cmp lRes, lVal
setnl al
dec al
movsx rax, al
ret
function VectorLessThanLong2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpgtq xmm0, xmm1, xmm0
vmovdqu xword [res], xmm0
ret
function VectorLessThanLong4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpcmpgtq ymm0, ymm1, ymm0
vmovdqu yword [res], ymm0
ret
function VectorLessThanLong8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vpcmpgtq msk1, zmm1, zmm0
vmovdqa64 zmm0{k1}{z}, [long8.m1]
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold векторное не равно>
function VectorNotEqualsDouble
vcmpsd xmm0, xmm0, xmm1, $04
vmovq rax, xmm0
ret
function VectorNotEqualsDouble2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vcmppd xmm0, xmm0, xmm1, $04
vmovdqu xword [res], xmm0
ret
function VectorNotEqualsDouble4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vcmppd ymm0, ymm0, ymm1, $04
vmovdqu yword [res], ymm0
ret
function VectorNotEqualsDouble8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vcmppd msk1, zmm0, zmm1, $04
vmovdqa64 zmm0{k1}{z}, [long8.m1]
vmovdqu64 zword [res], zmm0
ret
function VectorNotEqualsFloat
vcmpss xmm0, xmm0, xmm1, $04
vmovd eax, xmm0
ret
function VectorNotEqualsFloat2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vcmpps xmm0, xmm0, xmm1, $04
vmovq qword [res], xmm0
ret
function VectorNotEqualsFloat4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vcmpps xmm0, xmm0, xmm1, $04
vmovdqu xword [res], xmm0
ret
function VectorNotEqualsFloat8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vcmpps ymm0, ymm0, ymm1, $04
vmovdqu yword [res], ymm0
ret
function VectorNotEqualsByte
cmp bRes, bVal
sete al
dec al
movsx eax, al
ret
function VectorNotEqualsByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpcmpeqb xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [byte8.m1]
vmovd eax, xmm0
mov word [res], ax
ret
function VectorNotEqualsByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpcmpeqb xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [byte8.m1]
vmovd dword [res], xmm0
ret
function VectorNotEqualsByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpeqb xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [byte8.m1]
vmovq qword [res], xmm0
ret
function VectorNotEqualsShort
cmp sRes, sVal
sete al
dec al
movsx eax, al
ret
function VectorNotEqualsShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpcmpeqw xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [short8.m1]
vmovd dword [res], xmm0
ret
function VectorNotEqualsShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpeqw xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [short8.m1]
vmovq qword [res], xmm0
ret
function VectorNotEqualsShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpeqw xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [short8.m1]
vmovdqu xword [res], xmm0
ret
function VectorNotEqualsInt
cmp iRes, iVal
sete al
dec al
movsx eax, al
ret
function VectorNotEqualsInt2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpcmpeqd xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [int8.m1]
vmovq qword [res], xmm0
ret
function VectorNotEqualsInt4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpeqd xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [int8.m1]
vmovdqu xword [res], xmm0
ret
function VectorNotEqualsInt8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpcmpeqd ymm0, ymm0, ymm1
vpxor ymm0, ymm0, [int8.m1]
vmovdqu yword [res], ymm0
ret
function VectorNotEqualsLong
cmp lRes, lVal
sete al
dec al
movsx rax, al
ret
function VectorNotEqualsLong2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpcmpeqq xmm0, xmm0, xmm1
vpxor xmm0, xmm0, [long8.m1]
vmovdqu xword [res], xmm0
ret
