sha/asm/sha1-586.pl

409 my $Xi=4;			# 4xSIMD Xupdate round, start pre-seeded
410 my @X=map("xmm$_",(4..7,0..3));	# pre-seeded for $Xi=4
518 sub Xupdate_ssse3_16_31()		# recall that $Xi starts with 4
532 	  &movdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
551 	  &movdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
580 	  &movdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5);	# restore X[] from backtrace buffer
588 	  &movdqa	(@X[4],&QWP(112-16+16*(($Xi)/5),"esp"));	# K_XX_XX
593 	  &movdqa	(@X[1],@X[-2&7])	if ($Xi<7);
599   $Xi++;	push(@X,shift(@X));	# "rotate" X[]
608 	&movdqa	(@X[2],@X[-1&7])	if ($Xi==8);
617 	  &movdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);	# save X[] to backtrace buffer
620 	 if ($Xi%5) {
623 	  &movdqa	(@X[4],&QWP(112-16+16*($Xi/5),"esp"));
636 	  &movdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
656 	  &movdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19);	# restore X[] from backtrace buffer
662 	  &movdqa	(@X[3],@X[0])	if ($Xi<19);
667   $Xi++;	push(@X,shift(@X));	# "rotate" X[]
683 	  &movdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer IALU
702   $Xi=0;
713 	&pshufb	(@X[($Xi-3)&7],@X[2]);
716 	&paddd	(@X[($Xi-4)&7],@X[3]);
721 	&movdqa	(&QWP(0+16*$Xi,"esp"),@X[($Xi-4)&7]);	# X[]+K xfer to IALU
724 	&psubd	(@X[($Xi-4)&7],@X[3]);
727   $Xi++;
850 my $Xi=4;			# 4xSIMD Xupdate round, start pre-seeded
851 my @X=map("xmm$_",(4..7,0..3));	# pre-seeded for $Xi=4
939 sub Xupdate_avx_16_31()		# recall that $Xi starts with 4
952 	  &vmovdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
965 	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
996 	  &vmovdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5);	# restore X[] from backtrace buffer
1008 	  &vmovdqa	(@X[4],&QWP(112-16+16*(($Xi)/5),"esp"));	# K_XX_XX
1014   $Xi++;	push(@X,shift(@X));	# "rotate" X[]
1031 	  &vmovdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);	# save X[] to backtrace buffer
1034 	 if ($Xi%5) {
1037 	  &vmovdqa	(@X[4],&QWP(112-16+16*($Xi/5),"esp"));
1050 	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
1069 	  &vmovdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19);	# restore X[] from backtrace buffer
1079   $Xi++;	push(@X,shift(@X));	# "rotate" X[]
1095 	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer IALU
1114   $Xi=0;
1125 	&vpshufb	(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1128 	&vpaddd	(@X[$Xi&7],@X[($Xi-4)&7],@X[3]);
1133 	&vmovdqa	(&QWP(0+16*$Xi,"esp"),@X[$Xi&7]);	# X[]+K xfer to IALU
1138   $Xi++;