3 # ====================================================================
4 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
5 # project. The module is, however, dual licensed under OpenSSL and
6 # CRYPTOGAMS licenses depending on where you obtain it. For further
7 # details see http://www.openssl.org/~appro/cryptogams/.
8 # ====================================================================
10 # SHA512 block transform for x86. September 2007.
12 # Performance in clock cycles per processed byte (less is better):
14 # Pentium PIII P4 AMD K8 Core2
15 # gcc 100 75 116 54 66
17 # x86 asm 61 56 82 36 40
18 # SSE2 asm - - 38 24 20
19 # x86_64 asm(*) - - 30 10.0 10.5
21 # (*) x86_64 assembler performance is presented for reference
24 # IALU code-path is optimized for elder Pentiums. On vanilla Pentium
25 # performance improvement over compiler generated code reaches ~60%,
26 # while on PIII - ~35%. On newer ยต-archs improvement varies from 15%
27 # to 50%, but it's less important as they are expected to execute SSE2
28 # code-path, which is commonly ~2-3x faster [than compiler generated
29 # code]. SSE2 code-path is as fast as original sha512-sse2.pl, even
30 # though it does not use 128-bit operations. The latter means that
31 # SSE2-aware kernel is no longer required to execute the code. Another
32 # difference is that new code optimizes amount of writes, but at the
33 # cost of increased data cache "footprint" by 1/2KB.
35 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
36 push(@INC,"${dir}","${dir}../../perlasm");
39 &asm_init($ARGV[0],"sha512-586.pl",$ARGV[$#ARGV] eq "386");
42 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
44 &external_label("OPENSSL_ia32cap_P") if ($sse2);
46 $Tlo=&DWP(0,"esp"); $Thi=&DWP(4,"esp");
47 $Alo=&DWP(8,"esp"); $Ahi=&DWP(8+4,"esp");
48 $Blo=&DWP(16,"esp"); $Bhi=&DWP(16+4,"esp");
49 $Clo=&DWP(24,"esp"); $Chi=&DWP(24+4,"esp");
50 $Dlo=&DWP(32,"esp"); $Dhi=&DWP(32+4,"esp");
51 $Elo=&DWP(40,"esp"); $Ehi=&DWP(40+4,"esp");
52 $Flo=&DWP(48,"esp"); $Fhi=&DWP(48+4,"esp");
53 $Glo=&DWP(56,"esp"); $Ghi=&DWP(56+4,"esp");
54 $Hlo=&DWP(64,"esp"); $Hhi=&DWP(64+4,"esp");
59 $Csse2=&QWP(16,"esp");
60 $Dsse2=&QWP(24,"esp");
61 $Esse2=&QWP(32,"esp");
62 $Fsse2=&QWP(40,"esp");
63 $Gsse2=&QWP(48,"esp");
64 $Hsse2=&QWP(56,"esp");
67 $E="mm4"; # F-H are commonly loaded to respectively mm1-mm3 and
68 # mm5-mm7, but it's done on on-demand basis...
73 &movq ("mm5",$Fsse2); # load f
74 &movq ("mm6",$Gsse2); # load g
75 &movq ("mm7",$Hsse2); # load h
77 &movq ("mm1",$E); # %mm1 is sliding right
78 &movq ("mm2",$E); # %mm2 is sliding left
80 &movq ($Esse2,$E); # modulo-scheduled save e
82 &movq ("mm3","mm1"); # %mm3 is T1
91 &paddq ("mm7",QWP(0,$K512)); # h+=K512[i]
92 &pxor ("mm3","mm2"); # T1=Sigma1_512(e)
94 &pxor ("mm5","mm6"); # f^=g
95 &movq ("mm1",$Bsse2); # load b
96 &pand ("mm5",$E); # f&=e
97 &movq ("mm2",$Csse2); # load c