function VectorNotEqualsLong4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpcmpeqq ymm0, ymm0, ymm1
vpxor ymm0, ymm0, [long8.m1]
vmovdqu yword [res], ymm0
ret
function VectorNotEqualsLong8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vpcmpeqq msk1, zmm0, zmm1
knotw msk1, msk1
vmovdqa64 zmm0{k1}{z}, [long8.m1]
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold векторное верхнее умножение с насыщением>
function VectorHiMulByte
movsx eax, bRes
movsx iVal, bVal
imul eax, iVal
add eax, $7f
sar eax, $07
cmp eax, $7f
jle .L.0000
mov eax, $7f
.L.0000: ret
function VectorHiMulByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpmovsxbw xmm0, xmm0
vpmovsxbw xmm1, xmm1
vpmullw xmm0, xmm0, xmm1
vpaddw xmm0, xmm0, [short2.$007f]
vpsraw xmm0, xmm0, $07
vpacksswb xmm0, xmm0, [zero]
vmovd eax, xmm0
mov word [res], ax
ret
function VectorHiMulByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpmovsxbw xmm0, xmm0
vpmovsxbw xmm1, xmm1
vpmullw xmm0, xmm0, xmm1
vpaddw xmm0, xmm0, [short4.$007f]
vpsraw xmm0, xmm0, $07
vpacksswb xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function VectorHiMulByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpmovsxbw xmm0, xmm0
vpmovsxbw xmm1, xmm1
vpmullw xmm0, xmm0, xmm1
vpaddw xmm0, xmm0, [short8.$007f]
vpsraw xmm0, xmm0, $07
vpacksswb xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function VectorHiMulShort
movsx eax, sRes
movsx iVal, sVal
imul eax, iVal
add eax, $7fff
sar eax, $0f
cmp eax, $7fff
jle .L.0000
mov eax, $7fff
.L.0000: ret
function VectorHiMulShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpmovsxwd xmm0, xmm0
vpmovsxwd xmm1, xmm1
vpmulld xmm0, xmm0, xmm1
vpaddd xmm0, xmm0, [int2.$7fff]
vpsrad xmm0, xmm0, $0f
vpackssdw xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function VectorHiMulShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpmovsxwd xmm0, xmm0
vpmovsxwd xmm1, xmm1
vpmulld xmm0, xmm0, xmm1
vpaddd xmm0, xmm0, [int4.$7fff]
vpsrad xmm0, xmm0, $0f
vpackssdw xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function VectorHiMulShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpmovsxwd ymm0, xmm0
vpmovsxwd ymm1, xmm1
vpmulld ymm0, ymm0, ymm1
vpaddd ymm0, ymm0, [int8.$7fff]
vpsrad ymm0, ymm0, $0f
vextracti128 xmm1, ymm0, $01
vpackssdw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold векторное беззнаковое верхнее умножение с насыщением>
function VectorHiMuluByte
movzx eax, bRes
movzx iVal, bVal
imul eax, iVal
add eax, $ff
movsx eax, ah
ret
function VectorHiMuluByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpmovzxbw xmm0, xmm0
vpmovzxbw xmm1, xmm1
vpmullw xmm0, xmm0, xmm1
vpaddw xmm0, xmm0, [short2.$00ff]
vpsrlw xmm0, xmm0, $08
vpackuswb xmm0, xmm0, [zero]
vmovd eax, xmm0
mov word [res], ax
ret
function VectorHiMuluByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpmovzxbw xmm0, xmm0
vpmovzxbw xmm1, xmm1
vpmullw xmm0, xmm0, xmm1
vpaddw xmm0, xmm0, [short4.$00ff]
vpsrlw xmm0, xmm0, $08
vpackuswb xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function VectorHiMuluByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpmovzxbw xmm0, xmm0
vpmovzxbw xmm1, xmm1
vpmullw xmm0, xmm0, xmm1
vpaddw xmm0, xmm0, [short8.$00ff]
vpsrlw xmm0, xmm0, $08
vpackuswb xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function VectorHiMuluShort
movzx eax, sRes
movzx iVal, sVal
imul eax, iVal
add eax, $ffff
sar eax, $10
ret
function VectorHiMuluShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpmovzxwd xmm0, xmm0
vpmovzxwd xmm1, xmm1
vpmulld xmm0, xmm0, xmm1
vpaddd xmm0, xmm0, [int2.$ffff]
vpsrld xmm0, xmm0, $10
