include ksamd64.inc
EXTERNDEF s_sosemanukMulTables:FAR

.CODE
    ALIGN   8
Baseline_Add	PROC
	lea		rdx, [rdx+8*rcx]
	lea		r8, [r8+8*rcx]
	lea		r9, [r9+8*rcx]
	neg		rcx					; rcx is negative index
	jz		$1@Baseline_Add
	mov		rax,[r8+8*rcx]
	add		rax,[r9+8*rcx]
	mov		[rdx+8*rcx],rax
$0@Baseline_Add:
	mov		rax,[r8+8*rcx+8]
	adc		rax,[r9+8*rcx+8]
	mov		[rdx+8*rcx+8],rax
	lea		rcx,[rcx+2]			; advance index, avoid inc which causes slowdown on Intel Core 2
	jrcxz	$1@Baseline_Add		; loop until rcx overflows and becomes zero
	mov		rax,[r8+8*rcx]
	adc		rax,[r9+8*rcx]
	mov		[rdx+8*rcx],rax
	jmp		$0@Baseline_Add
$1@Baseline_Add:
	mov		rax, 0
	adc		rax, rax			; store carry into rax (return result register)
	ret
Baseline_Add ENDP

    ALIGN   8
Baseline_Sub	PROC
	lea		rdx, [rdx+8*rcx]
	lea		r8, [r8+8*rcx]
	lea		r9, [r9+8*rcx]
	neg		rcx					; rcx is negative index
	jz		$1@Baseline_Sub
	mov		rax,[r8+8*rcx]
	sub		rax,[r9+8*rcx]
	mov		[rdx+8*rcx],rax
$0@Baseline_Sub:
	mov		rax,[r8+8*rcx+8]
	sbb		rax,[r9+8*rcx+8]
	mov		[rdx+8*rcx+8],rax
	lea		rcx,[rcx+2]			; advance index, avoid inc which causes slowdown on Intel Core 2
	jrcxz	$1@Baseline_Sub		; loop until rcx overflows and becomes zero
	mov		rax,[r8+8*rcx]
	sbb		rax,[r9+8*rcx]
	mov		[rdx+8*rcx],rax
	jmp		$0@Baseline_Sub
$1@Baseline_Sub:
	mov		rax, 0
	adc		rax, rax			; store carry into rax (return result register)

	ret
Baseline_Sub ENDP

ALIGN 8
Salsa20_OperateKeystream PROC FRAME
mov r10, [rsp + 5*8]
alloc_stack(10*16 + 32*16 + 8)
save_xmm128 xmm6, 0200h
save_xmm128 xmm7, 0210h
save_xmm128 xmm8, 0220h
save_xmm128 xmm9, 0230h
save_xmm128 xmm10, 0240h
save_xmm128 xmm11, 0250h
save_xmm128 xmm12, 0260h
save_xmm128 xmm13, 0270h
save_xmm128 xmm14, 0280h
save_xmm128 xmm15, 0290h
.endprolog
cmp r8, 4
jl label5
movdqa xmm0, [r10 + 0*16]
movdqa xmm1, [r10 + 1*16]
movdqa xmm2, [r10 + 2*16]
movdqa xmm3, [r10 + 3*16]
pshufd xmm4, xmm0, 0*64+0*16+0*4+0
movdqa [rsp + (0*4+0)*16 + 256], xmm4
pshufd xmm4, xmm0, 1*64+1*16+1*4+1
movdqa [rsp + (0*4+1)*16 + 256], xmm4
pshufd xmm4, xmm0, 2*64+2*16+2*4+2
movdqa [rsp + (0*4+2)*16 + 256], xmm4
pshufd xmm4, xmm0, 3*64+3*16+3*4+3
movdqa [rsp + (0*4+3)*16 + 256], xmm4
pshufd xmm4, xmm1, 0*64+0*16+0*4+0
movdqa [rsp + (1*4+0)*16 + 256], xmm4
pshufd xmm4, xmm1, 2*64+2*16+2*4+2
movdqa [rsp + (1*4+2)*16 + 256], xmm4
pshufd xmm4, xmm1, 3*64+3*16+3*4+3
