GNU assembler (read SPARC Linux) support added.
[openssl.git] / crypto / md5 / asm / md5-sparcv9.S
1 .ident  "md5-sparcv9.S, Version 1.0"
2 .ident  "SPARC V9 ISA artwork by Andy Polyakov <appro@fy.chalmers.se>"
3 .file   "md5-sparcv9.S"
4
5 /*
6  * ====================================================================
7  * Copyright (c) 1999 Andy Polyakov <appro@fy.chalmers.se>.
8  *
9  * Rights for redistribution and usage in source and binary forms are
10  * granted as long as above copyright notices are retained. Warranty
11  * of any kind is (of course:-) disclaimed.
12  * ====================================================================
13  */
14
15 /*
16  * This is my modest contribution to OpenSSL project (see
17  * http://www.openssl.org/ for more information about it) and is an
18  * assembler implementation of MD5 block hash function. I've hand-coded
19  * this for the sole reason to reach UltraSPARC-specific "load in
20  * little-endian byte order" instruction. This gives up to 15%
21  * performance improvement for cases when input message is aligned at
22  * 32 bits boundary. The module was tested under both 32 *and* 64 bit
23  * kernels. For updates see http://fy.chalmers.se/~appro/hpe/.
24  *
25  * To compile with SC4.x/SC5.x:
26  *
27  *      cc -xarch=v[9|8plus] -DULTRASPARC -DMD5_BLOCK_DATA_ORDER \
28  *              -c md5-sparcv9.S
29  *
30  * and with gcc:
31  *
32  *      gcc -mcpu=ultrasparc -DULTRASPARC -DMD5_BLOCK_DATA_ORDER \
33  *              -c md5-sparcv9.S
34  *
35  * or if above fails (it does if you have gas):
36  *
37  *      gcc -E -DULTRASPARC -DMD5_BLOCK_DATA_ORDER md5_block.sparc.S | \
38  *              as -xarch=v8plus /dev/fd/0 -o md5-sparcv9.o
39  */
40
41 #define A       %o0
42 #define B       %o1
43 #define C       %o2
44 #define D       %o3
45 #define T1      %o4
46 #define T2      %o5
47
48 #define R0      %l0
49 #define R1      %l1
50 #define R2      %l2
51 #define R3      %l3
52 #define R4      %l4
53 #define R5      %l5
54 #define R6      %l6
55 #define R7      %l7
56 #define R8      %i3
57 #define R9      %i4
58 #define R10     %i5
59 #define R11     %g1
60 #define R12     %g2
61 #define R13     %g3
62 #define RX      %g4
63
64 #define Aptr    %i0+0
65 #define Bptr    %i0+4
66 #define Cptr    %i0+8
67 #define Dptr    %i0+12
68
69 #define Aval    R5      /* those not used at the end of the last round */
70 #define Bval    R6
71 #define Cval    R7
72 #define Dval    R8
73
74 #if defined(MD5_BLOCK_DATA_ORDER)
75 # if defined(ULTRASPARC)
76 #  define       LOAD                    lda
77 #  define       X(i)                    [%i1+i*4]%asi
78 #  define       md5_block               md5_block_asm_data_order_aligned
79 #  define       ASI_PRIMARY_LITTLE      0x88
80 # else
81 #  error "MD5_BLOCK_DATA_ORDER is supported only on UltraSPARC!"