vpackusdw xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function VectorHiMuluShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpmovzxwd xmm0, xmm0
vpmovzxwd xmm1, xmm1
vpmulld xmm0, xmm0, xmm1
vpaddd xmm0, xmm0, [int4.$ffff]
vpsrld xmm0, xmm0, $10
vpackusdw xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function VectorHiMuluShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpmovzxwd ymm0, xmm0
vpmovzxwd ymm1, xmm1
vpmulld ymm0, ymm0, ymm1
vpaddd ymm0, ymm0, [int8.$ffff]
vpsrld ymm0, ymm0, $10
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold векторная нижняя распаковка>
function VectorLoUnpackByte
movzx eax, bRes
ret
function VectorLoUnpackByte2
movzx iVal, word [val]
vmovd xmm0, iVal
vpmovzxbw xmm0, xmm0
vmovd dword [res], xmm0
ret
function VectorLoUnpackByte4
vmovd xmm0, [val]
vpmovzxbw xmm0, xmm0
vmovq qword [res], xmm0
ret
function VectorLoUnpackByte8
vmovq xmm0, [val]
vpmovzxbw xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function VectorLoUnpackShort
movzx eax, sRes
ret
function VectorLoUnpackShort2
vmovd xmm0, [val]
vpmovzxwd xmm0, xmm0
vmovq qword [res], xmm0
ret
function VectorLoUnpackShort4
vmovq xmm0, [val]
vpmovzxwd xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function VectorLoUnpackShort8
vmovdqu xmm0, [val]
vpmovzxwd ymm0, xmm0
vmovdqu yword [res], ymm0
ret
function VectorLoUnpackInt
mov eax, iRes
ret
function VectorLoUnpackInt2
vmovq xmm0, [val]
vpmovzxdq xmm0, xmm0
vmovdqu xword [res], xmm0
ret
function VectorLoUnpackInt4
vmovdqu xmm0, [val]
vpmovzxdq ymm0, xmm0
vmovdqu yword [res], ymm0
ret
function VectorLoUnpackInt8
vmovdqu ymm0, [val]
vpmovzxdq zmm0, ymm0
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold векторная верхняя распаковка>
function VectorUpUnpackByte
movzx eax, bRes
shl eax, $08
movsx eax, ax
ret
function VectorUpUnpackByte2
movzx iVal, word [val]
vmovd xmm0, iVal
vpmovzxbw xmm0, xmm0
vpsllw xmm0, xmm0, $08
vmovd dword [res], xmm0
ret
function VectorUpUnpackByte4
vmovd xmm0, [val]
vpmovzxbw xmm0, xmm0
vpsllw xmm0, xmm0, $08
vmovq qword [res], xmm0
ret
function VectorUpUnpackByte8
vmovq xmm0, [val]
vpmovzxbw xmm0, xmm0
vpsllw xmm0, xmm0, $08
vmovdqu xword [res], xmm0
ret
function VectorUpUnpackShort
movzx eax, sRes
shl eax, $10
ret
function VectorUpUnpackShort2
vmovd xmm0, [val]
vpmovzxwd xmm0, xmm0
vpslld xmm0, xmm0, $10
vmovq qword [res], xmm0
ret
function VectorUpUnpackShort4
vmovq xmm0, [val]
vpmovzxwd xmm0, xmm0
vpslld xmm0, xmm0, $10
vmovdqu xword [res], xmm0
ret
function VectorUpUnpackShort8
vmovdqu xmm0, [val]
vpmovzxwd ymm0, xmm0
vpslld ymm0, ymm0, $10
vmovdqu yword [res], ymm0
ret
function VectorUpUnpackInt
mov eax, iRes
shl rax, $20
ret
function VectorUpUnpackInt2
vmovq xmm0, [val]
vpmovzxdq xmm0, xmm0
vpsllq xmm0, xmm0, $20
vmovdqu xword [res], xmm0
ret
function VectorUpUnpackInt4
vmovdqu xmm0, [val]
vpmovzxdq ymm0, xmm0
vpsllq ymm0, ymm0, $20
vmovdqu yword [res], ymm0
ret
function VectorUpUnpackInt8
vmovdqu ymm0, [val]
vpmovzxdq zmm0, ymm0
vpsllq zmm0, zmm0, $20
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold векторное умножение>
function VectorMulDouble2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vmulpd xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorMulDouble4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vmulpd ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function VectorMulDouble8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vmulpd zmm0, zmm0, zmm1
vmovdqu64 zword [res], zmm0
ret
function VectorMulFloat2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vmulps xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorMulFloat4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vmulps xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorMulFloat8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vmulps ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function VectorMulByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpmovsxbw xmm0, xmm0