movdqa [rsp + (1*4+3)*16 + 256], xmm4
pshufd xmm4, xmm2, 1*64+1*16+1*4+1
movdqa [rsp + (2*4+1)*16 + 256], xmm4
pshufd xmm4, xmm2, 2*64+2*16+2*4+2
movdqa [rsp + (2*4+2)*16 + 256], xmm4
pshufd xmm4, xmm2, 3*64+3*16+3*4+3
movdqa [rsp + (2*4+3)*16 + 256], xmm4
pshufd xmm4, xmm3, 0*64+0*16+0*4+0
movdqa [rsp + (3*4+0)*16 + 256], xmm4
pshufd xmm4, xmm3, 1*64+1*16+1*4+1
movdqa [rsp + (3*4+1)*16 + 256], xmm4
pshufd xmm4, xmm3, 2*64+2*16+2*4+2
movdqa [rsp + (3*4+2)*16 + 256], xmm4
pshufd xmm4, xmm3, 3*64+3*16+3*4+3
movdqa [rsp + (3*4+3)*16 + 256], xmm4
label1:
mov eax, dword ptr [r10 + 8*4]
mov r11d, dword ptr [r10 + 5*4]
mov dword ptr [rsp + 8*16 + 0*4 + 256], eax
mov dword ptr [rsp + 5*16 + 0*4 + 256], r11d
add eax, 1
adc r11d, 0
mov dword ptr [rsp + 8*16 + 1*4 + 256], eax
mov dword ptr [rsp + 5*16 + 1*4 + 256], r11d
add eax, 1
adc r11d, 0
mov dword ptr [rsp + 8*16 + 2*4 + 256], eax
mov dword ptr [rsp + 5*16 + 2*4 + 256], r11d
add eax, 1
adc r11d, 0
mov dword ptr [rsp + 8*16 + 3*4 + 256], eax
mov dword ptr [rsp + 5*16 + 3*4 + 256], r11d
add eax, 1
adc r11d, 0
mov dword ptr [r10 + 8*4], eax
mov dword ptr [r10 + 5*4], r11d
movdqa xmm0, [rsp + 12*16 + 1*256]
movdqa xmm4, [rsp + 13*16 + 1*256]
movdqa xmm8, [rsp + 14*16 + 1*256]
movdqa xmm12, [rsp + 15*16 + 1*256]
movdqa xmm2, [rsp + 0*16 + 1*256]
movdqa xmm6, [rsp + 1*16 + 1*256]
movdqa xmm10, [rsp + 2*16 + 1*256]
movdqa xmm14, [rsp + 3*16 + 1*256]
paddd xmm0, xmm2
paddd xmm4, xmm6
paddd xmm8, xmm10
paddd xmm12, xmm14
movdqa xmm1, xmm0
movdqa xmm5, xmm4
movdqa xmm9, xmm8
movdqa xmm13, xmm12
pslld xmm0, 7
pslld xmm4, 7
pslld xmm8, 7
pslld xmm12, 7
psrld xmm1, 32-7
psrld xmm5, 32-7
psrld xmm9, 32-7
psrld xmm13, 32-7
pxor xmm0, [rsp + 4*16 + 1*256]
pxor xmm4, [rsp + 5*16 + 1*256]
pxor xmm8, [rsp + 6*16 + 1*256]
pxor xmm12, [rsp + 7*16 + 1*256]
pxor xmm0, xmm1
pxor xmm4, xmm5
pxor xmm8, xmm9
pxor xmm12, xmm13
movdqa [rsp + 4*16], xmm0
movdqa [rsp + 5*16], xmm4
movdqa [rsp + 6*16], xmm8
movdqa [rsp + 7*16], xmm12
movdqa xmm1, xmm0
movdqa xmm5, xmm4
movdqa xmm9, xmm8
movdqa xmm13, xmm12
paddd xmm0, xmm2
paddd xmm4, xmm6
paddd xmm8, xmm10
paddd xmm12, xmm14
movdqa xmm3, xmm0
movdqa xmm7, xmm4
movdqa xmm11, xmm8
movdqa xmm15, xmm12
pslld xmm0, 9
pslld xmm4, 9
pslld xmm8, 9
pslld xmm12, 9
psrld xmm3, 32-9
psrld xmm7, 32-9
psrld xmm11, 32-9
psrld xmm15, 32-9
pxor xmm0, [rsp + 8*16 + 1*256]