82 # endif
83 #else
84 # define        LOAD                    ld
85 # define        X(i)                    [%i1+i*4]
86 # define        md5_block               md5_block_asm_host_order
87 #endif
88
89 .section        ".text",#alloc,#execinstr
90
91 #if defined(__SUNPRO_C) && defined(__sparcv9)
92   /* They've said -xarch=v9 at command line */
93   .register     %g2,#scratch
94   .register     %g3,#scratch
95 # define        FRAME   -192
96 #elif defined(__GNUC__) && defined(__arch64__)
97   /* They've said -m64 at command line */
98   .register     %g2,#scratch
99   .register     %g3,#scratch
100 # define        FRAME   -192
101 #else
102 # define        FRAME   -96
103 #endif
104
105 .align  32
106
107 .global md5_block
108 md5_block:
109         save    %sp,FRAME,%sp
110
111         ld      [Dptr],D
112         ld      [Cptr],C
113         ld      [Bptr],B
114         ld      [Aptr],A
115 #ifdef ASI_PRIMARY_LITTLE
116         rd      %asi,%o7        ! How dare I? Well, I just do:-)
117         wr      %g0,ASI_PRIMARY_LITTLE,%asi
118 #endif
119         LOAD    X(0),R0
120
121 .Lmd5_block_loop:
122
123 !!!!!!!!Round 0
124
125         xor     C,D,T1
126         sethi   %hi(0xd76aa478),T2
127         and     T1,B,T1
128         or      T2,%lo(0xd76aa478),T2   !=
129         xor     T1,D,T1
130         add     T1,R0,T1
131         LOAD    X(1),R1
132         add     T1,T2,T1                !=
133         add     A,T1,A
134         sll     A,7,T2
135         srl     A,32-7,A
136         or      A,T2,A                  !=
137          xor     B,C,T1
138         add     A,B,A
139
140         sethi   %hi(0xe8c7b756),T2
141         and     T1,A,T1                 !=
142         or      T2,%lo(0xe8c7b756),T2
143         xor     T1,C,T1
144         LOAD    X(2),R2
145         add     T1,R1,T1                !=
146         add     T1,T2,T1
147         add     D,T1,D
148         sll     D,12,T2
149         srl     D,32-12,D               !=
150         or      D,T2,D
151          xor     A,B,T1
152         add     D,A,D
153
154         sethi   %hi(0x242070db),T2      !=
155         and     T1,D,T1
156         or      T2,%lo(0x242070db),T2
157         xor     T1,B,T1
158         add     T1,R2,T1                !=
159         LOAD    X(3),R3
160         add     T1,T2,T1
161         add     C,T1,C
162         sll     C,17,T2                 !=
163         srl     C,32-17,C
164         or      C,T2,C
165          xor     D,A,T1
166         add     C,D,C                   !=
167
168         sethi   %hi(0xc1bdceee),T2
169         and     T1,C,T1
170         or      T2,%lo(0xc1bdceee),T2
171         xor     T1,A,T1                 !=
172         add     T1,R3,T1
173         LOAD    X(4),R4
174         add     T1,T2,T1
175         add     B,T1,B                  !=
176         sll     B,22,T2
177         srl     B,32-22,B
178         or      B,T2,B
179          xor     C,D,T1                 !=
180         add     B,C,B
181
182         sethi   %hi(0xf57c0faf),T2
183         and     T1,B,T1
184         or      T2,%lo(0xf57c0faf),T2   !=
185         xor     T1,D,T1
186         add     T1,R4,T1
187         LOAD    X(5),R5
188         add     T1,T2,T1                !=
189         add     A,T1,A
190         sll     A,7,T2
191         srl     A,32-7,A
192         or      A,T2,A                  !=
193          xor     B,C,T1
194         add     A,B,A
195
196         sethi   %hi(0x4787c62a),T2
197         and     T1,A,T1                 !=
198         or      T2,%lo(0x4787c62a),T2
199         xor     T1,C,T1
200         LOAD    X(6),R6
201         add     T1,R5,T1                !=
202         add     T1,T2,T1
203         add     D,T1,D
204         sll     D,12,T2
205         srl     D,32-12,D               !=
206         or      D,T2,D
207          xor     A,B,T1
208         add     D,A,D
209
210         sethi   %hi(0xa8304613),T2      !=
211         and     T1,D,T1
212         or      T2,%lo(0xa8304613),T2
213         xor     T1,B,T1
214         add     T1,R6,T1                !=
215         LOAD    X(7),R7
216         add     T1,T2,T1
217         add     C,T1,C
218         sll     C,17,T2                 !=
219         srl     C,32-17,C
220         or      C,T2,C
221          xor     D,A,T1
222         add     C,D,C                   !=
