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.pl491
1 files changed, 491 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..d6d998f8ee
--- /dev/null
+++ b/src/lib/libcrypto/sha/asm/sha1-586.pl
@@ -0,0 +1,491 @@
1#!/usr/bin/perl
2
3$normal=0;
4
5push(@INC,"perlasm","../../perlasm");
6require "x86asm.pl";
7
8&asm_init($ARGV[0],"sha1-586.pl");
9
10$A="eax";
11$B="ebx";
12$C="ecx";
13$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("sha1_block_x86");
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; $i++)
57 {
58 &mov("eax",&DWP(($i+0)*4,$in,"",0)) unless $i == 0;
59 &bswap("eax");
60 &mov(&swtmp($i+0),"eax");
61 }
62
63 &comment("We now have the X array on the stack");
64 &comment("starting at sp-4");
65 }
66
67# Rules of engagement
68# F is always trashable at the start, the running total.
69# E becomes the next F so it can be trashed after it has been 'accumulated'
70# F becomes A in the next round. We don't need to access it much.
71# During the X update part, the result ends up in $X[$n0].
72
73sub BODY_00_15
74 {
75 local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
76
77return if $n & 1;
78 &comment("00_15 $n");
79
80 &mov($f,$c);
81
82 &mov($tmp1,$a);
83 &xor($f,$d); # F2
84
85 &rotl($tmp1,5); # A2
86
87 &and($f,$b); # F3
88 &add($tmp1,$e);
89
90 &rotr($b,1); # B1 <- F
91 &mov($e,&swtmp($n)); # G1
92
93 &rotr($b,1); # B1 <- F
94 &xor($f,$d); # F4
95
96 &lea($tmp1,&DWP($K,$tmp1,$e,1));
97
98############################
99# &BODY_40_59( 0,$K[2],$X,42,$A,$B,$C,$D,$E,$T);
100# &BODY_40_59( 0,$K[2],$X,43,$T,$A,$B,$C,$D,$E);
101$n++;
102 local($n0,$n1,$n2,$n3,$np)=&Na($n);
103 ($b,$c,$d,$e,$f,$a)=($a,$b,$c,$d,$e,$f);
104
105 &mov($f,$c);
106
107 &add($a,$tmp1); # MOVED DOWN
108 &xor($f,$d); # F2
109
110 &mov($tmp1,$a);
111 &and($f,$b); # F3
112
113 &rotl($tmp1,5); # A2
114
115 &add($tmp1,$e);
116 &mov($e,&swtmp($n)); # G1
117
118 &rotr($b,1); # B1 <- F
119 &xor($f,$d); # F4
120
121 &rotr($b,1); # B1 <- F
122 &lea($tmp1,&DWP($K,$tmp1,$e,1));
123
124 &add($f,$tmp1);
125 }
126
127sub BODY_16_19
128 {
129 local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
130 local($n0,$n1,$n2,$n3,$np)=&Na($n);
131
132return if $n & 1;
133 &comment("16_19 $n");
134
135 &nop() if ($pos < 0);
136&mov($tmp1,&swtmp($n0)); # X1
137 &mov($f,&swtmp($n1)); # X2
138&xor($f,$tmp1); # X3
139 &mov($tmp1,&swtmp($n2)); # X4
140&xor($f,$tmp1); # X5
141 &mov($tmp1,&swtmp($n3)); # X6
142&xor($f,$tmp1); # X7 - slot
143 &mov($tmp1,$c); # F1
144&rotl($f,1); # X8 - slot
145 &xor($tmp1,$d); # F2
