summaryrefslogtreecommitdiff
path: root/src/lib/libcrypto/sha/asm
diff options
context:
space:
mode:
Diffstat (limited to 'src/lib/libcrypto/sha/asm')
-rw-r--r--src/lib/libcrypto/sha/asm/sha1-586.pl540
1 files changed, 540 insertions, 0 deletions
diff --git a/src/lib/libcrypto/sha/asm/sha1-586.pl b/src/lib/libcrypto/sha/asm/sha1-586.pl
new file mode 100644
index 0000000000..fe51fd0794
--- /dev/null
+++ b/src/lib/libcrypto/sha/asm/sha1-586.pl
@@ -0,0 +1,540 @@
1#!/usr/local/bin/perl
2
3$normal=0;
4
5push(@INC,"perlasm","../../perlasm");
6require "x86asm.pl";
7
8&asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386");
9
10$A="eax";
11$B="ecx";
12$C="ebx";
13$D="edx";
14$E="edi";
15$T="esi";
16$tmp1="ebp";
17
18$off=9*4;
19
20@K=(0x5a827999,0x6ed9eba1,0x8f1bbcdc,0xca62c1d6);
21
22&sha1_block_data("sha1_block_asm_data_order");
23
24&asm_finish();
25
26sub Nn
27 {
28 local($p)=@_;
29 local(%n)=($A,$T,$B,$A,$C,$B,$D,$C,$E,$D,$T,$E);
30 return($n{$p});
31 }
32
33sub Np
34 {
35 local($p)=@_;
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);
38 return($n{$p});
39 }
40
41sub Na
42 {
43 local($n)=@_;
44 return( (($n )&0x0f),
45 (($n+ 2)&0x0f),
46 (($n+ 8)&0x0f),
47 (($n+13)&0x0f),
48 (($n+ 1)&0x0f));
49 }
50
51sub X_expand
52 {
53 local($in)=@_;
54
55 &comment("First, load the words onto the stack in network byte order");
56 for ($i=0; $i<16; $i+=2)
57 {
58 &mov($A,&DWP(($i+0)*4,$in,"",0));# unless $i == 0;
59 &mov($B,&DWP(($i+1)*4,$in,"",0));
60 &bswap($A);
61 &bswap($B);
62 &mov(&swtmp($i+0),$A);
63 &mov(&swtmp($i+1),$B);
64 }
65
66 &comment("We now have the X array on the stack");
67 &comment("starting at sp-4");
68 }
69
70# Rules of engagement
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].
75
76sub BODY_00_15
77 {
78 local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
79
80return if $n & 1;
81 &comment("00_15 $n");
82
83 &mov($f,$c);
84
85 &mov($tmp1,$a);
86 &xor($f,$d); # F2
87
88 &rotl($tmp1,5); # A2
89
90 &and($f,$b); # F3
91 &add($tmp1,$e);
92
93 &rotr($b,1); # B1 <- F
94 &mov($e,&swtmp($n)); # G1
95
96 &rotr($b,1); # B1 <- F
97 &xor($f,$d); # F4
98
99 &lea($tmp1,&DWP($K,$tmp1,$e,1));
100
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);
104$n++;
105 local($n0,$n1,$n2,$n3,$np)=&Na($n);
106 ($b,$c,$d,$e,$f,$a)=($a,$b,$c,$d,$e,$f);
107
108 &mov($f,$c);
109
110 &add($a,$tmp1); # MOVED DOWN
111 &xor($f,$d); # F2
112
113 &mov($tmp1,$a);
114 &and($f,$b); # F3
115
116 &rotl($tmp1,5); # A2
117
118 &add($tmp1,$e);
119 &mov($e,&swtmp($n)); # G1
120
121 &rotr($b,1); # B1 <- F
122 &xor($f,$d); # F4
123
124 &rotr($b,1); # B1 <- F
125 &lea($tmp1,&DWP($K,$tmp1,$e,1));
126
127 &add($f,$tmp1);
128 }
129
130sub BODY_16_19
131 {
132 local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
133 local($n0,$n1,$n2,$n3,$np)=&Na($n);
134
135return if $n & 1;
136 &comment("16_19 $n");
137
138 &nop() if ($pos < 0);
139&mov($tmp1,&swtmp($n0)); # X1
140 &mov($f,&swtmp($n1)); # X2
141&xor($f,$tmp1); # X3
142 &mov($tmp1,&swtmp($n2)); # X4
143&xor($f,$tmp1); # X5
144 &mov($tmp1,&swtmp($n3)); # X6
145&xor($f,$tmp1); # X7 - slot
146 &mov($tmp1,$c); # F1
147&rotl($f,1); # X8 - slot
148 &xor($tmp1,$d); # F2
149&mov(&swtmp($n0),$f); # X9 - anytime
150 &and($tmp1,$b); # F3
151&lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
152 &xor($tmp1,$d); # F4
153&mov($e,$a); # A1
154 &add($f,$tmp1); # tot+=F();
155
156&rotl($e,5); # A2
157
158&rotr($b,1); # B1 <- F
159 &add($f,$e); # tot+=a
160
161############################
162# &BODY_40_59( 0,$K[2],$X,42,$A,$B,$C,$D,$E,$T);
163# &BODY_40_59( 0,$K[2],$X,43,$T,$A,$B,$C,$D,$E);
