camellia/asm/cmll-x86.pl

48 $idx="esi";
84 	&xor	($t0,$idx);				# t0^=key[0]
86 	&movz	($idx,&HB($t0));			# (t0>>8)&0xff
87 	&mov	($t3,&DWP($SBOX3_3033,$Tbl,$idx,8));	# t3=SBOX3_3033[0]
88 	&movz	($idx,&LB($t0));			# (t0>>0)&0xff
89 	&xor	($t3,&DWP($SBOX4_4404,$Tbl,$idx,8));	# t3^=SBOX4_4404[0]
91 	&movz	($idx,&LB($t1));			# (t1>>0)&0xff
92 	&mov	($t2,&DWP($SBOX1_1110,$Tbl,$idx,8));	# t2=SBOX1_1110[1]
93 	&movz	($idx,&HB($t0));			# (t0>>24)&0xff
94 	&xor	($t3,&DWP($SBOX1_1110,$Tbl,$idx,8));	# t3^=SBOX1_1110[0]
95 	&movz	($idx,&HB($t1));			# (t1>>8)&0xff
96 	&xor	($t2,&DWP($SBOX4_4404,$Tbl,$idx,8));	# t2^=SBOX4_4404[1]
100 	&movz	($idx,&HB($t1));			# (t1>>24)&0xff
104 	&xor	($t2,&DWP($SBOX2_0222,$Tbl,$idx,8));	# t2^=SBOX2_0222[1]
105 	&movz	($idx,&LB($t1));			# (t1>>16)&0xff
108 	&xor	($t2,&DWP($SBOX3_3033,$Tbl,$idx,8));	# t2^=SBOX3_3033[1]
109 	&mov	($idx,&DWP($seed+($i+1)*$scale,$key));	# prefetch key[i+1]
123 	&mov	($idx,&wparam(1));	# load plaintext pointer
149 	&mov	(@T[0],&DWP(0,$idx));	# load plaintext
150 	&mov	(@T[1],&DWP(4,$idx));
151 	&mov	(@T[2],&DWP(8,$idx));
153 	&mov	(@T[3],&DWP(12,$idx));
162 	&mov	($idx,&wparam(3));	# load ciphertext pointer
166 	&mov	(&DWP(0,$idx),@T[0]);	# write ciphertext
167 	&mov	(&DWP(4,$idx),@T[1]);
168 	&mov	(&DWP(8,$idx),@T[2]);
169 	&mov	(&DWP(12,$idx),@T[3]);
187 	&mov	($idx,&wparam(0));	# load plaintext pointer
214 	&mov	(@T[0],&DWP(0,$idx));	# load plaintext
215 	&mov	(@T[1],&DWP(4,$idx));
216 	&mov	(@T[2],&DWP(8,$idx));
218 	&mov	(@T[3],&DWP(12,$idx));
227 	&mov	($idx,&wparam(1));	# load ciphertext pointer
231 	&mov	(&DWP(0,$idx),@T[0]);	# write ciphertext
232 	&mov	(&DWP(4,$idx),@T[1]);
233 	&mov	(&DWP(8,$idx),@T[2]);
234 	&mov	(&DWP(12,$idx),@T[3]);
243 	&mov	($idx,&DWP(16,$key));	# prefetch key[4]
257 	# @T[0-1] are preloaded, $idx is preloaded with key[0]
258 	&and	($idx,@T[0]);
260 	&rotl	($idx,1);
262 	&xor	(@T[1],$idx);
267 	&mov	($idx,&DWP(4,$key));
269 	&or	($idx,@T[1]);
271 	&xor	(@T[0],$idx);
275 	&mov	($idx,&DWP(16,$key));		# prefetch key[4]
284 	&xor	(@T[0],$idx);		# $idx is preloaded with key[0]
298 	&mov	($idx,&wparam(1));	# load ciphertext pointer
