modes/asm/ghash-x86.pl

253 	&mov	($inp,&wparam(0));		# load Xi
256 	&mov	($Zhh,&DWP(0,$inp));		# load Xi[16]
263 	&mov	(&DWP(0,"esp"),$Zhh);		# copy Xi[16] on stack
286 	&mov	($Zll,&wparam(0));		# load Xi
293 	&mov	($Zhh,&DWP(0,$Zll));		# load Xi[16]
324 	&mov	($inp,&wparam(0));	# load Xi
414 	&mov	($inp,&wparam(0));	# load Xi
424 	&mov	($inp,&wparam(0));	# load Xi
436 	&mov	($Zhh,&wparam(0));	# load Xi
448 	&mov	($Zll,&DWP(12,$Zhh));	# load Xi[16]
475 	&mov	($inp,&wparam(0));	# load Xi
585 	&mov	($inp,&wparam(0));	# load Xi
616     &mov	("eax",&wparam(0));		# Xi
630     &mov	(&DWP(528+16+0,"esp"),"eax");	# save Xi
672     &mov	("edx",&DWP(12,"eax"));		# load Xi
686     #&mov	(&DWP(528+12,"esp"),$dat);	# save inp^Xi
796     &mov	("eax",&DWP(528+16+0,"esp"));	# restore Xi
818 ($Xi,$Xhi)=("xmm0","xmm1");	$Hkey="xmm2";
825 my ($Xhi,$Xi,$Hkey)=@_;
827 	&movdqa		($Xhi,$Xi);		#
828 	&pshufd		($T1,$Xi,0b01001110);
830 	&pxor		($T1,$Xi);		#
833 	&pclmulqdq	($Xi,$Hkey,0x00);	#######
836 	&xorps		($T1,$Xi);		#
843 	&pxor		($Xi,$T2);		#
852 my ($Xhi,$Xi,$Hkey)=@_;
854 	&movdqa		($T1,$Xi);		#
855 	&movdqa		($Xhi,$Xi);
856 	&pclmulqdq	($Xi,$Hkey,0x00);	#######
863 	&pxor		($T2,$Xi);		#
870 	&pxor		($Xi,$T3);		#
884 my ($Xhi,$Xi) = @_;
887 	&movdqa		($T1,$Xi);		#
888 	&psllq		($Xi,1);
889 	&pxor		($Xi,$T1);		#
890 	&psllq		($Xi,5);		#
891 	&pxor		($Xi,$T1);		#
892 	&psllq		($Xi,57);		#
893 	&movdqa		($T2,$Xi);		#
894 	&pslldq		($Xi,8);
896 	&pxor		($Xi,$T1);
900 	&movdqa		($T2,$Xi);
901 	&psrlq		($Xi,5);
902 	&pxor		($Xi,$T2);		#
903 	&psrlq		($Xi,1);		#
904 	&pxor		($Xi,$T2);		#
906 	&psrlq		($Xi,1);		#
907 	&pxor		($Xi,$T2);		#
935 	&movdqa		($Xi,$Hkey);
936 	&clmul64x64_T2	($Xhi,$Xi,$Hkey);
937 	&reduction_alg9	($Xhi,$Xi);
940 	&movdqu		(&QWP(16,$Htbl),$Xi);	# save H^2
952 	&movdqu		($Xi,&QWP(0,$Xip));
955 	&pshufb		($Xi,$T3);
957 	&clmul64x64_T2	($Xhi,$Xi,$Hkey);
958 	&reduction_alg9	($Xhi,$Xi);
960 	&pshufb		($Xi,$T3);
961 	&movdqu		(&QWP(0,$Xip),$Xi);
975 	&movdqu		($Xi,&QWP(0,$Xip));
978 	&pshufb		($Xi,$T3);
984 	# Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
985 	#	[(H*Ii+1) + (H*Xi+1)] mod P =
986 	#	[(H*Ii+1) + H^2*(Ii+Xi)] mod P
992 	&pxor		($Xi,$T1);		# Ii+Xi
1002 	&clmul64x64_T2	($Xhi,$Xi,$Hkey);	# H^2*(Ii+Xi)
1006 	&pxor		($Xi,$Xn);		# (H*Ii+1) + H^2*(Ii+Xi)