164$n++;
165 local($n0,$n1,$n2,$n3,$np)=&Na($n);
166 ($b,$c,$d,$e,$f,$a)=($a,$b,$c,$d,$e,$f);
167
168
169&mov($f,&swtmp($n0)); # X1
170 &mov($tmp1,&swtmp($n1)); # X2
171&xor($f,$tmp1); # X3
172 &mov($tmp1,&swtmp($n2)); # X4
173&xor($f,$tmp1); # X5
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($f,1); # X8 - slot
178 &mov($tmp1,$c); # F1
179&xor($tmp1,$d); # F2
180 &mov(&swtmp($n0),$f); # X9 - anytime
181&and($tmp1,$b); # F3
182 &lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
183
184&xor($tmp1,$d); # F4
185 &mov($e,$a); # A1
186
187&rotl($e,5); # A2
188
189&rotr($b,1); # B1 <- F
190 &add($f,$e); # tot+=a
191
192&rotr($b,1); # B1 <- F
193 &add($f,$tmp1); # tot+=F();
194
195 }
196
197sub BODY_20_39
198 {
199 local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
200
201 &comment("20_39 $n");
202 local($n0,$n1,$n2,$n3,$np)=&Na($n);
203
204&mov($f,&swtmp($n0)); # X1
205 &mov($tmp1,&swtmp($n1)); # X2
206&xor($f,$tmp1); # X3
207 &mov($tmp1,&swtmp($n2)); # X4
208&xor($f,$tmp1); # X5
209 &mov($tmp1,&swtmp($n3)); # X6
210&xor($f,$tmp1); # X7 - slot
211 &mov($tmp1,$b); # F1
212&rotl($f,1); # X8 - slot
213 &xor($tmp1,$c); # F2
214&mov(&swtmp($n0),$f); # X9 - anytime
215 &xor($tmp1,$d); # F3
216
217&lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
218 &mov($e,$a); # A1
219
220&rotl($e,5); # A2
221
222if ($n != 79) # last loop
223 {
224 &rotr($b,1); # B1 <- F
225 &add($e,$tmp1); # tmp1=F()+a
226
227 &rotr($b,1); # B2 <- F
228 &add($f,$e); # tot+=tmp1;
229 }
230else
231 {
232 &add($e,$tmp1); # tmp1=F()+a
233 &mov($tmp1,&wparam(0));
234
235 &rotr($b,1); # B1 <- F
236 &add($f,$e); # tot+=tmp1;
237
238 &rotr($b,1); # B2 <- F
239 }
240 }
241
242sub BODY_40_59
243 {
244 local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
245
246 &comment("40_59 $n");
247 return if $n & 1;
248 local($n0,$n1,$n2,$n3,$np)=&Na($n);
249
250&mov($f,&swtmp($n0)); # X1
251 &mov($tmp1,&swtmp($n1)); # X2
252&xor($f,$tmp1); # X3
253 &mov($tmp1,&swtmp($n2)); # X4
254&xor($f,$tmp1); # X5
255 &mov($tmp1,&swtmp($n3)); # X6
256&xor($f,$tmp1); # X7 - slot
257 &mov($tmp1,$b); # F1
258&rotl($f,1); # X8 - slot
259 &or($tmp1,$c); # F2
260&mov(&swtmp($n0),$f); # X9 - anytime
261 &and($tmp1,$d); # F3
262
263&lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
264 &mov($e,$b); # F4
265
266&rotr($b,1); # B1 <- F
267 &and($e,$c); # F5
268
269&or($tmp1,$e); # F6
270 &mov($e,$a); # A1
271
272&rotl($e,5); # A2
273
274&add($tmp1,$e); # tmp1=F()+a
275
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);
279$n++;
280 local($n0,$n1,$n2,$n3,$np)=&Na($n);
281 ($b,$c,$d,$e,$f,$a)=($a,$b,$c,$d,$e,$f);
282
283 &mov($f,&swtmp($n0)); # X1
284&add($a,$tmp1); # tot+=tmp1; # moved was add f,tmp1
285 &mov($tmp1,&swtmp($n1)); # X2
286&xor($f,$tmp1); # X3
287 &mov($tmp1,&swtmp($n2)); # X4
288&xor($f,$tmp1); # X5
289 &mov($tmp1,&swtmp($n3)); # X6
290&rotr($c,1); # B2 <- F # moved was rotr b,1
291 &xor($f,$tmp1); # X7 - slot
292&rotl($f,1); # X8 - slot
293 &mov($tmp1,$b); # F1
294&mov(&swtmp($n0),$f); # X9 - anytime
295 &or($tmp1,$c); # F2
296&lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
297 &mov($e,$b); # F4
298&and($tmp1,$d); # F3
299 &and($e,$c); # F5
300
301&or($tmp1,$e); # F6
302 &mov($e,$a); # A1
303
304&rotl($e,5); # A2
305
306&rotr($b,1); # B1 <- F
307 &add($tmp1,$e); # tmp1=F()+a
308
309&rotr($b,1); # B2 <- F
310 &add($f,$tmp1); # tot+=tmp1;
311 }
312
313sub BODY_60_79
314 {
315 &BODY_20_39(@_);
316 }
317
318sub sha1_block_host
319 {
320 local($name, $sclabel)=@_;
321
322 &function_begin_B($name,"");
323
324 # parameter 1 is the MD5_CTX structure.
