sse2.esm

来自「开放源码的编译器open watcom 1.6.0版的源代码」· ESM 代码 · 共 601 行 · 第 1/2 页

.387
.386p
DGROUP		GROUP	_DATA
_TEXT		SEGMENT	DWORD PUBLIC USE32 'CODE'
		ASSUME CS:_TEXT, DS:DGROUP, SS:DGROUP
    addpd       xmm1,cs:0aH[bx+di]
    addpd       xmm1,cs:0aH[ebx+eax*2]
    addpd       xmm1,DGROUP:L$1
    addpd       xmm1,xmm7
    addsd       xmm1,cs:0aH[bx+di]
    addsd       xmm1,cs:0aH[ebx+eax*2]
    addsd       xmm1,DGROUP:L$1
    addsd       xmm1,xmm7
    andnpd      xmm1,cs:0aH[bx+di]
    andnpd      xmm1,cs:0aH[ebx+eax*2]
    andnpd      xmm1,DGROUP:L$1
    andnpd      xmm1,xmm7
    andpd       xmm1,cs:0aH[bx+di]
    andpd       xmm1,cs:0aH[ebx+eax*2]
    andpd       xmm1,DGROUP:L$1
    andpd       xmm1,xmm7
    clflush     cs:0aH[bx+di]
    clflush     cs:0aH[ebx+eax*2]
    clflush     DGROUP:L$1
    cmppd       xmm1,cs:0aH[bx+di],0
    cmppd       xmm1,cs:0aH[ebx+eax*2],0
    cmppd       xmm1,DGROUP:L$1,0
    cmppd       xmm1,xmm6,0
    cmpsd       xmm1,cs:0aH[bx+di],0
    cmpsd       xmm1,cs:0aH[ebx+eax*2],0
    cmpsd       xmm1,DGROUP:L$1,0
    cmpsd       xmm1,xmm6,0
    cmppd       xmm1,cs:0aH[bx+di],2
    cmppd       xmm1,cs:0aH[ebx+eax*2],2
    cmppd       xmm1,DGROUP:L$1,2
    cmppd       xmm1,xmm6,2
    cmpsd       xmm1,cs:0aH[bx+di],2
    cmpsd       xmm1,cs:0aH[ebx+eax*2],2
    cmpsd       xmm1,DGROUP:L$1,2
    cmpsd       xmm1,xmm6,2
    cmppd       xmm1,cs:0aH[bx+di],1
    cmppd       xmm1,cs:0aH[ebx+eax*2],1
    cmppd       xmm1,DGROUP:L$1,1
    cmppd       xmm1,xmm6,1
    cmpsd       xmm1,cs:0aH[bx+di],1
    cmpsd       xmm1,cs:0aH[ebx+eax*2],1
    cmpsd       xmm1,DGROUP:L$1,1
    cmpsd       xmm1,xmm6,1
    cmppd       xmm1,cs:0aH[bx+di],4
    cmppd       xmm1,cs:0aH[ebx+eax*2],4
    cmppd       xmm1,DGROUP:L$1,4
    cmppd       xmm1,xmm6,4
    cmpsd       xmm1,cs:0aH[bx+di],4
    cmpsd       xmm1,cs:0aH[ebx+eax*2],4
    cmpsd       xmm1,DGROUP:L$1,4
    cmpsd       xmm1,xmm6,4
    cmppd       xmm1,cs:0aH[bx+di],6
    cmppd       xmm1,cs:0aH[ebx+eax*2],6
    cmppd       xmm1,DGROUP:L$1,6
    cmppd       xmm1,xmm6,6
    cmpsd       xmm1,cs:0aH[bx+di],6
    cmpsd       xmm1,cs:0aH[ebx+eax*2],6
    cmpsd       xmm1,DGROUP:L$1,6
    cmpsd       xmm1,xmm6,6
    cmppd       xmm1,cs:0aH[bx+di],5
    cmppd       xmm1,cs:0aH[ebx+eax*2],5
    cmppd       xmm1,DGROUP:L$1,5
    cmppd       xmm1,xmm6,5
    cmpsd       xmm1,cs:0aH[bx+di],5
    cmpsd       xmm1,cs:0aH[ebx+eax*2],5
    cmpsd       xmm1,DGROUP:L$1,5
    cmpsd       xmm1,xmm6,5
    cmppd       xmm1,cs:0aH[bx+di],7
