chacha/asm/chacha-x86_64.pl

570 my ($xc,$xc_,$t0,$t1)=map("\"$_\"",$xt0,$xt1,$xt2,$xt3);
607 	"&movdqa	($t0,@x[$b0])",
609 	"&psrld		($t0,20)",
612 	"&por		(@x[$b0],$t0)",
614 	"&movdqa	($t0,'(%r11)')",	# .Lrot24(%rip)
621 	"&pshufb	(@x[$d0],$t0)",
622 	 "&pshufb	(@x[$d1],$t0)",
631 	 "&movdqa	($t0,@x[$b1])",
634 	 "&psrld	($t0,25)",
636 	 "&por		(@x[$b1],$t0)",
654 	"&movdqa	($t0,@x[$b2])",
656 	"&psrld		($t0,20)",
659 	"&por		(@x[$b2],$t0)",
661 	"&movdqa	($t0,'(%r11)')",	# .Lrot24(%rip)
668 	"&pshufb	(@x[$d2],$t0)",
669 	 "&pshufb	(@x[$d3],$t0)",
678 	 "&movdqa	($t0,@x[$b3])",
681 	 "&psrld	($t0,25)",
683 	 "&por		(@x[$b3],$t0)"
1153 my ($xc,$xc_,$t0,$t1)=map("\"$_\"",$xt0,$xt1,$xt2,$xt3);
1188 	"&vpslld	($t0,@x[$b0],12)",
1190 	"&vpor		(@x[$b0],$t0,@x[$b0])",
1191 	"&vbroadcasti128($t0,'(%r11)')",		# .Lrot24(%rip)
1200 	"&vpshufb	(@x[$d0],@x[$d0],$t0)",
1203 	 "&vpshufb	(@x[$d1],@x[$d1],$t0)",
1213 	 "&vpslld	($t0,@x[$b1],7)",
1215 	 "&vpor		(@x[$b1],$t0,@x[$b1])",
1231 	"&vpslld	($t0,@x[$b2],12)",
1233 	"&vpor		(@x[$b2],$t0,@x[$b2])",
1234 	"&vbroadcasti128($t0,'(%r11)')",		# .Lrot24(%rip)
1243 	"&vpshufb	(@x[$d2],@x[$d2],$t0)",
1246 	 "&vpshufb	(@x[$d3],@x[$d3],$t0)",
1256 	 "&vpslld	($t0,@x[$b3],7)",
1258 	 "&vpor		(@x[$b3],$t0,@x[$b3])"
1787 my ($t0,$t1,$t2,$t3) = map("%xmm$_",(4..7));
1875 	vpxor		0x00($inp),%x#$a,$t0	# xor with input
1881 	vmovdqu		$t0,0x00($out)		# write output
1889 	vextracti32x4	\$1,$a,$t0
1897 	vpxor		0x00($inp),$t0,$t0	# xor with input
1903 	vmovdqu		$t0,0x00($out)		# write output
1911 	vextracti32x4	\$2,$a,$t0
1919 	vpxor		0x00($inp),$t0,$t0	# xor with input
1925 	vmovdqu		$t0,0x00($out)		# write output
1933 	vextracti32x4	\$3,$a,$t0
1941 	vpxor		0x00($inp),$t0,$t0	# xor with input
1947 	vmovdqu		$t0,0x00($out)		# write output
1968 	vmovdqa		$t0,0x00(%rsp)