sha/asm/sha512-x86_64.pl

751 my ($t0,$t1,$t2,$t3, $t4,$t5) = map("%xmm$_",(4..9));
807 	movdqa	0x00($Tbl),$t0
810 	paddd	@X[0],$t0
817 	movdqa	$t0,0x00(%rsp)
833 	'&movdqa	($t0,@X[1]);',
835 	'&palignr	($t0,@X[0],$SZ)',	# X[1..4]
837 	'&movdqa	($t1,$t0)',
838 	'&movdqa	($t2,$t0);',
839 	'&psrld		($t0,$sigma0[2])',
844 	'&pxor		($t0,$t2)',
846 	'&pxor		($t0,$t1)',
848 	'&pxor		($t0,$t2);',
850 	'&pxor		($t0,$t1);',		# sigma0(X[1..4])
852 	'&paddd		(@X[0],$t0);',		# X[0..3] += sigma0(X[1..4])
887 	&movdqa		($t0,@X[1]);
896 	&palignr	($t0,@X[0],$SZ);	# X[1..4]
904 	&movdqa		($t1,$t0);
907 	&movdqa		($t2,$t0);
910 	&psrld		($t0,$sigma0[2]);
926 	&pxor		($t0,$t2);
933 	&pxor		($t0,$t1);
939 	&pxor		($t0,$t2);
945 	&pxor		($t0,$t1);		# sigma0(X[1..4])
952 	&paddd		(@X[0],$t0);		# X[0..3] += sigma0(X[1..4])
1144     my ($t0,$t1,$t2,$t3) = map("%xmm$_",(4..7));
1158 	vpaddd	0x00($Tbl),@X[0],$t0
1163 	vmovdqa	$t0,0x00(%rsp)
1183 	&vpalignr	($t0,@X[1],@X[0],$SZ);	# X[1..4]
1189 	&vprotd		($t1,$t0,8*$SZ-$sigma0[1]);
1192 	&vpsrld		($t0,$t0,$sigma0[2]);
1203 	&vpxor		($t0,$t0,$t1);
1211 	&vpxor		($t0,$t0,$t2);		# sigma0(X[1..4])
1217 	&vpaddd		(@X[0],@X[0],$t0);	# X[0..3] += sigma0(X[1..4])
1290     my ($t0,$t1,$t2,$t3) = map("%xmm$_",(8..11));
1311 	vpaddq	-0x80($Tbl),@X[0],$t0
1317 	vmovdqa	$t0,0x00(%rsp)
1318 	vpaddq	0x00($Tbl),@X[4],$t0
1325 	vmovdqa	$t0,0x40(%rsp)
1345 	&vpalignr	($t0,@X[1],@X[0],$SZ);	# X[1..2]
1351 	&vprotq		($t1,$t0,8*$SZ-$sigma0[1]);
1354 	&vpsrlq		($t0,$t0,$sigma0[2]);
1365 	&vpxor		($t0,$t0,$t1);
1373 	&vpxor		($t0,$t0,$t2);		# sigma0(X[1..2])
1379 	&vpaddq		(@X[0],@X[0],$t0);	# X[0..1] += sigma0(X[1..2])
1519     my ($t0,$t1,$t2,$t3, $t4,$t5) = map("%xmm$_",(4..9));
1536 	vpaddd	0x00($Tbl),@X[0],$t0
1541 	vmovdqa	$t0,0x00(%rsp)
1557 	'&vpalignr	($t0,@X[1],@X[0],$SZ)',	# X[1..4]
1559 	'&vpsrld	($t2,$t0,$sigma0[0]);',
1561 	'&vpsrld	($t3,$t0,$sigma0[2])',
1562 	'&vpslld	($t1,$t0,8*$SZ-$sigma0[1]);',
1563 	'&vpxor		($t0,$t3,$t2)',
1566 	'&vpxor		($t0,$t0,$t1)',
1568 	'&vpxor		($t0,$t0,$t2)',
1570 	'&vpxor		($t0,$t0,$t1)',		# sigma0(X[1..4])
1572 	'&vpaddd	(@X[0],@X[0],$t0)',	# X[0..3] += sigma0(X[1..4])
1619     my ($t0,$t1,$t2,$t3) = map("%xmm$_",(8..11));
1641 	vpaddq	-0x80($Tbl),@X[0],$t0
1647 	vmovdqa	$t0,0x00(%rsp)
1648 	vpaddq	0x00($Tbl),@X[4],$t0
1655 	vmovdqa	$t0,0x40(%rsp)
1671 	'&vpalignr	($t0,@X[1],@X[0],$SZ)',	# X[1..2]
1673 	'&vpsrlq	($t2,$t0,$sigma0[0])',
1675 	'&vpsrlq	($t3,$t0,$sigma0[2])',
1676 	'&vpsllq	($t1,$t0,8*$SZ-$sigma0[1]);',
1677 	 '&vpxor	($t0,$t3,$t2)',
1679 	 '&vpxor	($t0,$t0,$t1)',
1681 	 '&vpxor	($t0,$t0,$t2)',
1683 	'&vpxor		($t0,$t0,$t1)',		# sigma0(X[1..2])
1685 	'&vpaddq	(@X[0],@X[0],$t0)',	# X[0..1] += sigma0(X[1..2])
1875     my ($t0,$t1,$t2,$t3, $t4,$t5) = map("%ymm$_",(4..9));
1898 	vpaddd	0x00($Tbl),@X[0],$t0
1903 	vmovdqa	$t0,0x00(%rsp)
1952     my ($t0,$t1,$t2,$t3) = map("%ymm$_",(8..11));
1984 	vpaddq	-0x80($Tbl),@X[0],$t0
1990 	vmovdqa	$t0,0x00(%rsp)
1991 	vpaddq	0x00($Tbl),@X[4],$t0
1999 	vmovdqa	$t0,0x00(%rsp)