223
224         sethi   %hi(0xfd469501),T2
225         and     T1,C,T1
226         or      T2,%lo(0xfd469501),T2
227         xor     T1,A,T1                 !=
228         add     T1,R7,T1
229         LOAD    X(8),R8
230         add     T1,T2,T1
231         add     B,T1,B                  !=
232         sll     B,22,T2
233         srl     B,32-22,B
234         or      B,T2,B
235          xor     C,D,T1                 !=
236         add     B,C,B
237
238         sethi   %hi(0x698098d8),T2
239         and     T1,B,T1
240         or      T2,%lo(0x698098d8),T2   !=
241         xor     T1,D,T1
242         add     T1,R8,T1
243         LOAD    X(9),R9
244         add     T1,T2,T1                !=
245         add     A,T1,A
246         sll     A,7,T2
247         srl     A,32-7,A
248         or      A,T2,A                  !=
249          xor     B,C,T1
250         add     A,B,A
251
252         sethi   %hi(0x8b44f7af),T2
253         and     T1,A,T1                 !=
254         or      T2,%lo(0x8b44f7af),T2
255         xor     T1,C,T1
256         LOAD    X(10),R10
257         add     T1,R9,T1                !=
258         add     T1,T2,T1
259         add     D,T1,D
260         sll     D,12,T2
261         srl     D,32-12,D               !=
262         or      D,T2,D
263          xor     A,B,T1
264         add     D,A,D
265
266         sethi   %hi(0xffff5bb1),T2      !=
267         and     T1,D,T1
268         or      T2,%lo(0xffff5bb1),T2
269         xor     T1,B,T1
270         add     T1,R10,T1               !=
271         LOAD    X(11),R11
272         add     T1,T2,T1
273         add     C,T1,C
274         sll     C,17,T2                 !=
275         srl     C,32-17,C
276         or      C,T2,C
277          xor     D,A,T1
278         add     C,D,C                   !=
279
280         sethi   %hi(0x895cd7be),T2
281         and     T1,C,T1
282         or      T2,%lo(0x895cd7be),T2
283         xor     T1,A,T1                 !=
284         add     T1,R11,T1
285         LOAD    X(12),R12
286         add     T1,T2,T1
287         add     B,T1,B                  !=
288         sll     B,22,T2
289         srl     B,32-22,B
290         or      B,T2,B
291          xor     C,D,T1                 !=
292         add     B,C,B
293
294         sethi   %hi(0x6b901122),T2
295         and     T1,B,T1
296         or      T2,%lo(0x6b901122),T2   !=
297         xor     T1,D,T1
298         add     T1,R12,T1
299         LOAD    X(13),R13
300         add     T1,T2,T1                !=
301         add     A,T1,A
302         sll     A,7,T2
303         srl     A,32-7,A
304         or      A,T2,A                  !=
305          xor     B,C,T1
306         add     A,B,A
307
308         sethi   %hi(0xfd987193),T2
309         and     T1,A,T1                 !=
310         or      T2,%lo(0xfd987193),T2
311         xor     T1,C,T1
312         LOAD    X(14),RX
313         add     T1,R13,T1               !=
314         add     T1,T2,T1
315         add     D,T1,D
316         sll     D,12,T2
317         srl     D,32-12,D               !=
318         or      D,T2,D
319          xor     A,B,T1
320         add     D,A,D
321
322         sethi   %hi(0xa679438e),T2      !=
323         and     T1,D,T1
324         or      T2,%lo(0xa679438e),T2
325         xor     T1,B,T1
326         add     T1,RX,T1                !=
327         LOAD    X(15),RX
328         add     T1,T2,T1
329         add     C,T1,C
330         sll     C,17,T2                 !=
331         srl     C,32-17,C
332         or      C,T2,C
333          xor     D,A,T1
334         add     C,D,C                   !=
335
336         sethi   %hi(0x49b40821),T2
337         and     T1,C,T1
338         or      T2,%lo(0x49b40821),T2
339         xor     T1,A,T1                 !=
340         add     T1,RX,T1
341         !pre-LOADed     X(1),R1
342         add     T1,T2,T1
343         add     B,T1,B
344         sll     B,22,T2                 !=
345         srl     B,32-22,B
346         or      B,T2,B
347         add     B,C,B
348
349 !!!!!!!!Round 1
350
351         xor     B,C,T1                  !=
352         sethi   %hi(0xf61e2562),T2
353         and     T1,D,T1
354         or      T2,%lo(0xf61e2562),T2
355         xor     T1,C,T1                 !=
356         add     T1,R1,T1
357         !pre-LOADed     X(6),R6