    cmppd       xmm1,cs:0aH[ebx+eax*2],7
    cmppd       xmm1,DGROUP:L$1,7
    cmppd       xmm1,xmm6,7
    cmpsd       xmm1,cs:0aH[bx+di],7
    cmpsd       xmm1,cs:0aH[ebx+eax*2],7
    cmpsd       xmm1,DGROUP:L$1,7
    cmpsd       xmm1,xmm6,7
    cmppd       xmm1,cs:0aH[bx+di],0ffH
    cmppd       xmm1,cs:0aH[ebx+eax*2],0ffH
    cmppd       xmm1,DGROUP:L$1,0ffH
    cmppd       xmm1,xmm7,0ffH
    cmpsd       xmm1,cs:0aH[bx+di],0ffH
    cmpsd       xmm1,cs:0aH[ebx+eax*2],0ffH
    cmpsd       xmm1,DGROUP:L$1,0ffH
    cmpsd       xmm1,xmm7,0ffH
    cmppd       xmm1,cs:0aH[bx+di],3
    cmppd       xmm1,cs:0aH[ebx+eax*2],3
    cmppd       xmm1,DGROUP:L$1,3
    cmppd       xmm1,xmm6,3
    cmpsd       xmm1,cs:0aH[bx+di],3
    cmpsd       xmm1,cs:0aH[ebx+eax*2],3
    cmpsd       xmm1,DGROUP:L$1,3
    cmpsd       xmm1,xmm6,3
    comisd      xmm1,cs:0aH[bx+di]
    comisd      xmm1,cs:0aH[ebx+eax*2]
    comisd      xmm1,DGROUP:L$1
    comisd      xmm1,xmm7
    cvtdq2pd    xmm1,cs:0aH[bx+di]
    cvtdq2pd    xmm1,cs:0aH[ebx+eax*2]
    cvtdq2pd    xmm1,DGROUP:L$1
    cvtdq2pd    xmm1,xmm7
    cvtdq2ps    xmm1,cs:0aH[bx+di]
    cvtdq2ps    xmm1,cs:0aH[ebx+eax*2]
    cvtdq2ps    xmm1,DGROUP:L$1
    cvtdq2ps    xmm1,xmm7
    cvtpd2dq    xmm1,cs:0aH[bx+di]
    cvtpd2dq    xmm1,cs:0aH[ebx+eax*2]
    cvtpd2dq    xmm1,DGROUP:L$1
    cvtpd2dq    xmm1,xmm7
    cvtpd2pi    mm1,cs:0aH[bx+di]
    cvtpd2pi    mm1,cs:0aH[ebx+eax*2]
    cvtpd2pi    mm1,DGROUP:L$1
    cvtpd2pi    mm1,xmm7
    cvtpd2ps    xmm1,cs:0aH[bx+di]
    cvtpd2ps    xmm1,cs:0aH[ebx+eax*2]
    cvtpd2ps    xmm1,DGROUP:L$1
    cvtpd2ps    xmm1,xmm7
    cvtpi2pd    xmm1,cs:0aH[bx+di]
    cvtpi2pd    xmm1,cs:0aH[ebx+eax*2]
    cvtpi2pd    xmm1,mm7
    cvtpi2pd    xmm1,DGROUP:L$1
    cvtps2dq    xmm1,cs:0aH[bx+di]
    cvtps2dq    xmm1,cs:0aH[ebx+eax*2]
    cvtps2dq    xmm1,DGROUP:L$1
    cvtps2dq    xmm1,xmm7
    cvtps2pd    xmm1,cs:0aH[bx+di]
    cvtps2pd    xmm1,cs:0aH[ebx+eax*2]
    cvtps2pd    xmm1,DGROUP:L$1
    cvtps2pd    xmm1,xmm7
    cvtsd2si    ebx,cs:0aH[bx+di]
    cvtsd2si    ebx,cs:0aH[ebx+eax*2]
    cvtsd2si    ebx,DGROUP:L$1
    cvtsd2si    ebx,xmm7
    cvtsd2ss    xmm1,cs:0aH[bx+di]
    cvtsd2ss    xmm1,cs:0aH[ebx+eax*2]
    cvtsd2ss    xmm1,DGROUP:L$1
    cvtsd2ss    xmm1,xmm7
    cvtsi2sd    xmm5,cs:0aH[bx+di]
    cvtsi2sd    xmm5,cs:0aH[ebx+eax*2]
    cvtsi2sd    xmm5,ecx
    cvtsi2sd    xmm5,DGROUP:L$1
    cvtss2sd    xmm1,cs:0aH[bx+di]
    cvtss2sd    xmm1,cs:0aH[ebx+eax*2]
    cvtss2sd    xmm1,DGROUP:L$1
    cvtss2sd    xmm1,xmm7