146&mov(&swtmp($n0),$f); # X9 - anytime
147 &and($tmp1,$b); # F3
148&lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
149 &xor($tmp1,$d); # F4
150&mov($e,$a); # A1
151 &add($f,$tmp1); # tot+=F();
152
153&rotl($e,5); # A2
154
155&rotr($b,1); # B1 <- F
156 &add($f,$e); # tot+=a
157
158############################
159# &BODY_40_59( 0,$K[2],$X,42,$A,$B,$C,$D,$E,$T);
160# &BODY_40_59( 0,$K[2],$X,43,$T,$A,$B,$C,$D,$E);
161$n++;
162 local($n0,$n1,$n2,$n3,$np)=&Na($n);
163 ($b,$c,$d,$e,$f,$a)=($a,$b,$c,$d,$e,$f);
164
165
166&mov($f,&swtmp($n0)); # X1
167 &mov($tmp1,&swtmp($n1)); # X2
168&xor($f,$tmp1); # X3
169 &mov($tmp1,&swtmp($n2)); # X4
170&xor($f,$tmp1); # X5
171 &mov($tmp1,&swtmp($n3)); # X6
172&rotr($c,1); #&rotr($b,1); # B1 <- F # MOVED DOWN
173 &xor($f,$tmp1); # X7 - slot
174&rotl($f,1); # X8 - slot
175 &mov($tmp1,$c); # F1
176&xor($tmp1,$d); # F2
177 &mov(&swtmp($n0),$f); # X9 - anytime
178&and($tmp1,$b); # F3
179 &lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
180
181&xor($tmp1,$d); # F4
182 &mov($e,$a); # A1
183
184&rotl($e,5); # A2
185
186&rotr($b,1); # B1 <- F
187 &add($f,$e); # tot+=a
188
189&rotr($b,1); # B1 <- F
190 &add($f,$tmp1); # tot+=F();
191
192 }
193
194sub BODY_20_39
195 {
196 local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
197
198 &comment("20_39 $n");
199 local($n0,$n1,$n2,$n3,$np)=&Na($n);
200
201&mov($f,&swtmp($n0)); # X1
202 &mov($tmp1,&swtmp($n1)); # X2
203&xor($f,$tmp1); # X3
204 &mov($tmp1,&swtmp($n2)); # X4
205&xor($f,$tmp1); # X5
206 &mov($tmp1,&swtmp($n3)); # X6
207&xor($f,$tmp1); # X7 - slot
208 &mov($tmp1,$b); # F1
209&rotl($f,1); # X8 - slot
210 &xor($tmp1,$c); # F2
211&mov(&swtmp($n0),$f); # X9 - anytime
212 &xor($tmp1,$d); # F3
213
214&lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
215 &mov($e,$a); # A1
216
217&rotl($e,5); # A2
218
219if ($n != 79) # last loop
220 {
221 &rotr($b,1); # B1 <- F
222 &add($e,$tmp1); # tmp1=F()+a
223
224 &rotr($b,1); # B2 <- F
225 &add($f,$e); # tot+=tmp1;
226 }
227else
228 {
229 &add($e,$tmp1); # tmp1=F()+a
230 &mov($tmp1,&wparam(0));
231
232 &rotr($b,1); # B1 <- F
233 &add($f,$e); # tot+=tmp1;
234
235 &rotr($b,1); # B2 <- F
236 }
237 }
238
239sub BODY_40_59
240 {
241 local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
242
243 &comment("40_59 $n");
244 return if $n & 1;
245 local($n0,$n1,$n2,$n3,$np)=&Na($n);
246
247&mov($f,&swtmp($n0)); # X1
248 &mov($tmp1,&swtmp($n1)); # X2
249&xor($f,$tmp1); # X3
250 &mov($tmp1,&swtmp($n2)); # X4
251&xor($f,$tmp1); # X5
252 &mov($tmp1,&swtmp($n3)); # X6
