aes/asm/aes-586.pl

200 $s3="edx";
211 $__s3=&DWP(16,"esp");	# s3 backing store
247 	&mov	($v0,$s[3]);				# copy s3
264 	&xor	($s[3],&DWP(0,$te,$v0,8));		# s3>>0
267 	&xor	($s[2],&DWP(3,$te,$v0,8));		# s3>>8
270 	&xor	($s[1],&DWP(2,$te,$v1,8));		# s3>>16
272 	&xor	($s[0],&DWP(1,$te,$v0,8));		# s3>>24
308 	&rotr	($s3,16);			# 13,12,15,14
313 	&movz	($v0,&HB($s3));			# 13,12,15*,14
320 	&movz	($v0,&LB($s3));			# 13,12,15,14*
323 	&and	($s3,0xffff0000);		# 13,12, -, -
326 	&or	($s3,$s1);			# 13,12, 7, 6
333 	&movz	($v0,&HB($s3));			# 13,12, 7*, 6
337 	&mov	($v0,$s3);
346 	&movz	($v0,&LB($s3));			# 13,12, 7, 6*
347 	&shr	($s3,16);			#   ,  ,13,12
350 	&and	($s3,0xff);			#   ,  ,13,12*
351 	&mov	($s3,&DWP(0,$te,$s3,8));	# 12
352 	&xor	($s3,$s2);			# s[2]=t[3] collected
474 { my @s = ($s0,$s1,$s2,$s3);
506 	&xor	($s3,&DWP(12,$key));
525 		&enccompact(0,$tbl,$s0,$s1,$s2,$s3,1);
526 		&enccompact(1,$tbl,$s1,$s2,$s3,$s0,1);
527 		&enccompact(2,$tbl,$s2,$s3,$s0,$s1,1);
528 		&enccompact(3,$tbl,$s3,$s0,$s1,$s2,1);
539 		&xor	($s3,&DWP(12,$key));
545 	&enccompact(0,$tbl,$s0,$s1,$s2,$s3);
546 	&enccompact(1,$tbl,$s1,$s2,$s3,$s0);
547 	&enccompact(2,$tbl,$s2,$s3,$s0,$s1);
548 	&enccompact(3,$tbl,$s3,$s0,$s1,$s2);
553 	&xor	($s3,&DWP(28,$key));
574 # |     s3    |     s2    |     s1    |     s0    |
710 	&mov	($s3,&DWP(96-128,$tbl));
714 	&mov	($s3,&DWP(224-128,$tbl));
749 		&mov	($s3,&DWP(192-128,$tbl));
852 	&xor	($s3,&DWP(12,$key));
863 		    &encvert($tbl,$s0,$s1,$s2,$s3);
865 		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
866 		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
867 		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
868 		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
874 		&xor	($s3,&DWP(12,$key));
888 		    &encvert($tbl,$s0,$s1,$s2,$s3);
890 		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
891 		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
892 		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
893 		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
898 		&xor	($s3,&DWP(16*$i+12,$key));
905 		    &encvert($tbl,$s0,$s1,$s2,$s3);
907 		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
908 		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
909 		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
910 		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
915 		&xor	($s3,&DWP(16*$i+12,$key));
922 		    &encvert($tbl,$s0,$s1,$s2,$s3);
924 		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
925 		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
926 		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
927 		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
932 		&xor	($s3,&DWP(16*$i+12,$key));
941 	&enclast(0,$tbl,$s0,$s1,$s2,$s3);
942 	&enclast(1,$tbl,$s1,$s2,$s3,$s0);
943 	&enclast(2,$tbl,$s2,$s3,$s0,$s1);
944 	&enclast(3,$tbl,$s3,$s0,$s1,$s2);
950 	&xor	($s3,&DWP(12,$key));
1208 	&mov	($s3,&DWP(12,$acc));
1215 	&mov	(&DWP(12,$acc),$s3);
1266 { my @s = ($s0,$s1,$s2,$s3);
1332 	&xor	($s3,&DWP(12,$key));
1352 		&deccompact(0,$tbl,$s0,$s3,$s2,$s1,1);
1353 		&deccompact(1,$tbl,$s1,$s0,$s3,$s2,1);
1354 		&deccompact(2,$tbl,$s2,$s1,$s0,$s3,1);
1355 		&deccompact(3,$tbl,$s3,$s2,$s1,$s0,1);
1366 		&xor	($s3,&DWP(12,$key));
1372 	&deccompact(0,$tbl,$s0,$s3,$s2,$s1);
1373 	&deccompact(1,$tbl,$s1,$s0,$s3,$s2);
1374 	&deccompact(2,$tbl,$s2,$s1,$s0,$s3);
1375 	&deccompact(3,$tbl,$s3,$s2,$s1,$s0);
1380 	&xor	($s3,&DWP(28,$key));
1495 	&mov	($s3,&DWP(96-128,$tbl));
1499 	&mov	($s3,&DWP(224-128,$tbl));
1565 		&mov	($s3,&DWP(192-128,$tbl));
1668 	&xor	($s3,&DWP(12,$key));
1677 		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
1678 		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