325 # A 0
326 # B 4
327 # C 8
328 # D 12
329 # E 16
330
331 &mov("ecx", &wparam(2));
332 &push("esi");
333 &shl("ecx",6);
334 &mov("esi", &wparam(1));
335 &push("ebp");
336 &add("ecx","esi"); # offset to leave on
337 &push("ebx");
338 &mov("ebp", &wparam(0));
339 &push("edi");
340 &mov($D, &DWP(12,"ebp","",0));
341 &stack_push(18+9);
342 &mov($E, &DWP(16,"ebp","",0));
343 &mov($C, &DWP( 8,"ebp","",0));
344 &mov(&swtmp(17),"ecx");
345
346 &comment("First we need to setup the X array");
347
348 for ($i=0; $i<16; $i+=2)
349 {
350 &mov($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),$B);
354 }
355 &jmp($sclabel);
356 &function_end_B($name);
357 }
358
359
360sub sha1_block_data
361 {
362 local($name)=@_;
363
364 &function_begin_B($name,"");
365
366 # parameter 1 is the MD5_CTX structure.
367 # A 0
368 # B 4
369 # C 8
370 # D 12
371 # E 16
372
373 &mov("ecx", &wparam(2));
374 &push("esi");
375 &shl("ecx",6);
376 &mov("esi", &wparam(1));
377 &push("ebp");
378 &add("ecx","esi"); # offset to leave on
379 &push("ebx");
380 &mov("ebp", &wparam(0));
381 &push("edi");
382 &mov($D, &DWP(12,"ebp","",0));
383 &stack_push(18+9);
384 &mov($E, &DWP(16,"ebp","",0));
385 &mov($C, &DWP( 8,"ebp","",0));
386 &mov(&swtmp(17),"ecx");
387
388 &comment("First we need to setup the X array");
389
390 &set_label("start") unless $normal;
391
392 &X_expand("esi");
393 &mov(&wparam(1),"esi");
394
395 &set_label("shortcut", 0, 1);
396 &comment("");
397 &comment("Start processing");
398
399 # odd start
400 &mov($A, &DWP( 0,"ebp","",0));
401 &mov($B, &DWP( 4,"ebp","",0));
402 $X="esp";
403 &BODY_00_15(-2,$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,$E);
423
424 &BODY_20_39(-1,$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,$C);
444
445 &BODY_40_59(-1,$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,$A);
465
466 &BODY_60_79(-1,$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,$E);
486
487 &comment("End processing");
488 &comment("");
489 # D is the tmp value
490
491 # E -> A
492 # T -> B
493 # A -> C
494 # B -> D
495 # C -> E
496 # D -> T
497
498 # The last 2 have been moved into the last loop
499 # &mov($tmp1,&wparam(0));
500
501 &mov($D, &DWP(12,$tmp1,"",0));
502 &add($D,$B);
503 &mov($B, &DWP( 4,$tmp1,"",0));
504 &add($B,$T);
505 &mov($T, $A);
506 &mov($A, &DWP( 0,$tmp1,"",0));
507 &mov(&DWP(12,$tmp1,"",0),$D);
508
509 &add($A,$E);
510 &mov($E, &DWP(16,$tmp1,"",0));
511 &add($E,$C);
512 &mov($C, &DWP( 8,$tmp1,"",0));
513 &add($C,$T);
514
515 &mov(&DWP( 0,$tmp1,"",0),$A);
516 &mov("esi",&wparam(1));
517 &mov(&DWP( 8,$tmp1,"",0),$C);
518 &add("esi",64);
519 &mov("eax",&swtmp(17));
520 &mov(&DWP(16,$tmp1,"",0),$E);
521 &cmp("esi","eax");
522 &mov(&DWP( 4,$tmp1,"",0),$B);
523 &jl(&label("start"));
524
525 &stack_pop(18+9);
526 &pop("edi");
527 &pop("ebx");
528 &pop("ebp");
529 &pop("esi");
530 &ret();
531
532 # keep a note of shortcut label so it can be used outside
533 # block.
534 my $sclabel = &label("shortcut");
535
536 &function_end_B($name);
537 # Putting this here avoids problems with MASM in debugging mode
538 &sha1_block_host("sha1_block_asm_host_order", $sclabel);
539 }
540
generated by cgit v1.2.3-55-g6feb (git 2.39.1) at 2025-12-11 19:13:10 +0000