253&xor($f,$tmp1); # X7 - slot
254 &mov($tmp1,$b); # F1
255&rotl($f,1); # X8 - slot
256 &or($tmp1,$c); # F2
257&mov(&swtmp($n0),$f); # X9 - anytime
258 &and($tmp1,$d); # F3
259
260&lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
261 &mov($e,$b); # F4
262
263&rotr($b,1); # B1 <- F
264 &and($e,$c); # F5
265
266&or($tmp1,$e); # F6
267 &mov($e,$a); # A1
268
269&rotl($e,5); # A2
270
271&add($tmp1,$e); # tmp1=F()+a
272
273############################
274# &BODY_40_59( 0,$K[2],$X,42,$A,$B,$C,$D,$E,$T);
275# &BODY_40_59( 0,$K[2],$X,43,$T,$A,$B,$C,$D,$E);
276$n++;
277 local($n0,$n1,$n2,$n3,$np)=&Na($n);
278 ($b,$c,$d,$e,$f,$a)=($a,$b,$c,$d,$e,$f);
279
280 &mov($f,&swtmp($n0)); # X1
281&add($a,$tmp1); # tot+=tmp1; # moved was add f,tmp1
282 &mov($tmp1,&swtmp($n1)); # X2
283&xor($f,$tmp1); # X3
284 &mov($tmp1,&swtmp($n2)); # X4
285&xor($f,$tmp1); # X5
286 &mov($tmp1,&swtmp($n3)); # X6
287&rotr($c,1); # B2 <- F # moved was rotr b,1
288 &xor($f,$tmp1); # X7 - slot
289&rotl($f,1); # X8 - slot
290 &mov($tmp1,$b); # F1
291&mov(&swtmp($n0),$f); # X9 - anytime
292 &or($tmp1,$c); # F2
293&lea($f,&DWP($K,$f,$e,1)); # tot=X+K+e
294 &mov($e,$b); # F4
295&and($tmp1,$d); # F3
296 &and($e,$c); # F5
297
298&or($tmp1,$e); # F6
299 &mov($e,$a); # A1
300
301&rotl($e,5); # A2
302
303&rotr($b,1); # B1 <- F
304 &add($tmp1,$e); # tmp1=F()+a
305
306&rotr($b,1); # B2 <- F
307 &add($f,$tmp1); # tot+=tmp1;
308 }
309
310sub BODY_60_79
311 {
312 &BODY_20_39(@_);
313 }
314
315sub sha1_block
316 {
317 local($name)=@_;
318
319 &function_begin_B($name,"");
320
321 # parameter 1 is the MD5_CTX structure.
322 # A 0
323 # B 4
324 # C 8
325 # D 12
326 # E 16
327
328 &push("esi");
329 &push("ebp");
330 &mov("eax", &wparam(2));
331 &mov("esi", &wparam(1));
332 &add("eax", "esi"); # offset to leave on
333 &mov("ebp", &wparam(0));
334 &push("ebx");
335 &sub("eax", 64);
336 &push("edi");
337 &mov($B, &DWP( 4,"ebp","",0));
338 &stack_push(18);
339 &mov($D, &DWP(12,"ebp","",0));
340 &mov($E, &DWP(16,"ebp","",0));
341 &mov($C, &DWP( 8,"ebp","",0));
342 &mov(&swtmp(17),"eax");
343
344 &comment("First we need to setup the X array");
345 &mov("eax",&DWP(0,"esi","",0)); # pulled out of X_expand
346
347 &set_label("start") unless $normal;
348
349 &X_expand("esi");
350 &mov(&swtmp(16),"esi");
351
352 &comment("");
353 &comment("Start processing");
354
355 # odd start
356 &mov($A, &DWP( 0,"ebp","",0));
357 $X="esp";
358 &BODY_00_15(-2,$K[0],$X, 0,$A,$B,$C,$D,$E,$T);
359 &BODY_00_15( 0,$K[0],$X, 1,$T,$A,$B,$C,$D,$E);
360 &BODY_00_15( 0,$K[0],$X, 2,$E,$T,$A,$B,$C,$D);