vpmovsxbw xmm1, xmm1
vpmullw xmm0, xmm0, xmm1
vpand xmm0, xmm0, [short2.$00ff]
vpackuswb xmm0, xmm0, [zero]
vmovd eax, xmm0
mov word [res], ax
ret
function VectorMulByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpmovsxbw xmm0, xmm0
vpmovsxbw xmm1, xmm1
vpmullw xmm0, xmm0, xmm1
vpand xmm0, xmm0, [short4.$00ff]
vpackuswb xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function VectorMulByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpmovsxbw xmm0, xmm0
vpmovsxbw xmm1, xmm1
vpmullw xmm0, xmm0, xmm1
vpand xmm0, xmm0, [short8.$00ff]
vpackuswb xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function VectorMulShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpmullw xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorMulShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpmullw xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorMulShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpmullw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorMulInt2
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpmulld xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorMulInt4
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpmulld xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorMulInt8
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpmulld ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function VectorMulLong2
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpmullq xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorMulLong4
vmovdqu ymm0, [val1]
vmovdqu ymm1, [val2]
vpmullq ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
function VectorMulLong8
vmovdqu64 zmm0, [val1]
vmovdqu64 zmm1, [val2]
vpmullq zmm0, zmm0, zmm1
vmovdqu64 zword [res], zmm0
ret
; </fold>
; <fold векторное отрицание>
function VectorNegByte2
movzx iVal, word [val]
vmovd xmm0, iVal
vpxor xmm1, xmm0, xmm0
vpsubb xmm0, xmm1, xmm0
vmovd eax, xmm0
mov word [res], ax
ret
function VectorNegByte4
vmovd xmm0, [val]
vpxor xmm1, xmm0, xmm0
vpsubb xmm0, xmm1, xmm0
vmovd dword [res], xmm0
ret
function VectorNegByte8
vmovq xmm0, [val]
vpxor xmm1, xmm0, xmm0
vpsubb xmm0, xmm1, xmm0
vmovq qword [res], xmm0
ret
function VectorNegShort2
vmovd xmm0, [val]
vpxor xmm1, xmm0, xmm0
vpsubw xmm0, xmm1, xmm0
vmovd dword [res], xmm0
ret
function VectorNegShort4
vmovq xmm0, [val]
vpxor xmm1, xmm0, xmm0
vpsubw xmm0, xmm1, xmm0
vmovq qword [res], xmm0
ret
function VectorNegShort8
vmovdqu xmm0, [val]
vpxor xmm1, xmm0, xmm0
vpsubw xmm0, xmm1, xmm0
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold векторная упаковка>
function VectorPackShort
movsx eax, sRes
sar eax, $08
cmp eax, $00
je .L.0001
jg .L.0000
mov eax, $00000000
ret
.L.0000: mov eax, $ffffffff
ret
.L.0001: movsx eax, bRes
ret
function VectorPackShort2
vmovd xmm0, [val]
vpackuswb xmm0, xmm0, [zero]
vmovd eax, xmm0
mov word [res], ax
ret
function VectorPackShort4
vmovq xmm0, [val]
vpackuswb xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function VectorPackShort8
vmovdqu xmm0, [val]
vpackuswb xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function VectorPackInt
mov eax, iRes
sar eax, $10
cmp eax, $00
je .L.0001
jg .L.0000
mov eax, $00000000
ret
.L.0000: mov eax, $ffffffff
ret
.L.0001: movsx eax, sRes
ret
function VectorPackInt2
vmovq xmm0, [val]
vpackusdw xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function VectorPackInt4
vmovdqu xmm0, [val]
vpackusdw xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function VectorPackInt8