358         add     T1,T2,T1
359         add     A,T1,A
360         sll     A,5,T2                  !=
361         srl     A,32-5,A
362         or      A,T2,A
363         add     A,B,A
364
365         xor     A,B,T1                  !=
366         sethi   %hi(0xc040b340),T2
367         and     T1,C,T1
368         or      T2,%lo(0xc040b340),T2
369         xor     T1,B,T1                 !=
370         add     T1,R6,T1
371         !pre-LOADed     X(11),R11
372         add     T1,T2,T1
373         add     D,T1,D
374         sll     D,9,T2                  !=
375         srl     D,32-9,D
376         or      D,T2,D
377         add     D,A,D
378
379         xor     D,A,T1                  !=
380         sethi   %hi(0x265e5a51),T2
381         and     T1,B,T1
382         or      T2,%lo(0x265e5a51),T2
383         xor     T1,A,T1                 !=
384         add     T1,R11,T1
385         !pre-LOADed     X(0),R0
386         add     T1,T2,T1
387         add     C,T1,C
388         sll     C,14,T2                 !=
389         srl     C,32-14,C
390         or      C,T2,C
391         add     C,D,C
392
393         xor     C,D,T1                  !=
394         sethi   %hi(0xe9b6c7aa),T2
395         and     T1,A,T1
396         or      T2,%lo(0xe9b6c7aa),T2
397         xor     T1,D,T1                 !=
398         add     T1,R0,T1
399         !pre-LOADed     X(5),R5
400         add     T1,T2,T1
401         add     B,T1,B
402         sll     B,20,T2                 !=
403         srl     B,32-20,B
404         or      B,T2,B
405         add     B,C,B
406
407         xor     B,C,T1                  !=
408         sethi   %hi(0xd62f105d),T2
409         and     T1,D,T1
410         or      T2,%lo(0xd62f105d),T2
411         xor     T1,C,T1                 !=
412         add     T1,R5,T1
413         !pre-LOADed     X(10),R10
414         add     T1,T2,T1
415         add     A,T1,A
416         sll     A,5,T2                  !=
417         srl     A,32-5,A
418         or      A,T2,A
419         add     A,B,A
420
421         xor     A,B,T1                  !=
422         sethi   %hi(0x02441453),T2
423         and     T1,C,T1
424         or      T2,%lo(0x02441453),T2
425         xor     T1,B,T1                 !=
426         add     T1,R10,T1
427         LOAD    X(15),RX
428         add     T1,T2,T1
429         add     D,T1,D                  !=
430         sll     D,9,T2
431         srl     D,32-9,D
432         or      D,T2,D
433         add     D,A,D                   !=
434
435         xor     D,A,T1
436         sethi   %hi(0xd8a1e681),T2
437         and     T1,B,T1
438         or      T2,%lo(0xd8a1e681),T2   !=
439         xor     T1,A,T1
440         add     T1,RX,T1
441         !pre-LOADed     X(4),R4
442         add     T1,T2,T1
443         add     C,T1,C                  !=
444         sll     C,14,T2
445         srl     C,32-14,C
446         or      C,T2,C
447         add     C,D,C                   !=
448
449         xor     C,D,T1
450         sethi   %hi(0xe7d3fbc8),T2
451         and     T1,A,T1
452         or      T2,%lo(0xe7d3fbc8),T2   !=
453         xor     T1,D,T1
454         add     T1,R4,T1
455         !pre-LOADed     X(9),R9
456         add     T1,T2,T1
457         add     B,T1,B                  !=
458         sll     B,20,T2
459         srl     B,32-20,B
460         or      B,T2,B
461         add     B,C,B                   !=
462
463         xor     B,C,T1
464         sethi   %hi(0x21e1cde6),T2
465         and     T1,D,T1
466         or      T2,%lo(0x21e1cde6),T2   !=
467         xor     T1,C,T1
468         add     T1,R9,T1
469         LOAD    X(14),RX
470         add     T1,T2,T1                !=
471         add     A,T1,A
472         sll     A,5,T2
473         srl     A,32-5,A
474         or      A,T2,A                  !=
475         add     A,B,A
476
477         xor     A,B,T1
478         sethi   %hi(0xc33707d6),T2
479         and     T1,C,T1                 !=
480         or      T2,%lo(0xc33707d6),T2
481         xor     T1,B,T1
482         add     T1,RX,T1
483         !pre-LOADed     X(3),R3
484         add     T1,T2,T1                !=
485         add     D,T1,D
486         sll     D,9,T2
487         srl     D,32-9,D
488         or      D,T2,D                  !=
489         add     D,A,D
490