98 &pxor ("mm5","mm6"); # f^=g
99 &movq ($E,$Dsse2); # e = load d
100 &paddq ("mm3","mm5"); # T1+=Ch(e,f,g)
101 &movq (&QWP(0,"esp"),$A); # modulo-scheduled save a
102 &paddq ("mm3","mm7"); # T1+=h
104 &movq ("mm5",$A); # %mm5 is sliding right
105 &movq ("mm6",$A); # %mm6 is sliding left
106 &paddq ("mm3",&QWP(8*9,"esp")); # T1+=X[0]
108 &paddq ($E,"mm3"); # e += T1
110 &movq ("mm7","mm5"); # %mm7 is T2
120 &pxor ("mm7","mm6"); # T2=Sigma0_512(a)
122 &movq ("mm5",$A); # %mm5=a
123 &por ($A,"mm2"); # a=a|c
124 &movq ("mm6",&QWP(8*(9+16-14),"esp")) if ($prefetch);
125 &pand ("mm5","mm2"); # %mm5=a&c
126 &pand ($A,"mm1"); # a=(a|c)&b
127 &movq ("mm2",&QWP(8*(9+16-1),"esp")) if ($prefetch);
128 &por ("mm5",$A); # %mm5=(a&c)|((a|c)&b)
129 &paddq ("mm7","mm5"); # T2+=Maj(a,b,c)
130 &movq ($A,"mm3"); # a=T1
132 &mov (&LB("edx"),&BP(0,$K512));
133 &paddq ($A,"mm7"); # a+=T2
138 #define Sigma1(x) (ROTR((x),14) ^ ROTR((x),18) ^ ROTR((x),41))
139 # LO lo>>14^hi<<18 ^ lo>>18^hi<<14 ^ hi>>9^lo<<23
140 # HI hi>>14^lo<<18 ^ hi>>18^lo<<14 ^ lo>>9^hi<<23
145 &shr ("ecx",9) # lo>>9
147 &shr ("edx",9) # hi>>9
149 &shl ("esi",14); # lo<<14
151 &shl ("edi",14); # hi<<14
154 &shr ("ecx",14-9); # lo>>14
156 &shr ("edx",14-9); # hi>>14
158 &shl ("esi",18-14); # lo<<18
160 &shl ("edi",18-14); # hi<<18
163 &shr ("ecx",18-14); # lo>>18
165 &shr ("edx",18-14); # hi>>18
167 &shl ("esi",23-18); # lo<<23
169 &shl ("edi",23-18); # hi<<23
171 &xor ("ebx","edi"); # T1 = Sigma1(e)
178 &adc ("ebx",$Hhi); # T1 += h
183 &add ("eax",&DWP(8*(9+15)+0,"esp"));
184 &adc ("ebx",&DWP(8*(9+15)+4,"esp")); # T1 += X[0]
186 &xor ("edx","edi"); # Ch(e,f,g) = (f^g)&e)^g
188 &mov ("esi",&DWP(0,$K512));
189 &mov ("edi",&DWP(4,$K512)); # K[i]
191 &adc ("ebx","edx"); # T1 += Ch(e,f,g)
195 &adc ("ebx","edi"); # T1 += K[i]
197 &mov ($Thi,"ebx"); # put T1 away
199 &adc ("ebx","edx"); # d += T1
201 #define Sigma0(x) (ROTR((x),28) ^ ROTR((x),34) ^ ROTR((x),39))
202 # LO lo>>28^hi<<4 ^ hi>>2^lo<<30 ^ hi>>7^lo<<25
203 # HI hi>>28^lo<<4 ^ lo>>2^hi<<30 ^ lo>>7^hi<<25
210 &shr ("ecx",2) # lo>>2
212 &shr ("edx",2) # hi>>2
214 &shl ("esi",4); # lo<<4
216 &shl ("edi",4); # hi<<4
219 &shr ("ecx",7-2); # lo>>7
221 &shr ("edx",7-2); # hi>>7
223 &shl ("esi",25-4); # lo<<25
225 &shl ("edi",25-4); # hi<<25
228 &shr ("ecx",28-7); # lo>>28
230 &shr ("edx",28-7); # hi>>28
232 &shl ("esi",30-25); # lo<<30
234 &shl ("edi",30-25); # hi<<30
236 &xor ("ebx","edi"); # Sigma0(a)
243 &adc ("ebx",$Thi); # T1 = Sigma0(a)+T1
251 &or ("edx","edi"); # Maj(a,b,c) = ((a|b)&c)|(a&b)
254 &adc ("ebx","edx"); # T1 += Maj(a,b,c)
258 &mov (&LB("edx"),&BP(0,$K512)); # pre-fetch LSB of *K
260 &lea ($K512,&DWP(8,$K512)); # K++
264 &function_begin("sha512_block_data_order");
265 &mov ("esi",wparam(0)); # ctx
266 &mov ("edi",wparam(1)); # inp
267 &mov ("eax",wparam(2)); # num
268 &mov ("ebx","esp"); # saved sp
270 &call (&label("pic_point")); # make it PIC!
