Lines Matching refs:Xi
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 ($inp,&wparam(0)); # load Xi
426 &mov ($inp,&wparam(0)); # load Xi
438 &mov ($Zhh,&wparam(0)); # load Xi
452 &mov ($Zll,&DWP(12,$Zhh)); # load Xi[16]
479 &mov ($inp,&wparam(0)); # load Xi
589 &mov ($inp,&wparam(0)); # load Xi
622 &mov ("eax",&wparam(0)); # Xi
637 &mov (&DWP(528+16+0,"esp"),"eax"); # save Xi
679 &mov ("edx",&DWP(12,"eax")); # load Xi
693 #&mov (&DWP(528+12,"esp"),$dat); # save inp^Xi
803 &mov ("eax",&DWP(528+16+0,"esp")); # restore Xi
825 ($Xi,$Xhi)=("xmm0","xmm1"); $Hkey="xmm2";
832 my ($Xhi,$Xi,$Hkey)=@_;
834 &movdqa ($Xhi,$Xi); #
835 &pshufd ($T1,$Xi,0b01001110);
837 &pxor ($T1,$Xi); #
840 &pclmulqdq ($Xi,$Hkey,0x00); #######
843 &xorps ($T1,$Xi); #
850 &pxor ($Xi,$T2); #
859 my ($Xhi,$Xi,$Hkey)=@_;
861 &movdqa ($T1,$Xi); #
862 &movdqa ($Xhi,$Xi);
863 &pclmulqdq ($Xi,$Hkey,0x00); #######
870 &pxor ($T2,$Xi); #
877 &pxor ($Xi,$T3); #
891 my ($Xhi,$Xi) = @_;
894 &movdqa ($T1,$Xi) #
895 &psllq ($Xi,1);
896 &pxor ($Xi,$T1); #
897 &psllq ($Xi,5); #
898 &pxor ($Xi,$T1); #
899 &psllq ($Xi,57); #
900 &movdqa ($T2,$Xi); #
901 &pslldq ($Xi,8);
903 &pxor ($Xi,$T1);
907 &movdqa ($T2,$Xi);
908 &psrlq ($Xi,5);
909 &pxor ($Xi,$T2); #
910 &psrlq ($Xi,1); #
911 &pxor ($Xi,$T2); #
913 &psrlq ($Xi,1); #
914 &pxor ($Xi,$T2); #
944 &movdqa ($Xi,$Hkey);
945 &clmul64x64_T2 ($Xhi,$Xi,$Hkey);
946 &reduction_alg9 ($Xhi,$Xi);
949 &movdqu (&QWP(16,$Htbl),$Xi); # save H^2
963 &movdqu ($Xi,&QWP(0,$Xip));
966 &pshufb ($Xi,$T3);
968 &clmul64x64_T2 ($Xhi,$Xi,$Hkey);
969 &reduction_alg9 ($Xhi,$Xi);
971 &pshufb ($Xi,$T3);
972 &movdqu (&QWP(0,$Xip),$Xi);
988 &movdqu ($Xi,&QWP(0,$Xip));
991 &pshufb ($Xi,$T3);
997 # Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
998 # [(H*Ii+1) + (H*Xi+1)] mod P =
999 # [(H*Ii+1) + H^2*(Ii+Xi)] mod P
1005 &pxor ($Xi,$T1); # Ii+Xi
1015 &clmul64x64_T2 ($Xhi,$Xi,$Hkey); # H^2*(Ii+Xi)
1019 &pxor ($Xi,$Xn); # (H*Ii+1) + H^2*(Ii+Xi)
1028 &pxor ($Xhi,$T1); # "Ii+Xi", consume early
1030 &movdqa ($T1,$Xi) #&reduction_alg9($Xhi,$Xi); 1st phase
