5 push(@INC,"perlasm","../../perlasm");
8 &asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386");
20 @K=(0x5a827999,0x6ed9eba1,0x8f1bbcdc,0xca62c1d6);
22 &sha1_block_data("sha1_block_asm_data_order");
29 local(%n)=($A,$T,$B,$A,$C,$B,$D,$C,$E,$D,$T,$E);
36 local(%n)=($A,$T,$B,$A,$C,$B,$D,$C,$E,$D,$T,$E);
37 local(%n)=($A,$B,$B,$C,$C,$D,$D,$E,$E,$T,$T,$A);
55 &comment("First, load the words onto the stack in network byte order");
56 for ($i=0; $i<16; $i+=2)
58 &mov($A,&DWP(($i+0)*4,$in,"",0));# unless $i == 0;
59 &mov($B,&DWP(($i+1)*4,$in,"",0));
62 &mov(&swtmp($i+0),$A);
63 &mov(&swtmp($i+1),$B);
66 &comment("We now have the X array on the stack");
67 &comment("starting at sp-4");
71 # F is always trashable at the start, the running total.
72 # E becomes the next F so it can be trashed after it has been 'accumulated'
73 # F becomes A in the next round. We don't need to access it much.
74 # During the X update part, the result ends up in $X[$n0].
78 local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
93 &rotr($b,1); # B1 <- F
94 &mov($e,&swtmp($n)); # G1
96 &rotr($b,1); # B1 <- F
99 &lea($tmp1,&DWP($K,$tmp1,$e,1));
101 ############################
102 # &BODY_40_59( 0,$K[2],$X,42,$A,$B,$C,$D,$E,$T);
103 # &BODY_40_59( 0,$K[2],$X,43,$T,$A,$B,$C,$D,$E);
105 local($n0,$n1,$n2,$n3,$np)=&Na($n);
106 ($b,$c,$d,$e,$f,$a)=($a,$b,$c,$d,$e,$f);
110 &add($a,$tmp1); # MOVED DOWN
119 &mov($e,&swtmp($n)); # G1
121 &rotr($b,1); # B1 <- F
124 &rotr($b,1); # B1 <- F
125 &lea($tmp1,&DWP($K,$tmp1,$e,1));
132 local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
133 local($n0,$n1,$n2,$n3,$np)=&Na($n);
136 &comment("16_19 $n");
138 &nop() if ($pos < 0);
139 &mov($tmp1,&swtmp($n0)); # X1
140 &mov($f,&swtmp($n1)); # X2
142 &mov($tmp1,&swtmp($n2)); # X4
144 &mov($tmp1,&swtmp($n3)); # X6
145 &xor($f,$tmp1); # X7 - slot
147 &rotl($f,1); # X8 - slot
149 &mov(&swtmp($n0),$f); # X9 - anytime
151 &lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
154 &add($f,$tmp1); # tot+=F();
158 &rotr($b,1); # B1 <- F
159 &add($f,$e); # tot+=a
161 ############################
162 # &BODY_40_59( 0,$K[2],$X,42,$A,$B,$C,$D,$E,$T);
163 # &BODY_40_59( 0,$K[2],$X,43,$T,$A,$B,$C,$D,$E);
165 local($n0,$n1,$n2,$n3,$np)=&Na($n);
166 ($b,$c,$d,$e,$f,$a)=($a,$b,$c,$d,$e,$f);
169 &mov($f,&swtmp($n0)); # X1
170 &mov($tmp1,&swtmp($n1)); # X2
172 &mov($tmp1,&swtmp($n2)); # X4
174 &mov($tmp1,&swtmp($n3)); # X6
175 &rotr($c,1); #&rotr($b,1); # B1 <- F # MOVED DOWN
176 &xor($f,$tmp1); # X7 - slot
177 &rotl($f,1); # X8 - slot
180 &mov(&swtmp($n0),$f); # X9 - anytime
182 &lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
189 &rotr($b,1); # B1 <- F
190 &add($f,$e); # tot+=a