491         xor     D,A,T1
492         sethi   %hi(0xf4d50d87),T2
493         and     T1,B,T1                 !=
494         or      T2,%lo(0xf4d50d87),T2
495         xor     T1,A,T1
496         add     T1,R3,T1
497         !pre-LOADed     X(8),R8
498         add     T1,T2,T1                !=
499         add     C,T1,C
500         sll     C,14,T2
501         srl     C,32-14,C
502         or      C,T2,C                  !=
503         add     C,D,C
504
505         xor     C,D,T1
506         sethi   %hi(0x455a14ed),T2
507         and     T1,A,T1                 !=
508         or      T2,%lo(0x455a14ed),T2
509         xor     T1,D,T1
510         add     T1,R8,T1
511         !pre-LOADed     X(13),R13
512         add     T1,T2,T1                !=
513         add     B,T1,B
514         sll     B,20,T2
515         srl     B,32-20,B
516         or      B,T2,B                  !=
517         add     B,C,B
518
519         xor     B,C,T1
520         sethi   %hi(0xa9e3e905),T2
521         and     T1,D,T1                 !=
522         or      T2,%lo(0xa9e3e905),T2
523         xor     T1,C,T1
524         add     T1,R13,T1
525         !pre-LOADed     X(2),R2
526         add     T1,T2,T1                !=
527         add     A,T1,A
528         sll     A,5,T2
529         srl     A,32-5,A
530         or      A,T2,A                  !=
531         add     A,B,A
532
533         xor     A,B,T1
534         sethi   %hi(0xfcefa3f8),T2
535         and     T1,C,T1                 !=
536         or      T2,%lo(0xfcefa3f8),T2
537         xor     T1,B,T1
538         add     T1,R2,T1
539         !pre-LOADed     X(7),R7
540         add     T1,T2,T1                !=
541         add     D,T1,D
542         sll     D,9,T2
543         srl     D,32-9,D
544         or      D,T2,D                  !=
545         add     D,A,D
546
547         xor     D,A,T1
548         sethi   %hi(0x676f02d9),T2
549         and     T1,B,T1                 !=
550         or      T2,%lo(0x676f02d9),T2
551         xor     T1,A,T1
552         add     T1,R7,T1
553         !pre-LOADed     X(12),R12
554         add     T1,T2,T1                !=
555         add     C,T1,C
556         sll     C,14,T2
557         srl     C,32-14,C
558         or      C,T2,C                  !=
559         add     C,D,C
560
561         xor     C,D,T1
562         sethi   %hi(0x8d2a4c8a),T2
563         and     T1,A,T1                 !=
564         or      T2,%lo(0x8d2a4c8a),T2
565         xor     T1,D,T1
566         add     T1,R12,T1
567         !pre-LOADed     X(5),R5
568         add     T1,T2,T1                !=
569         add     B,T1,B
570         sll     B,20,T2
571         srl     B,32-20,B
572         or      B,T2,B                  !=
573         add     B,C,B
574
575 !!!!!!!!Round 2
576
577         xor     B,C,T1
578         sethi   %hi(0xfffa3942),T2
579         xor     T1,D,T1                 !=
580         or      T2,%lo(0xfffa3942),T2
581         add     T1,R5,T1
582         !pre-LOADed     X(8),R8
583         add     T1,T2,T1
584         add     A,T1,A                  !=
585         sll     A,4,T2
586         srl     A,32-4,A
587         or      A,T2,A
588         add     A,B,A                   !=
589
590         xor     A,B,T1
591         sethi   %hi(0x8771f681),T2
592         xor     T1,C,T1
593         or      T2,%lo(0x8771f681),T2   !=
594         add     T1,R8,T1
595         !pre-LOADed     X(11),R11
596         add     T1,T2,T1
597         add     D,T1,D
598         sll     D,11,T2                 !=
599         srl     D,32-11,D
600         or      D,T2,D
601         add     D,A,D
602
603         xor     D,A,T1                  !=
604         sethi   %hi(0x6d9d6122),T2
605         xor     T1,B,T1
606         or      T2,%lo(0x6d9d6122),T2
607         add     T1,R11,T1               !=
608         LOAD    X(14),RX
609         add     T1,T2,T1
610         add     C,T1,C
611         sll     C,16,T2                 !=
612         srl     C,32-16,C
613         or      C,T2,C
614         add     C,D,C
615
616         xor     C,D,T1                  !=
617         sethi   %hi(0xfde5380c),T2
618         xor     T1,A,T1
619         or      T2,%lo(0xfde5380c),T2
620         add     T1,RX,T1                !=
621         !pre-LOADed     X(1),R1
622         add     T1,T2,T1
623         add     B,T1,B