271 &set_label("pic_point");
273 &lea ($K512,&DWP(&label("K512")."-".&label("pic_point"),$K512));
280 &mov (&DWP(0,"esp"),"esi"); # ctx
281 &mov (&DWP(4,"esp"),"edi"); # inp
282 &mov (&DWP(8,"esp"),"eax"); # inp+num*128
283 &mov (&DWP(12,"esp"),"ebx"); # saved sp
286 &picmeup("edx","OPENSSL_ia32cap_P",$K512,&label("K512"));
287 &bt (&DWP(0,"edx"),26);
288 &jnc (&label("loop_x86"));
291 &movq ($A,&QWP(0,"esi"));
292 &movq ("mm1",&QWP(8,"esi"));
293 &movq ("mm2",&QWP(16,"esi"));
294 &movq ("mm3",&QWP(24,"esi"));
295 &movq ($E,&QWP(32,"esi"));
296 &movq ("mm5",&QWP(40,"esi"));
297 &movq ("mm6",&QWP(48,"esi"));
298 &movq ("mm7",&QWP(56,"esi"));
301 &set_label("loop_sse2",16);
303 &movq ($Bsse2,"mm1");
304 &movq ($Csse2,"mm2");
305 &movq ($Dsse2,"mm3");
307 &movq ($Fsse2,"mm5");
308 &movq ($Gsse2,"mm6");
309 &movq ($Hsse2,"mm7");
311 &mov ("ecx",&DWP(0,"edi"));
312 &mov ("edx",&DWP(4,"edi"));
316 &mov (&DWP(8*9+4,"esp"),"ecx");
317 &mov (&DWP(8*9+0,"esp"),"edx");
319 &set_label("00_14_sse2",16);
320 &mov ("eax",&DWP(0,"edi"));
321 &mov ("ebx",&DWP(4,"edi"));
325 &mov (&DWP(8*8+4,"esp"),"eax");
326 &mov (&DWP(8*8+0,"esp"),"ebx");
330 &cmp (&LB("edx"),0x35);
331 &jne (&label("00_14_sse2"));
335 &set_label("16_79_sse2",16);
336 #&movq ("mm2",&QWP(8*(9+16-1),"esp")); #prefetched in BODY_00_15
337 #&movq ("mm6",&QWP(8*(9+16-14),"esp"));
352 &psrlq ("mm6",61-19);
355 &movq ("mm2",&QWP(8*(9+16),"esp"));
362 &paddq ("mm2",&QWP(8*(9+16-9),"esp"));
364 &psllq ("mm1",63-56);
370 &paddq ("mm3","mm5");
371 &paddq ("mm3","mm2");
372 &movq (&QWP(8*9,"esp"),"mm3");
376 &cmp (&LB("edx"),0x17);
377 &jne (&label("16_79_sse2"));
380 &movq ("mm1",$Bsse2);
381 &movq ("mm2",$Csse2);
382 &movq ("mm3",$Dsse2);
384 &movq ("mm5",$Fsse2);
385 &movq ("mm6",$Gsse2);
386 &movq ("mm7",$Hsse2);
388 &paddq ($A,&QWP(0,"esi"));
389 &paddq ("mm1",&QWP(8,"esi"));
390 &paddq ("mm2",&QWP(16,"esi"));
391 &paddq ("mm3",&QWP(24,"esi"));
392 &paddq ($E,&QWP(32,"esi"));
393 &paddq ("mm5",&QWP(40,"esi"));
394 &paddq ("mm6",&QWP(48,"esi"));
395 &paddq ("mm7",&QWP(56,"esi"));
397 &movq (&QWP(0,"esi"),$A);
398 &movq (&QWP(8,"esi"),"mm1");
399 &movq (&QWP(16,"esi"),"mm2");
400 &movq (&QWP(24,"esi"),"mm3");
401 &movq (&QWP(32,"esi"),$E);
402 &movq (&QWP(40,"esi"),"mm5");
403 &movq (&QWP(48,"esi"),"mm6");
404 &movq (&QWP(56,"esi"),"mm7");
406 &add ("esp",8*80); # destroy frame
407 &sub ($K512,8*80); # rewind K
409 &cmp ("edi",&DWP(8*10+8,"esp")); # are we done yet?