192 &rotr($b,1); # B1 <- F
193 &add($f,$tmp1); # tot+=F();
199 local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
201 &comment("20_39 $n");
202 local($n0,$n1,$n2,$n3,$np)=&Na($n);
204 &mov($f,&swtmp($n0)); # X1
205 &mov($tmp1,&swtmp($n1)); # X2
207 &mov($tmp1,&swtmp($n2)); # X4
209 &mov($tmp1,&swtmp($n3)); # X6
210 &xor($f,$tmp1); # X7 - slot
212 &rotl($f,1); # X8 - slot
214 &mov(&swtmp($n0),$f); # X9 - anytime
217 &lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
222 if ($n != 79) # last loop
224 &rotr($b,1); # B1 <- F
225 &add($e,$tmp1); # tmp1=F()+a
227 &rotr($b,1); # B2 <- F
228 &add($f,$e); # tot+=tmp1;
232 &add($e,$tmp1); # tmp1=F()+a
233 &mov($tmp1,&wparam(0));
235 &rotr($b,1); # B1 <- F
236 &add($f,$e); # tot+=tmp1;
238 &rotr($b,1); # B2 <- F
244 local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
246 &comment("40_59 $n");
248 local($n0,$n1,$n2,$n3,$np)=&Na($n);
250 &mov($f,&swtmp($n0)); # X1
251 &mov($tmp1,&swtmp($n1)); # X2
253 &mov($tmp1,&swtmp($n2)); # X4
255 &mov($tmp1,&swtmp($n3)); # X6
256 &xor($f,$tmp1); # X7 - slot
258 &rotl($f,1); # X8 - slot
260 &mov(&swtmp($n0),$f); # X9 - anytime
263 &lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
266 &rotr($b,1); # B1 <- F
274 &add($tmp1,$e); # tmp1=F()+a
276 ############################
277 # &BODY_40_59( 0,$K[2],$X,42,$A,$B,$C,$D,$E,$T);
278 # &BODY_40_59( 0,$K[2],$X,43,$T,$A,$B,$C,$D,$E);
280 local($n0,$n1,$n2,$n3,$np)=&Na($n);
281 ($b,$c,$d,$e,$f,$a)=($a,$b,$c,$d,$e,$f);
283 &mov($f,&swtmp($n0)); # X1
284 &add($a,$tmp1); # tot+=tmp1; # moved was add f,tmp1
285 &mov($tmp1,&swtmp($n1)); # X2
287 &mov($tmp1,&swtmp($n2)); # X4
289 &mov($tmp1,&swtmp($n3)); # X6
290 &rotr($c,1); # B2 <- F # moved was rotr b,1
291 &xor($f,$tmp1); # X7 - slot
292 &rotl($f,1); # X8 - slot
294 &mov(&swtmp($n0),$f); # X9 - anytime
296 &lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
306 &rotr($b,1); # B1 <- F
307 &add($tmp1,$e); # tmp1=F()+a
309 &rotr($b,1); # B2 <- F
310 &add($f,$tmp1); # tot+=tmp1;
322 &function_begin_B($name,"");
324 # parameter 1 is the MD5_CTX structure.
331 &mov("ecx", &wparam(2));
334 &mov("esi", &wparam(1));
336 &add("ecx","esi"); # offset to leave on
338 &mov("ebp", &wparam(0));
340 &mov($D, &DWP(12,"ebp","",0));
342 &mov($E, &DWP(16,"ebp","",0));
343 &mov($C, &DWP( 8,"ebp","",0));
344 &mov(&swtmp(17),"ecx");
346 &comment("First we need to setup the X array");
348 for ($i=0; $i<16; $i+=2)
350 &mov($A,&DWP(($i+0)*4,"esi","",0));# unless $i == 0;
351 &mov($B,&DWP(($i+1)*4,"esi","",0));
352 &mov(&swtmp($i+0),$A);
353 &mov(&swtmp($i+1),$B);
355 &jmp(&label("shortcut"));
356 &function_end_B($name);
364 &function_begin_B($name,"");
366 # parameter 1 is the MD5_CTX structure.