624         sll     B,23,T2
625         srl     B,32-23,B               !=
626         or      B,T2,B
627         add     B,C,B
628
629         xor     B,C,T1
630         sethi   %hi(0xa4beea44),T2      !=
631         xor     T1,D,T1
632         or      T2,%lo(0xa4beea44),T2
633         add     T1,R1,T1
634         !pre-LOADed     X(4),R4
635         add     T1,T2,T1                !=
636         add     A,T1,A
637         sll     A,4,T2
638         srl     A,32-4,A
639         or      A,T2,A                  !=
640         add     A,B,A
641
642         xor     A,B,T1
643         sethi   %hi(0x4bdecfa9),T2
644         xor     T1,C,T1                 !=
645         or      T2,%lo(0x4bdecfa9),T2
646         add     T1,R4,T1
647         !pre-LOADed     X(7),R7
648         add     T1,T2,T1
649         add     D,T1,D                  !=
650         sll     D,11,T2
651         srl     D,32-11,D
652         or      D,T2,D
653         add     D,A,D                   !=
654
655         xor     D,A,T1
656         sethi   %hi(0xf6bb4b60),T2
657         xor     T1,B,T1
658         or      T2,%lo(0xf6bb4b60),T2   !=
659         add     T1,R7,T1
660         !pre-LOADed     X(10),R10
661         add     T1,T2,T1
662         add     C,T1,C
663         sll     C,16,T2                 !=
664         srl     C,32-16,C
665         or      C,T2,C
666         add     C,D,C
667
668         xor     C,D,T1                  !=
669         sethi   %hi(0xbebfbc70),T2
670         xor     T1,A,T1
671         or      T2,%lo(0xbebfbc70),T2
672         add     T1,R10,T1               !=
673         !pre-LOADed     X(13),R13
674         add     T1,T2,T1
675         add     B,T1,B
676         sll     B,23,T2
677         srl     B,32-23,B               !=
678         or      B,T2,B
679         add     B,C,B
680
681         xor     B,C,T1
682         sethi   %hi(0x289b7ec6),T2      !=
683         xor     T1,D,T1
684         or      T2,%lo(0x289b7ec6),T2
685         add     T1,R13,T1
686         !pre-LOADed     X(0),R0
687         add     T1,T2,T1                !=
688         add     A,T1,A
689         sll     A,4,T2
690         srl     A,32-4,A
691         or      A,T2,A                  !=
692         add     A,B,A
693
694         xor     A,B,T1
695         sethi   %hi(0xeaa127fa),T2
696         xor     T1,C,T1                 !=
697         or      T2,%lo(0xeaa127fa),T2
698         add     T1,R0,T1
699         !pre-LOADed     X(3),R3
700         add     T1,T2,T1
701         add     D,T1,D                  !=
702         sll     D,11,T2
703         srl     D,32-11,D
704         or      D,T2,D
705         add     D,A,D                   !=
706
707         xor     D,A,T1
708         sethi   %hi(0xd4ef3085),T2
709         xor     T1,B,T1
710         or      T2,%lo(0xd4ef3085),T2   !=
711         add     T1,R3,T1
712         !pre-LOADed     X(6),R6
713         add     T1,T2,T1
714         add     C,T1,C
715         sll     C,16,T2                 !=
716         srl     C,32-16,C
717         or      C,T2,C
718         add     C,D,C
719
720         xor     C,D,T1                  !=
721         sethi   %hi(0x04881d05),T2
722         xor     T1,A,T1
723         or      T2,%lo(0x04881d05),T2
724         add     T1,R6,T1                !=
725         !pre-LOADed     X(9),R9
726         add     T1,T2,T1
727         add     B,T1,B
728         sll     B,23,T2
729         srl     B,32-23,B               !=
730         or      B,T2,B
731         add     B,C,B
732
733         xor     B,C,T1
734         sethi   %hi(0xd9d4d039),T2      !=
735         xor     T1,D,T1
736         or      T2,%lo(0xd9d4d039),T2
737         add     T1,R9,T1
738         !pre-LOADed     X(12),R12
739         add     T1,T2,T1                !=
740         add     A,T1,A
741         sll     A,4,T2
742         srl     A,32-4,A
743         or      A,T2,A                  !=
744         add     A,B,A
745
746         xor     A,B,T1
747         sethi   %hi(0xe6db99e5),T2
748         xor     T1,C,T1                 !=
749         or      T2,%lo(0xe6db99e5),T2
750         add     T1,R12,T1
751         LOAD    X(15),RX
752         add     T1,T2,T1                !=
753         add     D,T1,D
754         sll     D,11,T2
755         srl     D,32-11,D
756         or      D,T2,D                  !=