410 &jb (&label("loop_sse2"));
413 &mov ("esp",&DWP(8*10+12,"esp")); # restore sp
416 &set_label("loop_x86",16);
417 # copy input block to stack reversing byte and qword order
418 for ($i=0;$i<8;$i++) {
419 &mov ("eax",&DWP($i*16+0,"edi"));
420 &mov ("ebx",&DWP($i*16+4,"edi"));
421 &mov ("ecx",&DWP($i*16+8,"edi"));
422 &mov ("edx",&DWP($i*16+12,"edi"));
433 &sub ("esp",9*8); # place for T,A,B,C,D,E,F,G,H
434 &mov (&DWP(8*(9+16)+4,"esp"),"edi");
436 # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
437 &lea ("edi",&DWP(8,"esp"));
439 &data_word(0xA5F3F689); # rep movsd
441 &set_label("00_15_x86",16);
444 &cmp (&LB("edx"),0x94);
445 &jne (&label("00_15_x86"));
447 &set_label("16_79_x86",16);
448 #define sigma0(x) (ROTR((x),1) ^ ROTR((x),8) ^ ((x)>>7))
449 # LO lo>>1^hi<<31 ^ lo>>8^hi<<24 ^ lo>>7^hi<<25
450 # HI hi>>1^lo<<31 ^ hi>>8^lo<<24 ^ hi>>7
451 &mov ("ecx",&DWP(8*(9+15+16-1)+0,"esp"));
452 &mov ("edx",&DWP(8*(9+15+16-1)+4,"esp"));
455 &shr ("ecx",1) # lo>>1
457 &shr ("edx",1) # hi>>1
459 &shl ("esi",24); # lo<<24
461 &shl ("edi",24); # hi<<24
464 &shr ("ecx",7-1); # lo>>7
466 &shr ("edx",7-1); # hi>>7
468 &shl ("esi",31-24); # lo<<31
470 &shl ("edi",25-24); # hi<<25
473 &shr ("ecx",8-7); # lo>>8
475 &shr ("edx",8-7); # hi>>8
477 &shl ("edi",31-25); # hi<<31
479 &xor ("eax","edi"); # T1 = sigma0(X[-15])
481 &mov (&DWP(0,"esp"),"eax");
482 &mov (&DWP(4,"esp"),"ebx"); # put T1 away
484 #define sigma1(x) (ROTR((x),19) ^ ROTR((x),61) ^ ((x)>>6))
485 # LO lo>>19^hi<<13 ^ hi>>29^lo<<3 ^ lo>>6^hi<<26
486 # HI hi>>19^lo<<13 ^ lo>>29^hi<<3 ^ hi>>6
487 &mov ("ecx",&DWP(8*(9+15+16-14)+0,"esp"));
488 &mov ("edx",&DWP(8*(9+15+16-14)+4,"esp"));
491 &shr ("ecx",6) # lo>>6
493 &shr ("edx",6) # hi>>6
495 &shl ("esi",3); # lo<<3
497 &shl ("edi",3); # hi<<3
500 &shr ("ecx",19-6); # lo>>19
502 &shr ("edx",19-6); # hi>>19
504 &shl ("esi",13-3); # lo<<13
506 &shl ("edi",13-3); # hi<<13
509 &shr ("ecx",29-19); # lo>>29
511 &shr ("edx",29-19); # hi>>29
513 &shl ("edi",26-13); # hi<<26
515 &xor ("eax","edi"); # sigma1(X[-2])
517 &mov ("ecx",&DWP(8*(9+15+16)+0,"esp"));
518 &mov ("edx",&DWP(8*(9+15+16)+4,"esp"));
519 &add ("eax",&DWP(0,"esp"));
520 &adc ("ebx",&DWP(4,"esp")); # T1 = sigma1(X[-2])+T1
521 &mov ("esi",&DWP(8*(9+15+16-9)+0,"esp"));
522 &mov ("edi",&DWP(8*(9+15+16-9)+4,"esp"));
524 &adc ("ebx","edx"); # T1 += X[-16]
526 &adc ("ebx","edi"); # T1 += X[-7]
527 &mov (&DWP(8*(9+15)+0,"esp"),"eax");
528 &mov (&DWP(8*(9+15)+4,"esp"),"ebx"); # save X[0]
532 &cmp (&LB("edx"),0x17);
533 &jne (&label("16_79_x86"));
535 &mov ("esi",&DWP(8*(9+16+80)+0,"esp"));# ctx