373 &mov("ecx", &wparam(2));
376 &mov("esi", &wparam(1));
378 &add("ecx","esi"); # offset to leave on
380 &mov("ebp", &wparam(0));
382 &mov($D, &DWP(12,"ebp","",0));
384 &mov($E, &DWP(16,"ebp","",0));
385 &mov($C, &DWP( 8,"ebp","",0));
386 &mov(&swtmp(17),"ecx");
388 &comment("First we need to setup the X array");
390 &set_label("start") unless $normal;
393 &mov(&wparam(1),"esi");
395 &set_label("shortcut", 1);
397 &comment("Start processing");
400 &mov($A, &DWP( 0,"ebp","",0));
401 &mov($B, &DWP( 4,"ebp","",0));
403 &BODY_00_15(-2,$K[0],$X, 0,$A,$B,$C,$D,$E,$T);
404 &BODY_00_15( 0,$K[0],$X, 1,$T,$A,$B,$C,$D,$E);
405 &BODY_00_15( 0,$K[0],$X, 2,$E,$T,$A,$B,$C,$D);
406 &BODY_00_15( 0,$K[0],$X, 3,$D,$E,$T,$A,$B,$C);
407 &BODY_00_15( 0,$K[0],$X, 4,$C,$D,$E,$T,$A,$B);
408 &BODY_00_15( 0,$K[0],$X, 5,$B,$C,$D,$E,$T,$A);
409 &BODY_00_15( 0,$K[0],$X, 6,$A,$B,$C,$D,$E,$T);
410 &BODY_00_15( 0,$K[0],$X, 7,$T,$A,$B,$C,$D,$E);
411 &BODY_00_15( 0,$K[0],$X, 8,$E,$T,$A,$B,$C,$D);
412 &BODY_00_15( 0,$K[0],$X, 9,$D,$E,$T,$A,$B,$C);
413 &BODY_00_15( 0,$K[0],$X,10,$C,$D,$E,$T,$A,$B);
414 &BODY_00_15( 0,$K[0],$X,11,$B,$C,$D,$E,$T,$A);
415 &BODY_00_15( 0,$K[0],$X,12,$A,$B,$C,$D,$E,$T);
416 &BODY_00_15( 0,$K[0],$X,13,$T,$A,$B,$C,$D,$E);
417 &BODY_00_15( 0,$K[0],$X,14,$E,$T,$A,$B,$C,$D);
418 &BODY_00_15( 1,$K[0],$X,15,$D,$E,$T,$A,$B,$C);
419 &BODY_16_19(-1,$K[0],$X,16,$C,$D,$E,$T,$A,$B);
420 &BODY_16_19( 0,$K[0],$X,17,$B,$C,$D,$E,$T,$A);
421 &BODY_16_19( 0,$K[0],$X,18,$A,$B,$C,$D,$E,$T);
422 &BODY_16_19( 1,$K[0],$X,19,$T,$A,$B,$C,$D,$E);
424 &BODY_20_39(-1,$K[1],$X,20,$E,$T,$A,$B,$C,$D);
425 &BODY_20_39( 0,$K[1],$X,21,$D,$E,$T,$A,$B,$C);
426 &BODY_20_39( 0,$K[1],$X,22,$C,$D,$E,$T,$A,$B);
427 &BODY_20_39( 0,$K[1],$X,23,$B,$C,$D,$E,$T,$A);
428 &BODY_20_39( 0,$K[1],$X,24,$A,$B,$C,$D,$E,$T);
429 &BODY_20_39( 0,$K[1],$X,25,$T,$A,$B,$C,$D,$E);
430 &BODY_20_39( 0,$K[1],$X,26,$E,$T,$A,$B,$C,$D);
431 &BODY_20_39( 0,$K[1],$X,27,$D,$E,$T,$A,$B,$C);
432 &BODY_20_39( 0,$K[1],$X,28,$C,$D,$E,$T,$A,$B);
433 &BODY_20_39( 0,$K[1],$X,29,$B,$C,$D,$E,$T,$A);
434 &BODY_20_39( 0,$K[1],$X,30,$A,$B,$C,$D,$E,$T);
435 &BODY_20_39( 0,$K[1],$X,31,$T,$A,$B,$C,$D,$E);
436 &BODY_20_39( 0,$K[1],$X,32,$E,$T,$A,$B,$C,$D);
437 &BODY_20_39( 0,$K[1],$X,33,$D,$E,$T,$A,$B,$C);
438 &BODY_20_39( 0,$K[1],$X,34,$C,$D,$E,$T,$A,$B);
439 &BODY_20_39( 0,$K[1],$X,35,$B,$C,$D,$E,$T,$A);
440 &BODY_20_39( 0,$K[1],$X,36,$A,$B,$C,$D,$E,$T);
441 &BODY_20_39( 0,$K[1],$X,37,$T,$A,$B,$C,$D,$E);
442 &BODY_20_39( 0,$K[1],$X,38,$E,$T,$A,$B,$C,$D);
443 &BODY_20_39( 1,$K[1],$X,39,$D,$E,$T,$A,$B,$C);
445 &BODY_40_59(-1,$K[2],$X,40,$C,$D,$E,$T,$A,$B);
446 &BODY_40_59( 0,$K[2],$X,41,$B,$C,$D,$E,$T,$A);