361 &BODY_00_15( 0,$K[0],$X, 3,$D,$E,$T,$A,$B,$C);
362 &BODY_00_15( 0,$K[0],$X, 4,$C,$D,$E,$T,$A,$B);
363 &BODY_00_15( 0,$K[0],$X, 5,$B,$C,$D,$E,$T,$A);
364 &BODY_00_15( 0,$K[0],$X, 6,$A,$B,$C,$D,$E,$T);
365 &BODY_00_15( 0,$K[0],$X, 7,$T,$A,$B,$C,$D,$E);
366 &BODY_00_15( 0,$K[0],$X, 8,$E,$T,$A,$B,$C,$D);
367 &BODY_00_15( 0,$K[0],$X, 9,$D,$E,$T,$A,$B,$C);
368 &BODY_00_15( 0,$K[0],$X,10,$C,$D,$E,$T,$A,$B);
369 &BODY_00_15( 0,$K[0],$X,11,$B,$C,$D,$E,$T,$A);
370 &BODY_00_15( 0,$K[0],$X,12,$A,$B,$C,$D,$E,$T);
371 &BODY_00_15( 0,$K[0],$X,13,$T,$A,$B,$C,$D,$E);
372 &BODY_00_15( 0,$K[0],$X,14,$E,$T,$A,$B,$C,$D);
373 &BODY_00_15( 1,$K[0],$X,15,$D,$E,$T,$A,$B,$C);
374 &BODY_16_19(-1,$K[0],$X,16,$C,$D,$E,$T,$A,$B);
375 &BODY_16_19( 0,$K[0],$X,17,$B,$C,$D,$E,$T,$A);
376 &BODY_16_19( 0,$K[0],$X,18,$A,$B,$C,$D,$E,$T);
377 &BODY_16_19( 1,$K[0],$X,19,$T,$A,$B,$C,$D,$E);
378
379 &BODY_20_39(-1,$K[1],$X,20,$E,$T,$A,$B,$C,$D);
380 &BODY_20_39( 0,$K[1],$X,21,$D,$E,$T,$A,$B,$C);
381 &BODY_20_39( 0,$K[1],$X,22,$C,$D,$E,$T,$A,$B);
382 &BODY_20_39( 0,$K[1],$X,23,$B,$C,$D,$E,$T,$A);
383 &BODY_20_39( 0,$K[1],$X,24,$A,$B,$C,$D,$E,$T);
384 &BODY_20_39( 0,$K[1],$X,25,$T,$A,$B,$C,$D,$E);
385 &BODY_20_39( 0,$K[1],$X,26,$E,$T,$A,$B,$C,$D);
386 &BODY_20_39( 0,$K[1],$X,27,$D,$E,$T,$A,$B,$C);
387 &BODY_20_39( 0,$K[1],$X,28,$C,$D,$E,$T,$A,$B);
388 &BODY_20_39( 0,$K[1],$X,29,$B,$C,$D,$E,$T,$A);
389 &BODY_20_39( 0,$K[1],$X,30,$A,$B,$C,$D,$E,$T);
390 &BODY_20_39( 0,$K[1],$X,31,$T,$A,$B,$C,$D,$E);
391 &BODY_20_39( 0,$K[1],$X,32,$E,$T,$A,$B,$C,$D);
392 &BODY_20_39( 0,$K[1],$X,33,$D,$E,$T,$A,$B,$C);
393 &BODY_20_39( 0,$K[1],$X,34,$C,$D,$E,$T,$A,$B);
394 &BODY_20_39( 0,$K[1],$X,35,$B,$C,$D,$E,$T,$A);
395 &BODY_20_39( 0,$K[1],$X,36,$A,$B,$C,$D,$E,$T);
396 &BODY_20_39( 0,$K[1],$X,37,$T,$A,$B,$C,$D,$E);
397 &BODY_20_39( 0,$K[1],$X,38,$E,$T,$A,$B,$C,$D);
398 &BODY_20_39( 1,$K[1],$X,39,$D,$E,$T,$A,$B,$C);
399
400 &BODY_40_59(-1,$K[2],$X,40,$C,$D,$E,$T,$A,$B);
401 &BODY_40_59( 0,$K[2],$X,41,$B,$C,$D,$E,$T,$A);
402 &BODY_40_59( 0,$K[2],$X,42,$A,$B,$C,$D,$E,$T);
403 &BODY_40_59( 0,$K[2],$X,43,$T,$A,$B,$C,$D,$E);
404 &BODY_40_59( 0,$K[2],$X,44,$E,$T,$A,$B,$C,$D);
405 &BODY_40_59( 0,$K[2],$X,45,$D,$E,$T,$A,$B,$C);
406 &BODY_40_59( 0,$K[2],$X,46,$C,$D,$E,$T,$A,$B);
407 &BODY_40_59( 0,$K[2],$X,47,$B,$C,$D,$E,$T,$A);
408 &BODY_40_59( 0,$K[2],$X,48,$A,$B,$C,$D,$E,$T);
409 &BODY_40_59( 0,$K[2],$X,49,$T,$A,$B,$C,$D,$E);
410 &BODY_40_59( 0,$K[2],$X,50,$E,$T,$A,$B,$C,$D);
411 &BODY_40_59( 0,$K[2],$X,51,$D,$E,$T,$A,$B,$C);