1679 		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
1680 		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
1685 		&xor	($s3,&DWP(12,$key));
1698 		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
1699 		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
1700 		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
1701 		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
1705 		&xor	($s3,&DWP(16*$i+12,$key));
1711 		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
1712 		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
1713 		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
1714 		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
1718 		&xor	($s3,&DWP(16*$i+12,$key));
1724 		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
1725 		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
1726 		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
1727 		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
1731 		&xor	($s3,&DWP(16*$i+12,$key));
1735 	&declast(0,$tbl,$s0,$s3,$s2,$s1);
1736 	&declast(1,$tbl,$s1,$s0,$s3,$s2);
1737 	&declast(2,$tbl,$s2,$s1,$s0,$s3);
1738 	&declast(3,$tbl,$s3,$s2,$s1,$s0);
1744 	&xor	($s3,&DWP(12,$key));
1997 	&mov	($s3,&DWP(12,$acc));
2004 	&mov	(&DWP(12,$acc),$s3);
2016 #             12(%esp)		# s3 backing store	16(%esp)
2067 	&mov	($s3,$acc);
2070 	&and	($s3,0xfff);		# p = %esp&0xfff
2072 	&cmp	($s3,$s1);		# if (p>=e) %esp =- (p-e);
2074 	&sub	($s3,$s1);
2075 	&sub	($acc,$s3);
2078 	&sub	($s3,$s0);
2079 	&and	($s3,0xfff);
2080 	&add	($s3,384);
2081 	&sub	($acc,$s3);
2084 	&lea	($s3,&wparam(0));	# obtain pointer to parameter block
2090 	&mov	($s0,&DWP(0,$s3));	# load inp
2091 	&mov	($s1,&DWP(4,$s3));	# load out
2092 	#&mov	($s2,&DWP(8,$s3));	# load len
2093 	&mov	($key,&DWP(12,$s3));	# load key
2094 	&mov	($acc,&DWP(16,$s3));	# load ivp
2095 	&mov	($s3,&DWP(20,$s3));	# load enc flag
2134 	&cmp	($s3,0);
2143 		&mov	($s3,&DWP(12,$key));
2148 		&xor	($s3,&DWP(12,$acc));
2159 		&mov	(&DWP(12,$key),$s3);
2164 		&lea	($s3,&DWP(16,$key));	# advance out
2165 		&mov	($_out,$s3);		# save out
2171 	&mov	($s3,&DWP(12,$key));
2175 	&mov	(&DWP(12,$acc),$s3);
2205 		&mov	($s3,&DWP(12,$acc));
2215 		&xor	($s3,&DWP(12,$key));
2223 		&mov	(&DWP(12,$key),$s3);
2239 	&mov	($s3,&DWP(12,$key));
2243 	&mov	(&DWP(12,$acc),$s3);
2251 		&mov	($s3,&DWP(12,$acc));
2257 		&mov	(&DWP(12,$key),$s3);
2267 		&xor	($s3,&DWP(12,$key));
2272 		&mov	(&DWP(12,$acc),$s3);
2281 		&mov	($s3,&DWP(12,$acc));
2286 		&mov	(&DWP(12,$key),$s3);
2334 	&lea	($s3,&wparam(0));	# pointer to parameter block
2342 	&mov	($s0,&DWP(0,$s3));	# load inp
2343 	&mov	($s1,&DWP(4,$s3));	# load out
2344 	#&mov	($s2,&DWP(8,$s3));	# load len
2345 	#&mov	($key,&DWP(12,$s3));	# load key
2346 	&mov	($acc,&DWP(16,$s3));	# load ivp
2347 	&mov	($s3,&DWP(20,$s3));	# load enc flag
2358 	&cmp	($s3,0);
2363 	&mov	($s3,$s1);
2389 		&lea	($s3,&DWP(16,$key));	# advance out
2390 		&mov	($_out,$s3);		# save out
2412 		&mov	($s3,&DWP(12,$key));
2417 		&xor	($s3,&DWP(12,$acc));
2428 		&mov	(&DWP(12,$key),$s3);
2433 		&lea	($s3,&DWP(16,$key));	# advance out
2434 		&mov	($_out,$s3);		# save out
2443 	&mov	($s3,&DWP(12,$key));
2447 	&mov	(&DWP(12,$acc),$s3);
2456 	&mov	($key eq "edi"? $key:"",$s3);	# load out to edi
2473 	&mov	($acc,$s3);			# output as input
2545 		&mov	($s3,&DWP(12,$acc));
2551 		&mov	(&DWP(12,$key),$s3);
2561 		&xor	($s3,&DWP(12,$key));
2572 		&mov	(&DWP(12,$acc),$s3);
2581 		&mov	($s3,&DWP(12,$acc));
2586 		&mov	(&DWP(12,$key),$s3);
2602 	&mov	(&DWP(12,$acc),$s3);
2608 	&mov	($s3,&DWP(12,$acc));
2613 	&mov	(&DWP(12,$key),$s3);
2969 		&deckey	(0,$key,$s0,$s1,$s2,$s3);
2970 		&deckey	(1,$key,$s1,$s2,$s3,$s0);
2971 		&deckey	(2,$key,$s2,$s3,$s0,$s1);
2972 		&deckey	(3,$key,$s3,$s0,$s1,$s2);