324 	&mov	(@T[0],&DWP(0,$idx));	# load ciphertext
325 	&mov	(@T[1],&DWP(4,$idx));
326 	&mov	(@T[2],&DWP(8,$idx));
328 	&mov	(@T[3],&DWP(12,$idx));
337 	&mov	($idx,&wparam(3));	# load plaintext pointer
341 	&mov	(&DWP(0,$idx),@T[0]);	# write plaintext
342 	&mov	(&DWP(4,$idx),@T[1]);
343 	&mov	(&DWP(8,$idx),@T[2]);
344 	&mov	(&DWP(12,$idx),@T[3]);
362 	&mov	($idx,&wparam(0));	# load ciphertext pointer
389 	&mov	(@T[0],&DWP(0,$idx));	# load ciphertext
390 	&mov	(@T[1],&DWP(4,$idx));
391 	&mov	(@T[2],&DWP(8,$idx));
393 	&mov	(@T[3],&DWP(12,$idx));
402 	&mov	($idx,&wparam(1));	# load plaintext pointer
406 	&mov	(&DWP(0,$idx),@T[0]);	# write plaintext
407 	&mov	(&DWP(4,$idx),@T[1]);
408 	&mov	(&DWP(8,$idx),@T[2]);
409 	&mov	(&DWP(12,$idx),@T[3]);
418 	&mov	($idx,&DWP(-8,$key));	# prefetch key[-2]
432 	# @T[0-1] are preloaded, $idx is preloaded with key[2]
433 	&and	($idx,@T[0]);
435 	&rotl	($idx,1);
437 	&xor	(@T[1],$idx);
442 	&mov	($idx,&DWP(12,$key));
444 	&or	($idx,@T[1]);
446 	&xor	(@T[0],$idx);
450 	&mov	($idx,&DWP(-8,$key));	# prefetch key[4]
459 	&xor	(@T[2],$idx);		# $idx is preloaded with key[2]
475 	&mov	($idx,$i0);
479 	&shld	($i3,$idx,$rot);
497 	&mov	($idx,$i1);
498 	&shr	($idx,32-$rot);
500 	&or	($i0,$idx);
501 	&mov	($idx,$i2);
504 	&shr	($idx,32-$rot);
505 	&or	($i1,$idx);
507 	&mov	($idx,$i3);
508 	&shr	($idx,32-$rot);
511 	&or	($i2,$idx);
553 	&mov	($idx,&wparam(1));
556 	&mov	(@T[0],&DWP(0,$idx));		# load 0-127 bits
557 	&mov	(@T[1],&DWP(4,$idx));
558 	&mov	(@T[2],&DWP(8,$idx));
559 	&mov	(@T[3],&DWP(12,$idx));
571 	&mov	(@T[0],&DWP(16,$idx));		# load 128-191 bits
572 	&mov	(@T[1],&DWP(20,$idx));
575 	&mov	(@T[2],&DWP(24,$idx));		# load 192-255 bits
576 	&mov	(@T[3],&DWP(28,$idx));
603 	&mov	($idx,&DWP($step*8,$key));	# prefetch SIGMA[0]
613 	&mov	($idx,&wparam(2));
614 	&xor	(@T[0],&DWP(0*8+0,$idx));	# ^KL
615 	&xor	(@T[1],&DWP(0*8+4,$idx));
616 	&xor	(@T[2],&DWP(1*8+0,$idx));
617 	&xor	(@T[3],&DWP(1*8+4,$idx));
619 	&mov	($idx,&DWP($step*8,$key));	# prefetch SIGMA[4]
629 	&mov	($idx,&wparam(0));
630 	&cmp	($idx,128);
662 	&mov	($idx,&wparam(2));
663 	&_saveround	(6,$idx,@T);		# temporary storage for KA!