pxor xmm4, [rsp + 9*16 + 1*256]
pxor xmm8, [rsp + 10*16 + 1*256]
pxor xmm12, [rsp + 11*16 + 1*256]
pxor xmm0, xmm3
pxor xmm4, xmm7
pxor xmm8, xmm11
pxor xmm12, xmm15
movdqa [rsp + 8*16], xmm0
movdqa [rsp + 9*16], xmm4
movdqa [rsp + 10*16], xmm8
movdqa [rsp + 11*16], xmm12
movdqa xmm3, xmm0
movdqa xmm7, xmm4
movdqa xmm11, xmm8
movdqa xmm15, xmm12
paddd xmm0, xmm1
paddd xmm4, xmm5
paddd xmm8, xmm9
paddd xmm12, xmm13
movdqa xmm1, xmm0
movdqa xmm5, xmm4
movdqa xmm9, xmm8
movdqa xmm13, xmm12
pslld xmm0, 13
pslld xmm4, 13
pslld xmm8, 13
pslld xmm12, 13
psrld xmm1, 32-13
psrld xmm5, 32-13
psrld xmm9, 32-13
psrld xmm13, 32-13
pxor xmm0, [rsp + 12*16 + 1*256]
pxor xmm4, [rsp + 13*16 + 1*256]
pxor xmm8, [rsp + 14*16 + 1*256]
pxor xmm12, [rsp + 15*16 + 1*256]
pxor xmm0, xmm1
pxor xmm4, xmm5
pxor xmm8, xmm9
pxor xmm12, xmm13
movdqa [rsp + 12*16], xmm0
movdqa [rsp + 13*16], xmm4
movdqa [rsp + 14*16], xmm8
movdqa [rsp + 15*16], xmm12
paddd xmm0, xmm3
paddd xmm4, xmm7
paddd xmm8, xmm11
paddd xmm12, xmm15
movdqa xmm3, xmm0
movdqa xmm7, xmm4
movdqa xmm11, xmm8
movdqa xmm15, xmm12
pslld xmm0, 18
pslld xmm4, 18
pslld xmm8, 18
pslld xmm12, 18
psrld xmm3, 32-18
psrld xmm7, 32-18
psrld xmm11, 32-18
psrld xmm15, 32-18
pxor xmm0, xmm2
pxor xmm4, xmm6
pxor xmm8, xmm10
pxor xmm12, xmm14
pxor xmm0, xmm3
pxor xmm4, xmm7
pxor xmm8, xmm11
pxor xmm12, xmm15
movdqa [rsp + 0*16], xmm0
movdqa [rsp + 1*16], xmm4
movdqa [rsp + 2*16], xmm8
movdqa [rsp + 3*16], xmm12
mov rax, r9
jmp label2
labelSSE2_Salsa_Output:
movdqa xmm0, xmm4
punpckldq xmm4, xmm5
movdqa xmm1, xmm6
punpckldq xmm6, xmm7
movdqa xmm2, xmm4
punpcklqdq xmm4, xmm6
punpckhqdq xmm2, xmm6
punpckhdq xmm0, xmm5
punpckhdq xmm1, xmm7
movdqa xmm6, xmm0
punpcklqdq xmm0, xmm1
punpckhqdq xmm6, xmm1
test rdx, rdx
jz labelSSE2_Salsa_Output_A3
test rdx, 15
jnz labelSSE2_Salsa_Output_A7
pxor xmm4, [rdx+0*16]
pxor xmm2, [rdx+4*16]
pxor xmm0, [rdx+8*16]
pxor xmm6, [rdx+12*16]
add rdx, 1*16
jmp labelSSE2_Salsa_Output_A3
labelSSE2_Salsa_Output_A7:
movdqu xmm1, [rdx+0*16]
pxor xmm4, xmm1
movdqu xmm1, [rdx+4*16]
pxor xmm2, xmm1
movdqu xmm1, [rdx+8*16]
pxor xmm0, xmm1
movdqu xmm1, [rdx+12*16]
pxor xmm6, xmm1
add rdx, 1*16
labelSSE2_Salsa_Output_A3:
test rcx, 15
jnz labelSSE2_Salsa_Output_A8
movdqa [rcx+0*16], xmm4
movdqa [rcx+4*16], xmm2
movdqa [rcx+8*16], xmm0
movdqa [rcx+12*16], xmm6
jmp labelSSE2_Salsa_Output_A9
