modes/asm/ghash-x86.pl

838 ($T1,$T2,$T3)=("xmm3","xmm4","xmm5");
847 	&pshufd		($T1,$Xi,0b01001110);
849 	&pxor		($T1,$Xi);		#
855 	&pclmulqdq	($T1,$HK,0x00);		#######
856 	&xorps		($T1,$Xi);		#
857 	&xorps		($T1,$Xhi);		#
859 	&movdqa		($T2,$T1);		#
860 	&psrldq		($T1,8);
862 	&pxor		($Xhi,$T1);
874 	&movdqa		($T1,$Xi);		#
878 	&pshufd		($T2,$T1,0b01001110);	#
880 	&pxor		($T2,$T1);		#
908 	&movdqa		($T1,$Xi);
910 	&pxor		($T1,$Xi);		#
912 	&pxor		($Xi,$T1);		#
914 	&movdqa		($T1,$Xi);		#
916 	&psrldq		($T1,8);		#
918 	&pxor		($Xhi,$T1);		#
945 	&movdqa		($T1,$Hkey);
948 	&psrlq		($T1,63);
950 	&pslldq		($T1,8);
951 	&por		($Hkey,$T1);		# H<<=1
962 	&pshufd		($T1,$Hkey,0b01001110);
964 	&pxor		($T1,$Hkey);		# Karatsuba pre-processing
968 	&palignr	($T2,$T1,8);		# low part is H.lo^H.hi
1022 	&movdqu		($T1,&QWP(0,$inp));	# Ii
1024 	&pshufb		($T1,$T3);
1027 	&pxor		($Xi,$T1);		# Ii+Xi
1029 	&pshufd		($T1,$Xn,0b01001110);	# H*Ii+1
1031 	&pxor		($T1,$Xn);		#
1036 	&pclmulqdq	($T1,$T3,0x00);		#######
1059 	&pxor		($T1,$Xi);		# aggregated Karatsuba post-processing
1061 	&pxor		($T1,$Xhi);		#
1064 	&pxor		($T2,$T1);		#
1066 	&movdqa		($T1,$T2);		#
1068 	&pslldq		($T1,8);		#
1070 	&pxor		($Xi,$T1);		#
1076 	  &movdqa	($T1,$Xi);
1078 	  &pxor		($T1,$Xi);		#
1080 	  &pxor		($Xi,$T1);		#
1084 	  &movdqa	($T1,$Xi);		#
1086 	  &psrldq	($T1,8);		#
1088 	  &pxor		($Xhi,$T1);		#
1089 	&pshufd		($T1,$Xhn,0b01001110);
1092 	&pxor		($T1,$Xhn);
1101 	&pclmulqdq	($T1,$T3,0x00);		#######
1119 	&pxor		($T1,$Xi);		# aggregated Karatsuba post-processing
1120 	&pxor		($T1,$Xhi);		#
1122 	&pxor		($T2,$T1);		#
1124 	&movdqa		($T1,$T2);		#
1126 	&pslldq		($T1,8);		#
1128 	&pxor		($Xi,$T1);		#
1137 	&movdqu		($T1,&QWP(0,$inp));	# Ii
1138 	&pshufb		($T1,$T3);
1139 	&pxor		($Xi,$T1);		# Ii+Xi
1156 	&movdqa		($T1,$Xi);		#
1160 	&psrld		($T1,31);
1162 	&movdqa		($T3,$T1);
1163 	&pslldq		($T1,4);
1167 	&por		($Xi,$T1);
1171 	&movdqa		($T1,$Xi);
1174 	&pslld		($T1,31);
1177 	&pxor		($T1,$T2);
1178 	&pxor		($T1,$Xi);		#
1179 	&movdqa		($T2,$T1);		#
1180 	&pslldq		($T1,12);
1182 	&pxor		($T3,$T1);
1187 	&movdqa		($T1,$T3);
1189 	&psrld		($T1,2);
1191 	&pxor		($Xi,$T1);
1267 	&movdqu		($T1,&QWP(0,$inp));	# Ii
1269 	&pshufb		($T1,$T3);
1271 	&pxor		($Xi,$T1);		# Ii+Xi
1291 	&movdqu		($T1,&QWP(0,$inp));	# Ii
1293 	&pshufb		($T1,$T3);
1295 	&pxor		($Xi,$T1);		# Ii+Xi
1318 	&movdqu		($T1,&QWP(0,$inp));	# Ii
1319 	&pshufb		($T1,$T3);
1320 	&pxor		($Xi,$T1);		# Ii+Xi