    cvttpd2dq   xmm1,cs:0aH[bx+di]
    cvttpd2dq   xmm1,cs:0aH[ebx+eax*2]
    cvttpd2dq   xmm1,DGROUP:L$1
    cvttpd2dq   xmm1,xmm7
    cvttpd2pi   mm1,cs:0aH[bx+di]
    cvttpd2pi   mm1,cs:0aH[ebx+eax*2]
    cvttpd2pi   mm1,DGROUP:L$1
    cvttpd2pi   mm1,xmm7
    cvttps2dq   xmm1,cs:0aH[bx+di]
    cvttps2dq   xmm1,cs:0aH[ebx+eax*2]
    cvttps2dq   xmm1,DGROUP:L$1
    cvttps2dq   xmm1,xmm7
    cvttsd2si   ebx,cs:0aH[bx+di]
    cvttsd2si   ebx,cs:0aH[ebx+eax*2]
    cvttsd2si   ebx,DGROUP:L$1
    cvttsd2si   ebx,xmm7
    divpd       xmm1,cs:0aH[bx+di]
    divpd       xmm1,cs:0aH[ebx+eax*2]
    divpd       xmm1,DGROUP:L$1
    divpd       xmm1,xmm7
    divsd       xmm1,cs:0aH[bx+di]
    divsd       xmm1,cs:0aH[ebx+eax*2]
    divsd       xmm1,DGROUP:L$1
    divsd       xmm1,xmm7
    lfence      
    maskmovdqu  xmm1,xmm5
    maxpd       xmm1,cs:0aH[bx+di]
    maxpd       xmm1,cs:0aH[ebx+eax*2]
    maxpd       xmm1,DGROUP:L$1
    maxpd       xmm1,xmm7
    maxsd       xmm1,cs:0aH[bx+di]
    maxsd       xmm1,cs:0aH[ebx+eax*2]
    maxsd       xmm1,DGROUP:L$1
    maxsd       xmm1,xmm7
    mfence      
    minpd       xmm1,cs:0aH[bx+di]
    minpd       xmm1,cs:0aH[ebx+eax*2]
    minpd       xmm1,DGROUP:L$1
    minpd       xmm1,xmm7
    minsd       xmm1,cs:0aH[bx+di]
    minsd       xmm1,cs:0aH[ebx+eax*2]
    minsd       xmm1,DGROUP:L$1
    minsd       xmm1,xmm7
    movapd      cs:0aH[bx+di],xmm7
    movapd      cs:0aH[ebx+eax*2],xmm7
    movapd      DGROUP:L$1,xmm7
    movapd      xmm1,cs:0aH[bx+di]
    movapd      xmm1,cs:0aH[ebx+eax*2]
    movapd      xmm1,DGROUP:L$1
    movapd      xmm1,xmm7
    movd        cs:0aH[bx+di],xmm1
    movd        cs:0aH[ebx+eax*2],xmm1
    movd        ebx,xmm3
    movd        DGROUP:L$1,xmm1
    movd        xmm1,cs:0aH[bx+di]
    movd        xmm1,cs:0aH[ebx+eax*2]
    movd        xmm1,DGROUP:L$1
    movd        xmm7,ebx
    movdq2q     mm7,xmm1
    movdqa      cs:0aH[bx+di],xmm1
    movdqa      cs:0aH[ebx+eax*2],xmm1
    movdqa      DGROUP:L$1,xmm1
    movdqa      xmm1,cs:0aH[bx+di]
    movdqa      xmm1,cs:0aH[ebx+eax*2]
    movdqa      xmm1,DGROUP:L$1
    movdqa      xmm1,xmm7
    movdqu      cs:0aH[bx+di],xmm3
    movdqu      cs:0aH[ebx+eax*2],xmm3
    movdqu      DGROUP:L$1,xmm3
    movdqu      xmm1,cs:0aH[bx+di]
    movdqu      xmm1,cs:0aH[ebx+eax*2]
    movdqu      xmm1,DGROUP:L$1
    movdqu      xmm1,xmm7
    movhpd      cs:0aH[bx+di],xmm3
    movhpd      cs:0aH[ebx+eax*2],xmm3
    movhpd      DGROUP:L$1,xmm3
    movhpd      xmm1,cs:0aH[bx+di]