labelSSE2_Salsa_Output_A8:
movdqu [rcx+0*16], xmm4
movdqu [rcx+4*16], xmm2
movdqu [rcx+8*16], xmm0
movdqu [rcx+12*16], xmm6
labelSSE2_Salsa_Output_A9:
add rcx, 1*16
ret
label6:
movdqa xmm0, [rsp + 12*16 + 0*256]
movdqa xmm4, [rsp + 13*16 + 0*256]
movdqa xmm8, [rsp + 14*16 + 0*256]
movdqa xmm12, [rsp + 15*16 + 0*256]
movdqa xmm2, [rsp + 0*16 + 0*256]
movdqa xmm6, [rsp + 1*16 + 0*256]
movdqa xmm10, [rsp + 2*16 + 0*256]
movdqa xmm14, [rsp + 3*16 + 0*256]
paddd xmm0, xmm2
paddd xmm4, xmm6
paddd xmm8, xmm10
paddd xmm12, xmm14
movdqa xmm1, xmm0
movdqa xmm5, xmm4
movdqa xmm9, xmm8
movdqa xmm13, xmm12
pslld xmm0, 7
pslld xmm4, 7
pslld xmm8, 7
pslld xmm12, 7
psrld xmm1, 32-7
psrld xmm5, 32-7
psrld xmm9, 32-7
psrld xmm13, 32-7
pxor xmm0, [rsp + 4*16 + 0*256]
pxor xmm4, [rsp + 5*16 + 0*256]
pxor xmm8, [rsp + 6*16 + 0*256]
pxor xmm12, [rsp + 7*16 + 0*256]
pxor xmm0, xmm1
pxor xmm4, xmm5
pxor xmm8, xmm9
pxor xmm12, xmm13
movdqa [rsp + 4*16], xmm0
movdqa [rsp + 5*16], xmm4
movdqa [rsp + 6*16], xmm8
movdqa [rsp + 7*16], xmm12
movdqa xmm1, xmm0
movdqa xmm5, xmm4
movdqa xmm9, xmm8
movdqa xmm13, xmm12
paddd xmm0, xmm2
paddd xmm4, xmm6
paddd xmm8, xmm10
paddd xmm12, xmm14
movdqa xmm3, xmm0
movdqa xmm7, xmm4
movdqa xmm11, xmm8
movdqa xmm15, xmm12
pslld xmm0, 9
pslld xmm4, 9
pslld xmm8, 9
pslld xmm12, 9
psrld xmm3, 32-9
psrld xmm7, 32-9
psrld xmm11, 32-9
psrld xmm15, 32-9
pxor xmm0, [rsp + 8*16 + 0*256]
pxor xmm4, [rsp + 9*16 + 0*256]
pxor xmm8, [rsp + 10*16 + 0*256]
pxor xmm12, [rsp + 11*16 + 0*256]
pxor xmm0, xmm3
pxor xmm4, xmm7
pxor xmm8, xmm11
pxor xmm12, xmm15
movdqa [rsp + 8*16], xmm0
movdqa [rsp + 9*16], xmm4
movdqa [rsp + 10*16], xmm8
movdqa [rsp + 11*16], xmm12
movdqa xmm3, xmm0
movdqa xmm7, xmm4
movdqa xmm11, xmm8
movdqa xmm15, xmm12
paddd xmm0, xmm1
paddd xmm4, xmm5
paddd xmm8, xmm9
paddd xmm12, xmm13
movdqa xmm1, xmm0
movdqa xmm5, xmm4
movdqa xmm9, xmm8
movdqa xmm13, xmm12
pslld xmm0, 13
pslld xmm4, 13
pslld xmm8, 13
pslld xmm12, 13
psrld xmm1, 32-13
psrld xmm5, 32-13
psrld xmm9, 32-13
psrld xmm13, 32-13
pxor xmm0, [rsp + 12*16 + 0*256]
pxor xmm4, [rsp + 13*16 + 0*256]
pxor xmm8, [rsp + 14*16 + 0*256]
pxor xmm12, [rsp + 15*16 + 0*256]
pxor xmm0, xmm1
pxor xmm4, xmm5
pxor xmm8, xmm9
pxor xmm12, xmm13
movdqa [rsp + 12*16], xmm0
movdqa [rsp + 13*16], xmm4
movdqa [rsp + 14*16], xmm8
movdqa [rsp + 15*16], xmm12
paddd xmm0, xmm3