vmovdqu ymm0, [val]
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorPackLong
mov rax, lRes
sar rax, $20
cmp eax, $00
je .L.0001
jg .L.0000
mov eax, $00000000
ret
.L.0000: mov eax, $ffffffff
ret
.L.0001: mov eax, iRes
ret
function VectorPackLong2
vmovdqu xmm0, [val]
vpxor xmm1, xmm0, xmm0
vpcmpgtq xmm1, xmm1, xmm0
vpcmpgtq xmm2, xmm0, [long8.maskd]
vpandn xmm0, xmm1, xmm0
vpor xmm0, xmm2, xmm0
vshufps xmm0, xmm0, xmm0, $d8
vmovq qword [res], xmm0
ret
function VectorPackLong4
vmovdqu ymm0, [val]
vpxor ymm1, ymm0, ymm0
vpcmpgtq ymm1, ymm1, ymm0
vpcmpgtq ymm2, ymm0, [long8.maskd]
vpandn ymm0, ymm1, ymm0
vpor ymm0, ymm2, ymm0
vshufps ymm0, ymm0, ymm0, $d8
vextracti128 xmm1, ymm0, $01
vpand xmm0, xmm0, [int8.m1_0]
vpslldq xmm1, xmm1, $08
vpor xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
function VectorPackLong8
vmovdqu64 zmm0, [val]
vpxorq zmm1, zmm0, zmm0
vpcmpgtq msk1, zmm1, zmm0
vpcmpgtq msk2, zmm0, [long8.maskd]
vmovdqa64 zmm2, [long8.m1]
vmovdqa64 zmm1{k1}{z}, zmm2
vmovdqa64 zmm2{k2}{z}, zmm2
vpandnq zmm0, zmm1, zmm0
vporq zmm0, zmm2, zmm0
vshufps zmm0, zmm0, zmm0, $d8
vshufi64x2 zmm0, zmm0, zmm0, $d8
vextracti64x4 ymm1, zmm0, $01
vpand ymm0, ymm0, [int8.m1_0]
vpslldq ymm1, ymm1, $08
vpor ymm0, ymm0, ymm1
vmovdqu yword [res], ymm0
ret
; </fold>
; <fold векторное сложение с насыщением>
function VectorSaturatedAddByte
movsx eax, bRes
movsx iVal, bVal
add eax, iVal
cmp eax, -$80
jge .L.0000
mov eax, -$80
ret
.L.0000: cmp eax, $7f
jle .L.0001
mov eax, $7f
.L.0001: ret
function VectorSaturatedAddByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpaddsb xmm0, xmm0, xmm1
vmovd eax, xmm0
mov word [res], ax
ret
function VectorSaturatedAddByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpaddsb xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorSaturatedAddByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpaddsb xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorSaturatedAddShort
movsx eax, sRes
movsx iVal, sVal
add eax, iVal
cmp eax, -$8000
jge .L.0000
mov eax, -$8000
ret
.L.0000: cmp eax, $7fff
jle .L.0001
mov eax, $7fff
.L.0001: ret
function VectorSaturatedAddShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpaddsw xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorSaturatedAddShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpaddsw xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorSaturatedAddShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpaddsw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold векторное беззнаковое сложение с насыщением>
function VectorSaturatedAdduByte
movzx eax, bRes
movzx iVal, bVal
add eax, iVal
cmp eax, $ff
jbe .L.0000
mov eax, $ff
.L.0000: movsx eax, al
ret
function VectorSaturatedAdduByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpaddusb xmm0, xmm0, xmm1
vmovd eax, xmm0
mov word [res], ax
ret
function VectorSaturatedAdduByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpaddusb xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorSaturatedAdduByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpaddusb xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorSaturatedAdduShort
movzx eax, sRes
movzx iVal, sVal
add eax, iVal
cmp eax, $ffff
jbe .L.0000
mov eax, $ffff
.L.0000: movsx eax, ax
ret
function VectorSaturatedAdduShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpaddusw xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorSaturatedAdduShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpaddusw xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorSaturatedAdduShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpaddusw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold векторное вычитание с насыщением>