    movhpd      xmm1,cs:0aH[ebx+eax*2]
    movhpd      xmm1,DGROUP:L$1
    movlpd      cs:0aH[bx+di],xmm7
    movlpd      cs:0aH[ebx+eax*2],xmm7
    movlpd      DGROUP:L$1,xmm7
    movlpd      xmm1,cs:0aH[bx+di]
    movlpd      xmm1,cs:0aH[ebx+eax*2]
    movlpd      xmm1,DGROUP:L$1
    movmskpd    ebx,xmm1
    movntdq     cs:0aH[bx+di],xmm1
    movntdq     cs:0aH[ebx+eax*2],xmm1
    movntdq     DGROUP:L$1,xmm1
    movnti      cs:0aH[bx+di],ebx
    movnti      cs:0aH[ebx+eax*2],ebx
    movnti      DGROUP:L$2,ebx
    movntpd     cs:0aH[bx+di],xmm1
    movntpd     cs:0aH[ebx+eax*2],xmm1
    movntpd     DGROUP:L$1,xmm1
    movq        cs:0aH[bx+di],xmm1
    movq        cs:0aH[ebx+eax*2],xmm1
    movq        DGROUP:L$1,xmm1
    movq        xmm1,cs:0aH[bx+di]
    movq        xmm1,cs:0aH[ebx+eax*2]
    movq        xmm1,DGROUP:L$1
    movq        xmm1,xmm7
    movq2dq     xmm1,mm7
    movsd       cs:0aH[bx+di],xmm1
    movsd       cs:0aH[ebx+eax*2],xmm1
    movsd       DGROUP:L$1,xmm1
    movsd       xmm1,cs:0aH[bx+di]
    movsd       xmm1,cs:0aH[ebx+eax*2]
    movsd       xmm1,DGROUP:L$1
    movsd       xmm1,xmm7
    movupd      cs:0aH[bx+di],xmm1
    movupd      cs:0aH[ebx+eax*2],xmm1
    movupd      DGROUP:L$1,xmm1
    movupd      xmm1,cs:0aH[bx+di]
    movupd      xmm1,cs:0aH[ebx+eax*2]
    movupd      xmm1,DGROUP:L$1
    movupd      xmm1,xmm7
    mulpd       xmm1,cs:0aH[bx+di]
    mulpd       xmm1,cs:0aH[ebx+eax*2]
    mulpd       xmm1,DGROUP:L$1
    mulpd       xmm1,xmm7
    mulsd       xmm1,cs:0aH[bx+di]
    mulsd       xmm1,cs:0aH[ebx+eax*2]
    mulsd       xmm1,DGROUP:L$1
    mulsd       xmm1,xmm7
    orpd        xmm1,cs:0aH[bx+di]
    orpd        xmm1,cs:0aH[ebx+eax*2]
    orpd        xmm1,DGROUP:L$1
    orpd        xmm1,xmm7
    packssdw    xmm1,cs:0aH[bx+di]
    packssdw    xmm1,cs:0aH[ebx+eax*2]
    packssdw    xmm1,DGROUP:L$1
    packssdw    xmm1,xmm7
    packsswb    xmm1,cs:0aH[bx+di]
    packsswb    xmm1,cs:0aH[ebx+eax*2]
    packsswb    xmm1,DGROUP:L$1
    packsswb    xmm1,xmm7
    packuswb    xmm1,cs:0aH[bx+di]
    packuswb    xmm1,cs:0aH[ebx+eax*2]
    packuswb    xmm1,DGROUP:L$1
    packuswb    xmm1,xmm7
    paddb       xmm1,cs:0aH[bx+di]
    paddb       xmm1,cs:0aH[ebx+eax*2]
    paddb       xmm1,DGROUP:L$1
    paddb       xmm1,xmm7
    paddd       xmm1,cs:0aH[bx+di]
    paddd       xmm1,cs:0aH[ebx+eax*2]
    paddd       xmm1,DGROUP:L$1
    paddd       xmm1,xmm7
    paddq       xmm1,cs:0aH[bx+di]
    paddq       xmm1,cs:0aH[ebx+eax*2]
    paddq       xmm1,DGROUP:L$1
    paddq       xmm1,xmm7