757         add     D,A,D
758
759         xor     D,A,T1
760         sethi   %hi(0x1fa27cf8),T2
761         xor     T1,B,T1                 !=
762         or      T2,%lo(0x1fa27cf8),T2
763         add     T1,RX,T1
764         !pre-LOADed     X(2),R2
765         add     T1,T2,T1
766         add     C,T1,C                  !=
767         sll     C,16,T2
768         srl     C,32-16,C
769         or      C,T2,C
770         add     C,D,C                   !=
771
772         xor     C,D,T1
773         sethi   %hi(0xc4ac5665),T2
774         xor     T1,A,T1
775         or      T2,%lo(0xc4ac5665),T2   !=
776         add     T1,R2,T1
777         !pre-LOADed     X(0),R0
778         add     T1,T2,T1
779         add     B,T1,B
780         sll     B,23,T2                 !=
781         srl     B,32-23,B
782         or      B,T2,B
783         add     B,C,B
784
785 !!!!!!!!Round 3
786
787         orn     B,D,T1                  !=
788         sethi   %hi(0xf4292244),T2
789         xor     T1,C,T1
790         or      T2,%lo(0xf4292244),T2
791         add     T1,R0,T1                !=
792         !pre-LOADed     X(7),R7
793         add     T1,T2,T1
794         add     A,T1,A
795         sll     A,6,T2
796         srl     A,32-6,A                !=
797         or      A,T2,A
798         add     A,B,A
799
800         orn     A,C,T1
801         sethi   %hi(0x432aff97),T2      !=
802         xor     T1,B,T1
803         or      T2,%lo(0x432aff97),T2
804         LOAD    X(14),RX
805         add     T1,R7,T1                !=
806         add     T1,T2,T1
807         add     D,T1,D
808         sll     D,10,T2
809         srl     D,32-10,D               !=
810         or      D,T2,D
811         add     D,A,D
812
813         orn     D,B,T1
814         sethi   %hi(0xab9423a7),T2      !=
815         xor     T1,A,T1
816         or      T2,%lo(0xab9423a7),T2
817         add     T1,RX,T1
818         !pre-LOADed     X(5),R5
819         add     T1,T2,T1                !=
820         add     C,T1,C
821         sll     C,15,T2
822         srl     C,32-15,C
823         or      C,T2,C                  !=
824         add     C,D,C
825
826         orn     C,A,T1
827         sethi   %hi(0xfc93a039),T2
828         xor     T1,D,T1                 !=
829         or      T2,%lo(0xfc93a039),T2
830         add     T1,R5,T1
831         !pre-LOADed     X(12),R12
832         add     T1,T2,T1
833         add     B,T1,B                  !=
834         sll     B,21,T2
835         srl     B,32-21,B
836         or      B,T2,B
837         add     B,C,B                   !=
838
839         orn     B,D,T1
840         sethi   %hi(0x655b59c3),T2
841         xor     T1,C,T1
842         or      T2,%lo(0x655b59c3),T2   !=
843         add     T1,R12,T1
844         !pre-LOADed     X(3),R3
845         add     T1,T2,T1
846         add     A,T1,A
847         sll     A,6,T2                  !=
848         srl     A,32-6,A
849         or      A,T2,A
850         add     A,B,A
851
852         orn     A,C,T1                  !=
853         sethi   %hi(0x8f0ccc92),T2
854         xor     T1,B,T1
855         or      T2,%lo(0x8f0ccc92),T2
856         add     T1,R3,T1                !=
857         !pre-LOADed     X(10),R10
858         add     T1,T2,T1
859         add     D,T1,D
860         sll     D,10,T2
861         srl     D,32-10,D               !=
862         or      D,T2,D
863         add     D,A,D
864
865         orn     D,B,T1
866         sethi   %hi(0xffeff47d),T2      !=
867         xor     T1,A,T1
868         or      T2,%lo(0xffeff47d),T2
869         add     T1,R10,T1
870         !pre-LOADed     X(1),R1
871         add     T1,T2,T1                !=
872         add     C,T1,C
873         sll     C,15,T2
874         srl     C,32-15,C
875         or      C,T2,C                  !=
876         add     C,D,C
877
878         orn     C,A,T1
879         sethi   %hi(0x85845dd1),T2
880         xor     T1,D,T1                 !=
881         or      T2,%lo(0x85845dd1),T2
882         add     T1,R1,T1
883         !pre-LOADed     X(8),R8
884         add     T1,T2,T1
885         add     B,T1,B                  !=
886         sll     B,21,T2
887         srl     B,32-21,B
888         or      B,T2,B
889         add     B,C,B                   !=
890
891         orn     B,D,T1