447 &BODY_40_59( 0,$K[2],$X,42,$A,$B,$C,$D,$E,$T);
448 &BODY_40_59( 0,$K[2],$X,43,$T,$A,$B,$C,$D,$E);
449 &BODY_40_59( 0,$K[2],$X,44,$E,$T,$A,$B,$C,$D);
450 &BODY_40_59( 0,$K[2],$X,45,$D,$E,$T,$A,$B,$C);
451 &BODY_40_59( 0,$K[2],$X,46,$C,$D,$E,$T,$A,$B);
452 &BODY_40_59( 0,$K[2],$X,47,$B,$C,$D,$E,$T,$A);
453 &BODY_40_59( 0,$K[2],$X,48,$A,$B,$C,$D,$E,$T);
454 &BODY_40_59( 0,$K[2],$X,49,$T,$A,$B,$C,$D,$E);
455 &BODY_40_59( 0,$K[2],$X,50,$E,$T,$A,$B,$C,$D);
456 &BODY_40_59( 0,$K[2],$X,51,$D,$E,$T,$A,$B,$C);
457 &BODY_40_59( 0,$K[2],$X,52,$C,$D,$E,$T,$A,$B);
458 &BODY_40_59( 0,$K[2],$X,53,$B,$C,$D,$E,$T,$A);
459 &BODY_40_59( 0,$K[2],$X,54,$A,$B,$C,$D,$E,$T);
460 &BODY_40_59( 0,$K[2],$X,55,$T,$A,$B,$C,$D,$E);
461 &BODY_40_59( 0,$K[2],$X,56,$E,$T,$A,$B,$C,$D);
462 &BODY_40_59( 0,$K[2],$X,57,$D,$E,$T,$A,$B,$C);
463 &BODY_40_59( 0,$K[2],$X,58,$C,$D,$E,$T,$A,$B);
464 &BODY_40_59( 1,$K[2],$X,59,$B,$C,$D,$E,$T,$A);
466 &BODY_60_79(-1,$K[3],$X,60,$A,$B,$C,$D,$E,$T);
467 &BODY_60_79( 0,$K[3],$X,61,$T,$A,$B,$C,$D,$E);
468 &BODY_60_79( 0,$K[3],$X,62,$E,$T,$A,$B,$C,$D);
469 &BODY_60_79( 0,$K[3],$X,63,$D,$E,$T,$A,$B,$C);
470 &BODY_60_79( 0,$K[3],$X,64,$C,$D,$E,$T,$A,$B);
471 &BODY_60_79( 0,$K[3],$X,65,$B,$C,$D,$E,$T,$A);
472 &BODY_60_79( 0,$K[3],$X,66,$A,$B,$C,$D,$E,$T);
473 &BODY_60_79( 0,$K[3],$X,67,$T,$A,$B,$C,$D,$E);
474 &BODY_60_79( 0,$K[3],$X,68,$E,$T,$A,$B,$C,$D);
475 &BODY_60_79( 0,$K[3],$X,69,$D,$E,$T,$A,$B,$C);
476 &BODY_60_79( 0,$K[3],$X,70,$C,$D,$E,$T,$A,$B);
477 &BODY_60_79( 0,$K[3],$X,71,$B,$C,$D,$E,$T,$A);
478 &BODY_60_79( 0,$K[3],$X,72,$A,$B,$C,$D,$E,$T);
479 &BODY_60_79( 0,$K[3],$X,73,$T,$A,$B,$C,$D,$E);
480 &BODY_60_79( 0,$K[3],$X,74,$E,$T,$A,$B,$C,$D);
481 &BODY_60_79( 0,$K[3],$X,75,$D,$E,$T,$A,$B,$C);
482 &BODY_60_79( 0,$K[3],$X,76,$C,$D,$E,$T,$A,$B);
483 &BODY_60_79( 0,$K[3],$X,77,$B,$C,$D,$E,$T,$A);
484 &BODY_60_79( 0,$K[3],$X,78,$A,$B,$C,$D,$E,$T);
485 &BODY_60_79( 2,$K[3],$X,79,$T,$A,$B,$C,$D,$E);
487 &comment("End processing");
498 # The last 2 have been moved into the last loop
499 # &mov($tmp1,&wparam(0));
501 &mov($D, &DWP(12,$tmp1,"",0));
503 &mov($B, &DWP( 4,$tmp1,"",0));
506 &mov($A, &DWP( 0,$tmp1,"",0));
507 &mov(&DWP(12,$tmp1,"",0),$D);
510 &mov($E, &DWP(16,$tmp1,"",0));
512 &mov($C, &DWP( 8,$tmp1,"",0));
515 &mov(&DWP( 0,$tmp1,"",0),$A);
516 &mov("esi",&wparam(1));
517 &mov(&DWP( 8,$tmp1,"",0),$C);
519 &mov("eax",&swtmp(17));
520 &mov(&DWP(16,$tmp1,"",0),$E);
522 &mov(&DWP( 4,$tmp1,"",0),$B);
523 &jl(&label("start"));
532 # it has to reside within sha1_block_asm_host_order body
533 # because it calls &jmp(&label("shortcut"));
534 &sha1_block_host("sha1_block_asm_host_order");
536 &function_end_B($name);