1015 	 &pxor		($Xhi,$T1);		# "Ii+Xi", consume early
1017 	  &movdqa	($T1,$Xi);		#&reduction_alg9($Xhi,$Xi); 1st phase
1018 	  &psllq	($Xi,1);
1019 	  &pxor		($Xi,$T1);		#
1020 	  &psllq	($Xi,5);		#
1021 	  &pxor		($Xi,$T1);		#
1023 	  &psllq	($Xi,57);		#
1024 	  &movdqa	($T2,$Xi);		#
1025 	  &pslldq	($Xi,8);
1027 	  &pxor		($Xi,$T1);
1035 	  &movdqa	($T2,$Xi);		# 2nd phase
1036 	  &psrlq	($Xi,5);
1037 	  &pxor		($Xi,$T2);		#
1038 	  &psrlq	($Xi,1);		#
1039 	  &pxor		($Xi,$T2);		#
1041 	  &psrlq	($Xi,1);		#
1042 	  &pxor		($Xi,$T2);		#
1061 	&clmul64x64_T2	($Xhi,$Xi,$Hkey);	# H^2*(Ii+Xi)
1063 	&pxor		($Xi,$Xn);		# (H*Ii+1) + H^2*(Ii+Xi)
1066 	&reduction_alg9	($Xhi,$Xi);
1075 	&pxor		($Xi,$T1);		# Ii+Xi
1077 	&clmul64x64_T2	($Xhi,$Xi,$Hkey);	# H*(Ii+Xi)
1078 	&reduction_alg9	($Xhi,$Xi);
1081 	&pshufb		($Xi,$T3);
1082 	&movdqu		(&QWP(0,$Xip),$Xi);
1089 my ($Xhi,$Xi)=@_;
1092 	&movdqa		($T1,$Xi);		#
1094 	&pslld		($Xi,1);
1103 	&por		($Xi,$T1);
1107 	&movdqa		($T1,$Xi);
1108 	&movdqa		($T2,$Xi);
1109 	&movdqa		($T3,$Xi);		#
1112 	&pslld		($Xi,25);		#
1114 	&pxor		($T1,$Xi);		#
1122 	&movdqa		($Xi,$T3);
1124 	&psrld		($Xi,1);		#
1127 	&pxor		($Xi,$T1);
1129 	&pxor		($Xi,$T3);		#
1130 	&pxor		($Xi,$Xhi);		#
1144 	&movdqa		($Xi,$Hkey);
1145 	&clmul64x64_T3	($Xhi,$Xi,$Hkey);
1146 	&reduction_alg5	($Xhi,$Xi);
1149 	&movdqu		(&QWP(16,$Htbl),$Xi);	# save H^2
1161 	&movdqu		($Xi,&QWP(0,$Xip));
1164 	&pshufb		($Xi,$Xn);
1166 	&clmul64x64_T3	($Xhi,$Xi,$Hkey);
1167 	&reduction_alg5	($Xhi,$Xi);
1169 	&pshufb		($Xi,$Xn);
1170 	&movdqu		(&QWP(0,$Xip),$Xi);
1184 	&movdqu		($Xi,&QWP(0,$Xip));
1187 	&pshufb		($Xi,$T3);
1193 	# Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
1194 	#	[(H*Ii+1) + (H*Xi+1)] mod P =
1195 	#	[(H*Ii+1) + H^2*(Ii+Xi)] mod P
1201 	&pxor		($Xi,$T1);		# Ii+Xi
1211 	&clmul64x64_T3	($Xhi,$Xi,$Hkey);	# H^2*(Ii+Xi)
1214 	&pxor		($Xi,$Xn);		# (H*Ii+1) + H^2*(Ii+Xi)
1217 	&reduction_alg5	($Xhi,$Xi);
1225 	&pxor		($Xi,$T1);		# Ii+Xi
1235 	&clmul64x64_T3	($Xhi,$Xi,$Hkey);	# H^2*(Ii+Xi)
1237 	&pxor		($Xi,$Xn);		# (H*Ii+1) + H^2*(Ii+Xi)
1240 	&reduction_alg5	($Xhi,$Xi);
1250 	&pxor		($Xi,$T1);		# Ii+Xi
1252 	&clmul64x64_T3	($Xhi,$Xi,$Hkey);	# H*(Ii+Xi)
1253 	&reduction_alg5	($Xhi,$Xi);
1257 	&pshufb		($Xi,$T3);
1258 	&movdqu		(&QWP(0,$Xip),$Xi);