function VectorSaturatedSubByte
movsx eax, bRes
movsx iVal, bVal
sub eax, iVal
cmp eax, -$80
jge .L.0000
mov eax, -$80
ret
.L.0000: cmp eax, $7f
jle .L.0001
mov eax, $7f
.L.0001: ret
function VectorSaturatedSubByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpsubsb xmm0, xmm0, xmm1
vmovd eax, xmm0
mov word [res], ax
ret
function VectorSaturatedSubByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpsubsb xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorSaturatedSubByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpsubsb xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorSaturatedSubShort
movsx eax, sRes
movsx iVal, sVal
sub eax, iVal
cmp eax, -$8000
jge .L.0000
mov eax, -$8000
ret
.L.0000: cmp eax, $7fff
jle .L.0001
mov eax, $7fff
.L.0001: ret
function VectorSaturatedSubShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpsubsw xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorSaturatedSubShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpsubsw xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorSaturatedSubShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpsubsw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold векторное беззнаковое вычитание с насыщением>
function VectorSaturatedSubuByte
movzx eax, bRes
movzx iVal, bVal
sub eax, iVal
cmp eax, $ff
jbe .L.0000
mov eax, $00
.L.0000: movsx eax, al
ret
function VectorSaturatedSubuByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpsubusb xmm0, xmm0, xmm1
vmovd eax, xmm0
mov word [res], ax
ret
function VectorSaturatedSubuByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpsubusb xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorSaturatedSubuByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpsubusb xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorSaturatedSubuShort
movzx eax, sRes
movzx iVal, sVal
sub eax, iVal
cmp eax, $ffff
jbe .L.0000
mov eax, $0000
.L.0000: movsx eax, ax
ret
function VectorSaturatedSubuShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpsubusw xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorSaturatedSubuShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpsubusw xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorSaturatedSubuShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpsubusw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold векторный сдвиг влево>
function VectorShlByte2
movzx iVal1, word [val1]
and iVal2, $1f
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpmovsxbd xmm0, xmm0
vpslld xmm0, xmm0, xmm1
vpand xmm0, xmm0, [int8.maskb]
vpackusdw xmm0, xmm0, [zero]
vpackuswb xmm0, xmm0, [zero]
vmovd eax, xmm0
mov word [res], ax
ret
function VectorShlByte4
and iVal2, $1f
vmovd xmm0, [val1]
vmovd xmm1, iVal2
vpmovsxbd xmm0, xmm0
vpslld xmm0, xmm0, xmm1
vpand xmm0, xmm0, [int8.maskb]
vpackusdw xmm0, xmm0, [zero]
vpackuswb xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function VectorShlByte8
and iVal2, $1f
vmovq xmm0, [val1]
vmovd xmm1, iVal2
vpmovsxbd ymm0, xmm0
vpslld ymm0, ymm0, xmm1
vpand ymm0, ymm0, [int8.maskb]
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vpackuswb xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function VectorShlShort2
and iVal2, $1f
vmovd xmm0, [val1]
vmovd xmm1, iVal2
vpmovsxwd xmm0, xmm0
vpslld xmm0, xmm0, xmm1
vpand xmm0, xmm0, [int8.maskw]
vpackusdw xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function VectorShlShort4
and iVal2, $1f
vmovq xmm0, [val1]
vmovd xmm1, iVal2
vpmovsxwd xmm0, xmm0
vpslld xmm0, xmm0, xmm1
vpand xmm0, xmm0, [int8.maskw]
vpackusdw xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function VectorShlShort8