412 &BODY_40_59( 0,$K[2],$X,52,$C,$D,$E,$T,$A,$B);
413 &BODY_40_59( 0,$K[2],$X,53,$B,$C,$D,$E,$T,$A);
414 &BODY_40_59( 0,$K[2],$X,54,$A,$B,$C,$D,$E,$T);
415 &BODY_40_59( 0,$K[2],$X,55,$T,$A,$B,$C,$D,$E);
416 &BODY_40_59( 0,$K[2],$X,56,$E,$T,$A,$B,$C,$D);
417 &BODY_40_59( 0,$K[2],$X,57,$D,$E,$T,$A,$B,$C);
418 &BODY_40_59( 0,$K[2],$X,58,$C,$D,$E,$T,$A,$B);
419 &BODY_40_59( 1,$K[2],$X,59,$B,$C,$D,$E,$T,$A);
420
421 &BODY_60_79(-1,$K[3],$X,60,$A,$B,$C,$D,$E,$T);
422 &BODY_60_79( 0,$K[3],$X,61,$T,$A,$B,$C,$D,$E);
423 &BODY_60_79( 0,$K[3],$X,62,$E,$T,$A,$B,$C,$D);
424 &BODY_60_79( 0,$K[3],$X,63,$D,$E,$T,$A,$B,$C);
425 &BODY_60_79( 0,$K[3],$X,64,$C,$D,$E,$T,$A,$B);
426 &BODY_60_79( 0,$K[3],$X,65,$B,$C,$D,$E,$T,$A);
427 &BODY_60_79( 0,$K[3],$X,66,$A,$B,$C,$D,$E,$T);
428 &BODY_60_79( 0,$K[3],$X,67,$T,$A,$B,$C,$D,$E);
429 &BODY_60_79( 0,$K[3],$X,68,$E,$T,$A,$B,$C,$D);
430 &BODY_60_79( 0,$K[3],$X,69,$D,$E,$T,$A,$B,$C);
431 &BODY_60_79( 0,$K[3],$X,70,$C,$D,$E,$T,$A,$B);
432 &BODY_60_79( 0,$K[3],$X,71,$B,$C,$D,$E,$T,$A);
433 &BODY_60_79( 0,$K[3],$X,72,$A,$B,$C,$D,$E,$T);
434 &BODY_60_79( 0,$K[3],$X,73,$T,$A,$B,$C,$D,$E);
435 &BODY_60_79( 0,$K[3],$X,74,$E,$T,$A,$B,$C,$D);
436 &BODY_60_79( 0,$K[3],$X,75,$D,$E,$T,$A,$B,$C);
437 &BODY_60_79( 0,$K[3],$X,76,$C,$D,$E,$T,$A,$B);
438 &BODY_60_79( 0,$K[3],$X,77,$B,$C,$D,$E,$T,$A);
439 &BODY_60_79( 0,$K[3],$X,78,$A,$B,$C,$D,$E,$T);
440 &BODY_60_79( 2,$K[3],$X,79,$T,$A,$B,$C,$D,$E);
441
442 &comment("End processing");
443 &comment("");
444 # D is the tmp value
445
446 # E -> A
447 # T -> B
448 # A -> C
449 # B -> D
450 # C -> E
451 # D -> T
452
453 # The last 2 have been moved into the last loop
454 # &mov($tmp1,&wparam(0));
455
456 &mov($D, &DWP(12,$tmp1,"",0));
457 &add($D,$B);
458 &mov($B, &DWP( 4,$tmp1,"",0));
459 &add($B,$T);
460 &mov($T, $A);
461 &mov($A, &DWP( 0,$tmp1,"",0));
462 &mov(&DWP(12,$tmp1,"",0),$D);
463
464 &add($A,$E);
465 &mov($E, &DWP(16,$tmp1,"",0));
466 &add($E,$C);
467 &mov($C, &DWP( 8,$tmp1,"",0));
468 &add($C,$T);
469
470 &mov(&DWP( 0,$tmp1,"",0),$A);
471 &mov("esi",&swtmp(16));
472 &mov(&DWP( 8,$tmp1,"",0),$C); # This is for looping
473 &add("esi",64);
474 &mov("eax",&swtmp(17));
475 &mov(&DWP(16,$tmp1,"",0),$E);
476 &cmp("eax","esi");
477 &mov(&DWP( 4,$tmp1,"",0),$B); # This is for looping
478 &jl(&label("end"));
479 &mov("eax",&DWP(0,"esi","",0)); # Pulled down from
480 &jmp(&label("start"));
481
482 &set_label("end");
483 &stack_pop(18);
484 &pop("edi");
485 &pop("ebx");
486 &pop("ebp");
487 &pop("esi");
488 &ret();
489 &function_end_B($name);
490 }
491