paddd xmm4, xmm7
paddd xmm8, xmm11
paddd xmm12, xmm15
movdqa xmm3, xmm0
movdqa xmm7, xmm4
movdqa xmm11, xmm8
movdqa xmm15, xmm12
pslld xmm0, 18
pslld xmm4, 18
pslld xmm8, 18
pslld xmm12, 18
psrld xmm3, 32-18
psrld xmm7, 32-18
psrld xmm11, 32-18
psrld xmm15, 32-18
pxor xmm0, xmm2
pxor xmm4, xmm6
pxor xmm8, xmm10
pxor xmm12, xmm14
pxor xmm0, xmm3
pxor xmm4, xmm7
pxor xmm8, xmm11
pxor xmm12, xmm15
movdqa [rsp + 0*16], xmm0
movdqa [rsp + 1*16], xmm4
movdqa [rsp + 2*16], xmm8
movdqa [rsp + 3*16], xmm12
label2:
movdqa xmm0, [rsp + 7*16 + 0*256]
movdqa xmm4, [rsp + 4*16 + 0*256]
movdqa xmm8, [rsp + 5*16 + 0*256]
movdqa xmm12, [rsp + 6*16 + 0*256]
movdqa xmm2, [rsp + 0*16 + 0*256]
movdqa xmm6, [rsp + 1*16 + 0*256]
movdqa xmm10, [rsp + 2*16 + 0*256]
movdqa xmm14, [rsp + 3*16 + 0*256]
paddd xmm0, xmm2
paddd xmm4, xmm6
paddd xmm8, xmm10
paddd xmm12, xmm14
movdqa xmm1, xmm0
movdqa xmm5, xmm4
movdqa xmm9, xmm8
movdqa xmm13, xmm12
pslld xmm0, 7
pslld xmm4, 7
pslld xmm8, 7
pslld xmm12, 7
psrld xmm1, 32-7
psrld xmm5, 32-7
psrld xmm9, 32-7
psrld xmm13, 32-7
pxor xmm0, [rsp + 13*16 + 0*256]
pxor xmm4, [rsp + 14*16 + 0*256]
pxor xmm8, [rsp + 15*16 + 0*256]
pxor xmm12, [rsp + 12*16 + 0*256]
pxor xmm0, xmm1
pxor xmm4, xmm5
pxor xmm8, xmm9
pxor xmm12, xmm13
movdqa [rsp + 13*16], xmm0
movdqa [rsp + 14*16], xmm4
movdqa [rsp + 15*16], xmm8
movdqa [rsp + 12*16], xmm12
movdqa xmm1, xmm0
movdqa xmm5, xmm4
movdqa xmm9, xmm8
movdqa xmm13, xmm12
paddd xmm0, xmm2
paddd xmm4, xmm6
paddd xmm8, xmm10
paddd xmm12, xmm14
movdqa xmm3, xmm0
movdqa xmm7, xmm4
movdqa xmm11, xmm8
movdqa xmm15, xmm12
pslld xmm0, 9
pslld xmm4, 9
pslld xmm8, 9
pslld xmm12, 9
psrld xmm3, 32-9
psrld xmm7, 32-9
psrld xmm11, 32-9
psrld xmm15, 32-9
pxor xmm0, [rsp + 10*16 + 0*256]
pxor xmm4, [rsp + 11*16 + 0*256]
pxor xmm8, [rsp + 8*16 + 0*256]
pxor xmm12, [rsp + 9*16 + 0*256]
pxor xmm0, xmm3
pxor xmm4, xmm7
pxor xmm8, xmm11
pxor xmm12, xmm15
movdqa [rsp + 10*16], xmm0
movdqa [rsp + 11*16], xmm4
movdqa [rsp + 8*16], xmm8
movdqa [rsp + 9*16], xmm12
movdqa xmm3, xmm0
movdqa xmm7, xmm4
movdqa xmm11, xmm8
movdqa xmm15, xmm12
paddd xmm0, xmm1
paddd xmm4, xmm5
paddd xmm8, xmm9
paddd xmm12, xmm13
movdqa xmm1, xmm0
movdqa xmm5, xmm4
movdqa xmm9, xmm8
movdqa xmm13, xmm12
pslld xmm0, 13
pslld xmm4, 13
pslld xmm8, 13
pslld xmm12, 13
psrld xmm1, 32-13
psrld xmm5, 32-13
psrld xmm9, 32-13
psrld xmm13, 32-13