and iVal2, $1f
vmovdqu xmm0, [val1]
vmovd xmm1, iVal2
vpmovsxwd ymm0, xmm0
vpslld ymm0, ymm0, xmm1
vpand ymm0, ymm0, [int8.maskw]
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold векторный сдвиг вправо>
function VectorShrByte2
movzx iVal1, word [val1]
and iVal2, $1f
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpmovsxbw xmm0, xmm0
vpsraw xmm0, xmm0, xmm1
vpand xmm0, xmm0, [short2.$00ff]
vpackuswb xmm0, xmm0, [zero]
vmovd eax, xmm0
mov word [res], ax
ret
function VectorShrByte4
and iVal2, $1f
vmovd xmm0, [val1]
vmovd xmm1, iVal2
vpmovsxbw xmm0, xmm0
vpsraw xmm0, xmm0, xmm1
vpand xmm0, xmm0, [short4.$00ff]
vpackuswb xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function VectorShrByte8
and iVal2, $1f
vmovq xmm0, [val1]
vmovd xmm1, iVal2
vpmovsxbw xmm0, xmm0
vpsraw xmm0, xmm0, xmm1
vpand xmm0, xmm0, [short8.$00ff]
vpackuswb xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function VectorShrShort2
and iVal2, $1f
vmovd xmm0, [val1]
vmovd xmm1, iVal2
vpsraw xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorShrShort4
and iVal2, $1f
vmovq xmm0, [val1]
vmovd xmm1, iVal2
vpsraw xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorShrShort8
and iVal2, $1f
vmovdqu xmm0, [val1]
vmovd xmm1, iVal2
vpsraw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold векторный беззнаковый сдвиг вправо>
function VectorShruByte2
movzx iVal1, word [val1]
and iVal2, $1f
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpmovsxbd xmm0, xmm0
vpsrld xmm0, xmm0, xmm1
vpand xmm0, xmm0, [int8.maskb]
vpackusdw xmm0, xmm0, [zero]
vpackuswb xmm0, xmm0, [zero]
vmovd eax, xmm0
mov word [res], ax
ret
function VectorShruByte4
and iVal2, $1f
vmovd xmm0, [val1]
vmovd xmm1, iVal2
vpmovsxbd xmm0, xmm0
vpsrld xmm0, xmm0, xmm1
vpand xmm0, xmm0, [int8.maskb]
vpackusdw xmm0, xmm0, [zero]
vpackuswb xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function VectorShruByte8
and iVal2, $1f
vmovq xmm0, [val1]
vmovd xmm1, iVal2
vpmovsxbd ymm0, xmm0
vpsrld ymm0, ymm0, xmm1
vpand ymm0, ymm0, [int8.maskb]
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vpackuswb xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function VectorShruShort2
and iVal2, $1f
vmovd xmm0, [val1]
vmovd xmm1, iVal2
vpmovsxwd xmm0, xmm0
vpsrld xmm0, xmm0, xmm1
vpand xmm0, xmm0, [int8.maskw]
vpackusdw xmm0, xmm0, [zero]
vmovd dword [res], xmm0
ret
function VectorShruShort4
and iVal2, $1f
vmovq xmm0, [val1]
vmovd xmm1, iVal2
vpmovsxwd xmm0, xmm0
vpsrld xmm0, xmm0, xmm1
vpand xmm0, xmm0, [int8.maskw]
vpackusdw xmm0, xmm0, [zero]
vmovq qword [res], xmm0
ret
function VectorShruShort8
and iVal2, $1f
vmovdqu xmm0, [val1]
vmovd xmm1, iVal2
vpmovsxwd ymm0, xmm0
vpsrld ymm0, ymm0, xmm1
vpand ymm0, ymm0, [int8.maskw]
vextracti128 xmm1, ymm0, $01
vpackusdw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
; </fold>
; <fold векторное вычитание>
function VectorSubByte2
movzx iVal1, word [val1]
movzx iVal2, word [val2]
vmovd xmm0, iVal1
vmovd xmm1, iVal2
vpsubb xmm0, xmm0, xmm1
vmovd eax, xmm0
mov word [res], ax
ret
function VectorSubByte4
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpsubb xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorSubByte8
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpsubb xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorSubShort2
vmovd xmm0, [val1]
vmovd xmm1, [val2]
vpsubw xmm0, xmm0, xmm1
vmovd dword [res], xmm0
ret
function VectorSubShort4
vmovq xmm0, [val1]
vmovq xmm1, [val2]
vpsubw xmm0, xmm0, xmm1
vmovq qword [res], xmm0
ret
function VectorSubShort8
vmovdqu xmm0, [val1]
vmovdqu xmm1, [val2]
vpsubw xmm0, xmm0, xmm1
vmovdqu xword [res], xmm0
ret
; </fold>