665 	&xor	(@T[0],&DWP(4*8+0,$idx));	# KA^KR
666 	&xor	(@T[1],&DWP(4*8+4,$idx));
667 	&xor	(@T[2],&DWP(5*8+0,$idx));
668 	&xor	(@T[3],&DWP(5*8+4,$idx));
670 	&mov	($idx,&DWP($step*8,$key));	# prefetch SIGMA[8]
836 	&lea	($idx,&DWP(-64,"esp"));
837 	&and	($idx,-64);
842 	&sub	($key,$idx);
845 	&sub	($idx,$key);
849 	&exch	("esp",$idx);
851 	&mov	($_esp,$idx);		# save %esp
864 	&mov	($idx,32);
871 		&dec	($idx);
875 	&mov	($idx,$_inp);
897 		&xor	($s0,&DWP(0,$idx));	# xor input data
898 		&xor	($s1,&DWP(4,$idx));
899 		&xor	($s2,&DWP(8,$idx));
901 		&xor	($s3,&DWP(12,$idx));
909 		&mov	($idx,$_inp);		# load inp
923 		&lea	($idx,&DWP(16,$idx));
924 		&mov	($_inp,$idx);		# save inp
935 	&mov	($idx,$_ivp);		# load ivp
938 	&mov	(&DWP(0,$idx),$s0);	# save ivec
939 	&mov	(&DWP(4,$idx),$s1);
940 	&mov	(&DWP(8,$idx),$s2);
941 	&mov	(&DWP(12,$idx),$s3);
955 	&cmp	($key,$idx);			# compare with inp
969 	&mov	($idx,$_out);			# output as input
982 	&cmp	($idx,$_out);
989 		&mov	($s0,&DWP(0,$idx));	# read input
990 		&mov	($s1,&DWP(4,$idx));
991 		&mov	($s2,&DWP(8,$idx));
993 		&mov	($s3,&DWP(12,$idx));
1002 		&mov	($idx,$_len);		# load len
1013 		&sub	($idx,16);
1015 		&mov	($_len,$idx);		# save len
1016 		&mov	($idx,$_inp);		# load inp
1024 		&mov	($_tmp,$idx);		# save ivp
1025 		&lea	($idx,&DWP(16,$idx));
1026 		&mov	($_inp,$idx);		# save inp
1034 	&mov	($idx,$_ivp);		# load user ivp
1039 	&mov	(&DWP(0,$idx),$s0);	# copy back to user
1040 	&mov	(&DWP(4,$idx),$s1);
1041 	&mov	(&DWP(8,$idx),$s2);
1042 	&mov	(&DWP(12,$idx),$s3);
1051 	&lea	($s2 eq "ecx" ? $s2 : "",&DWP(16,$idx));
1052 	&mov	($idx eq "esi" ? $idx : "",$key);
1061 		&mov	($s0,&DWP(0,$idx));	# read input
1062 		&mov	($s1,&DWP(4,$idx));
1063 		&mov	($s2,&DWP(8,$idx));
1064 		&mov	($s3,&DWP(12,$idx));
1079 		&mov	($idx,$_out);		# load out
1090 		&mov	(&DWP(0,$idx),$s0);	# write output
1091 		&mov	(&DWP(4,$idx),$s1);
1092 		&mov	(&DWP(8,$idx),$s2);
1093 		&mov	(&DWP(12,$idx),$s3);
1095 		&lea	($idx,&DWP(16,$idx));
1096 		&mov	($_out,$idx);		# save out
1098 		&lea	($idx,$ivec);
1099 		&mov	($s0,&DWP(0,$idx));	# read temp
1100 		&mov	($s1,&DWP(4,$idx));
1101 		&mov	($s2,&DWP(8,$idx));
1102 		&mov	($s3,&DWP(12,$idx));
1109 		&mov	($idx,$_inp);		# load inp
1111 		&lea	($idx,&DWP(16,$idx));
1112 		&mov	($_inp,$idx);		# save inp
1124 	&lea	($idx eq "esi" ? $idx : "",$ivec);
1126 	&lea	($idx,&DWP(16,$idx,$s2));