pxor xmm0, [rsp + 7*16 + 0*256]
pxor xmm4, [rsp + 4*16 + 0*256]
pxor xmm8, [rsp + 5*16 + 0*256]
pxor xmm12, [rsp + 6*16 + 0*256]
pxor xmm0, xmm1
pxor xmm4, xmm5
pxor xmm8, xmm9
pxor xmm12, xmm13
movdqa [rsp + 7*16], xmm0
movdqa [rsp + 4*16], xmm4
movdqa [rsp + 5*16], xmm8
movdqa [rsp + 6*16], xmm12
paddd xmm0, xmm3
paddd xmm4, xmm7
paddd xmm8, xmm11
paddd xmm12, xmm15
movdqa xmm3, xmm0
movdqa xmm7, xmm4
movdqa xmm11, xmm8
movdqa xmm15, xmm12
pslld xmm0, 18
pslld xmm4, 18
pslld xmm8, 18
pslld xmm12, 18
psrld xmm3, 32-18
psrld xmm7, 32-18
psrld xmm11, 32-18
psrld xmm15, 32-18
pxor xmm0, xmm2
pxor xmm4, xmm6
pxor xmm8, xmm10
pxor xmm12, xmm14
pxor xmm0, xmm3
pxor xmm4, xmm7
pxor xmm8, xmm11
pxor xmm12, xmm15
movdqa [rsp + 0*16], xmm0
movdqa [rsp + 1*16], xmm4
movdqa [rsp + 2*16], xmm8
movdqa [rsp + 3*16], xmm12
sub eax, 2
jnz label6
movdqa xmm4, [rsp + 0*16 + 256]
paddd xmm4, [rsp + 0*16]
movdqa xmm5, [rsp + 13*16 + 256]
paddd xmm5, [rsp + 13*16]
movdqa xmm6, [rsp + 10*16 + 256]
paddd xmm6, [rsp + 10*16]
movdqa xmm7, [rsp + 7*16 + 256]
paddd xmm7, [rsp + 7*16]
call labelSSE2_Salsa_Output
movdqa xmm4, [rsp + 4*16 + 256]
paddd xmm4, [rsp + 4*16]
movdqa xmm5, [rsp + 1*16 + 256]
paddd xmm5, [rsp + 1*16]
movdqa xmm6, [rsp + 14*16 + 256]
paddd xmm6, [rsp + 14*16]
movdqa xmm7, [rsp + 11*16 + 256]
paddd xmm7, [rsp + 11*16]
call labelSSE2_Salsa_Output
movdqa xmm4, [rsp + 8*16 + 256]
paddd xmm4, [rsp + 8*16]
movdqa xmm5, [rsp + 5*16 + 256]
paddd xmm5, [rsp + 5*16]
movdqa xmm6, [rsp + 2*16 + 256]
paddd xmm6, [rsp + 2*16]
movdqa xmm7, [rsp + 15*16 + 256]
paddd xmm7, [rsp + 15*16]
call labelSSE2_Salsa_Output
movdqa xmm4, [rsp + 12*16 + 256]
paddd xmm4, [rsp + 12*16]
movdqa xmm5, [rsp + 9*16 + 256]
paddd xmm5, [rsp + 9*16]
movdqa xmm6, [rsp + 6*16 + 256]
paddd xmm6, [rsp + 6*16]
movdqa xmm7, [rsp + 3*16 + 256]
paddd xmm7, [rsp + 3*16]
call labelSSE2_Salsa_Output
test rdx, rdx
jz label9
add rdx, 12*16
label9:
add rcx, 12*16
sub r8, 4
cmp r8, 4
jge label1
label5:
sub r8, 1
jl label4
movdqa xmm0, [r10 + 0*16]
movdqa xmm1, [r10 + 1*16]
movdqa xmm2, [r10 + 2*16]
movdqa xmm3, [r10 + 3*16]
mov rax, r9
label0:
movdqa xmm4, xmm3
paddd xmm4, xmm0
movdqa xmm5, xmm4
pslld xmm4, 7
psrld xmm5, 32-7
pxor xmm1, xmm4
pxor xmm1, xmm5
movdqa xmm4, xmm0
paddd xmm4, xmm1
movdqa xmm5, xmm4
pslld xmm4, 9
psrld xmm5, 32-9
pxor xmm2, xmm4
pxor xmm2, xmm5