1031 &psllq ($Xi,1);
1032 &pxor ($Xi,$T1); #
1033 &psllq ($Xi,5); #
1034 &pxor ($Xi,$T1); #
1036 &psllq ($Xi,57); #
1037 &movdqa ($T2,$Xi); #
1038 &pslldq ($Xi,8);
1040 &pxor ($Xi,$T1);
1048 &movdqa ($T2,$Xi); # 2nd phase
1049 &psrlq ($Xi,5);
1050 &pxor ($Xi,$T2); #
1051 &psrlq ($Xi,1); #
1052 &pxor ($Xi,$T2); #
1054 &psrlq ($Xi,1); #
1055 &pxor ($Xi,$T2); #
1074 &clmul64x64_T2 ($Xhi,$Xi,$Hkey); # H^2*(Ii+Xi)
1076 &pxor ($Xi,$Xn); # (H*Ii+1) + H^2*(Ii+Xi)
1079 &reduction_alg9 ($Xhi,$Xi);
1088 &pxor ($Xi,$T1); # Ii+Xi
1090 &clmul64x64_T2 ($Xhi,$Xi,$Hkey); # H*(Ii+Xi)
1091 &reduction_alg9 ($Xhi,$Xi);
1094 &pshufb ($Xi,$T3);
1095 &movdqu (&QWP(0,$Xip),$Xi);
1102 my ($Xhi,$Xi)=@_;
1105 &movdqa ($T1,$Xi); #
1107 &pslld ($Xi,1);
1116 &por ($Xi,$T1);
1120 &movdqa ($T1,$Xi);
1121 &movdqa ($T2,$Xi);
1122 &movdqa ($T3,$Xi); #
1125 &pslld ($Xi,25); #
1127 &pxor ($T1,$Xi); #
1135 &movdqa ($Xi,$T3);
1137 &psrld ($Xi,1); #
1140 &pxor ($Xi,$T1);
1142 &pxor ($Xi,$T3); #
1143 &pxor ($Xi,$Xhi); #
1159 &movdqa ($Xi,$Hkey);
1160 &clmul64x64_T3 ($Xhi,$Xi,$Hkey);
1161 &reduction_alg5 ($Xhi,$Xi);
1164 &movdqu (&QWP(16,$Htbl),$Xi); # save H^2
1178 &movdqu ($Xi,&QWP(0,$Xip));
1181 &pshufb ($Xi,$Xn);
1183 &clmul64x64_T3 ($Xhi,$Xi,$Hkey);
1184 &reduction_alg5 ($Xhi,$Xi);
1186 &pshufb ($Xi,$Xn);
1187 &movdqu (&QWP(0,$Xip),$Xi);
1203 &movdqu ($Xi,&QWP(0,$Xip));
1206 &pshufb ($Xi,$T3);
1212 # Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
1213 # [(H*Ii+1) + (H*Xi+1)] mod P =
1214 # [(H*Ii+1) + H^2*(Ii+Xi)] mod P
1220 &pxor ($Xi,$T1); # Ii+Xi
1230 &clmul64x64_T3 ($Xhi,$Xi,$Hkey); # H^2*(Ii+Xi)
1233 &pxor ($Xi,$Xn); # (H*Ii+1) + H^2*(Ii+Xi)
1236 &reduction_alg5 ($Xhi,$Xi);
1244 &pxor ($Xi,$T1); # Ii+Xi
1254 &clmul64x64_T3 ($Xhi,$Xi,$Hkey); # H^2*(Ii+Xi)
1256 &pxor ($Xi,$Xn); # (H*Ii+1) + H^2*(Ii+Xi)
1259 &reduction_alg5 ($Xhi,$Xi);
1269 &pxor ($Xi,$T1); # Ii+Xi
1271 &clmul64x64_T3 ($Xhi,$Xi,$Hkey); # H*(Ii+Xi)
1272 &reduction_alg5 ($Xhi,$Xi);
1276 &pshufb ($Xi,$T3);
1277 &movdqu (&QWP(0,$Xip),$Xi);