892         sethi   %hi(0x6fa87e4f),T2
893         xor     T1,C,T1
894         or      T2,%lo(0x6fa87e4f),T2   !=
895         add     T1,R8,T1
896         LOAD    X(15),RX
897         add     T1,T2,T1
898         add     A,T1,A                  !=
899         sll     A,6,T2
900         srl     A,32-6,A
901         or      A,T2,A
902         add     A,B,A                   !=
903
904         orn     A,C,T1
905         sethi   %hi(0xfe2ce6e0),T2
906         xor     T1,B,T1
907         or      T2,%lo(0xfe2ce6e0),T2   !=
908         add     T1,RX,T1
909         !pre-LOADed     X(6),R6
910         add     T1,T2,T1
911         add     D,T1,D
912         sll     D,10,T2                 !=
913         srl     D,32-10,D
914         or      D,T2,D
915         add     D,A,D
916
917         orn     D,B,T1                  !=
918         sethi   %hi(0xa3014314),T2
919         xor     T1,A,T1
920         or      T2,%lo(0xa3014314),T2
921         add     T1,R6,T1                !=
922         !pre-LOADed     X(13),R13
923         add     T1,T2,T1
924         add     C,T1,C
925         sll     C,15,T2
926         srl     C,32-15,C               !=
927         or      C,T2,C
928         add     C,D,C
929
930         orn     C,A,T1
931         sethi   %hi(0x4e0811a1),T2      !=
932         xor     T1,D,T1
933         or      T2,%lo(0x4e0811a1),T2
934         !pre-LOADed     X(4),R4
935          ld      [Aptr],Aval
936         add     T1,R13,T1               !=
937         add     T1,T2,T1
938         add     B,T1,B
939         sll     B,21,T2
940         srl     B,32-21,B               !=
941         or      B,T2,B
942         add     B,C,B
943
944         orn     B,D,T1
945         sethi   %hi(0xf7537e82),T2      !=
946         xor     T1,C,T1
947         or      T2,%lo(0xf7537e82),T2
948         !pre-LOADed     X(11),R11
949          ld      [Dptr],Dval
950         add     T1,R4,T1                !=
951         add     T1,T2,T1
952         add     A,T1,A
953         sll     A,6,T2
954         srl     A,32-6,A                !=
955         or      A,T2,A
956         add     A,B,A
957
958         orn     A,C,T1
959         sethi   %hi(0xbd3af235),T2      !=
960         xor     T1,B,T1
961         or      T2,%lo(0xbd3af235),T2
962         !pre-LOADed     X(2),R2
963          ld      [Cptr],Cval
964         add     T1,R11,T1               !=
965         add     T1,T2,T1
966         add     D,T1,D
967         sll     D,10,T2
968         srl     D,32-10,D               !=
969         or      D,T2,D
970         add     D,A,D
971
972         orn     D,B,T1
973         sethi   %hi(0x2ad7d2bb),T2      !=
974         xor     T1,A,T1
975         or      T2,%lo(0x2ad7d2bb),T2
976         !pre-LOADed     X(9),R9
977          ld      [Bptr],Bval
978         add     T1,R2,T1                !=
979          add     Aval,A,Aval
980         add     T1,T2,T1
981          st      Aval,[Aptr]
982         add     C,T1,C                  !=
983         sll     C,15,T2
984          add     Dval,D,Dval
985         srl     C,32-15,C
986         or      C,T2,C                  !=
987          st      Dval,[Dptr]
988         add     C,D,C
989
990         orn     C,A,T1
991         sethi   %hi(0xeb86d391),T2      !=
992         xor     T1,D,T1
993         or      T2,%lo(0xeb86d391),T2
994         add     T1,R9,T1
995         !pre-LOADed     X(0),R0
996          mov     Aval,A                 !=
997         add     T1,T2,T1
998          mov     Dval,D
999         add     B,T1,B
1000         sll     B,21,T2                 !=
1001          add     Cval,C,Cval
1002         srl     B,32-21,B
1003          st      Cval,[Cptr]
1004         or      B,T2,B                  !=
1005         add     B,C,B
1006
1007         deccc   %i2
1008         mov     Cval,C
1009         add     B,Bval,B                !=
1010         inc     64,%i1
1011         nop
1012         st      B,[Bptr]
1013         nop                             !=
1014
1015 #ifdef  ULTRASPARC
1016         bg,a,pt %icc,.Lmd5_block_loop
1017 #else
1018         bg,a    .Lmd5_block_loop
1019 #endif
1020         LOAD    X(0),R0
1021
1022 #ifdef ASI_PRIMARY_LITTLE
1023         wr      %g0,%o7,%asi
1024 #endif
1025         ret
1026         restore %g0,0,%o0
1027
1028 .type   md5_block,#function
1029 .size   md5_block,(.-md5_block)