536 &mov ("edi",&DWP(8*(9+16+80)+4,"esp"));# inp
537 for($i=0;$i<4;$i++) {
538 &mov ("eax",&DWP($i*16+0,"esi"));
539 &mov ("ebx",&DWP($i*16+4,"esi"));
540 &mov ("ecx",&DWP($i*16+8,"esi"));
541 &mov ("edx",&DWP($i*16+12,"esi"));
542 &add ("eax",&DWP(8+($i*16)+0,"esp"));
543 &adc ("ebx",&DWP(8+($i*16)+4,"esp"));
544 &mov (&DWP($i*16+0,"esi"),"eax");
545 &mov (&DWP($i*16+4,"esi"),"ebx");
546 &add ("ecx",&DWP(8+($i*16)+8,"esp"));
547 &adc ("edx",&DWP(8+($i*16)+12,"esp"));
548 &mov (&DWP($i*16+8,"esi"),"ecx");
549 &mov (&DWP($i*16+12,"esi"),"edx");
551 &add ("esp",8*(9+16+80)); # destroy frame
552 &sub ($K512,8*80); # rewind K
554 &cmp ("edi",&DWP(8,"esp")); # are we done yet?
555 &jb (&label("loop_x86"));
557 &mov ("esp",&DWP(12,"esp")); # restore sp
560 &set_label("K512",64); # Yes! I keep it in the code segment!
561 &data_word(0xd728ae22,0x428a2f98); # u64
562 &data_word(0x23ef65cd,0x71374491); # u64
563 &data_word(0xec4d3b2f,0xb5c0fbcf); # u64
564 &data_word(0x8189dbbc,0xe9b5dba5); # u64
565 &data_word(0xf348b538,0x3956c25b); # u64
566 &data_word(0xb605d019,0x59f111f1); # u64
567 &data_word(0xaf194f9b,0x923f82a4); # u64
568 &data_word(0xda6d8118,0xab1c5ed5); # u64
569 &data_word(0xa3030242,0xd807aa98); # u64
570 &data_word(0x45706fbe,0x12835b01); # u64
571 &data_word(0x4ee4b28c,0x243185be); # u64
572 &data_word(0xd5ffb4e2,0x550c7dc3); # u64
573 &data_word(0xf27b896f,0x72be5d74); # u64
574 &data_word(0x3b1696b1,0x80deb1fe); # u64
575 &data_word(0x25c71235,0x9bdc06a7); # u64
576 &data_word(0xcf692694,0xc19bf174); # u64
577 &data_word(0x9ef14ad2,0xe49b69c1); # u64
578 &data_word(0x384f25e3,0xefbe4786); # u64
579 &data_word(0x8b8cd5b5,0x0fc19dc6); # u64
580 &data_word(0x77ac9c65,0x240ca1cc); # u64
581 &data_word(0x592b0275,0x2de92c6f); # u64
582 &data_word(0x6ea6e483,0x4a7484aa); # u64
583 &data_word(0xbd41fbd4,0x5cb0a9dc); # u64
584 &data_word(0x831153b5,0x76f988da); # u64
585 &data_word(0xee66dfab,0x983e5152); # u64
586 &data_word(0x2db43210,0xa831c66d); # u64
587 &data_word(0x98fb213f,0xb00327c8); # u64
588 &data_word(0xbeef0ee4,0xbf597fc7); # u64
589 &data_word(0x3da88fc2,0xc6e00bf3); # u64
590 &data_word(0x930aa725,0xd5a79147); # u64
591 &data_word(0xe003826f,0x06ca6351); # u64
592 &data_word(0x0a0e6e70,0x14292967); # u64
593 &data_word(0x46d22ffc,0x27b70a85); # u64
594 &data_word(0x5c26c926,0x2e1b2138); # u64
595 &data_word(0x5ac42aed,0x4d2c6dfc); # u64
596 &data_word(0x9d95b3df,0x53380d13); # u64
597 &data_word(0x8baf63de,0x650a7354); # u64
598 &data_word(0x3c77b2a8,0x766a0abb); # u64
599 &data_word(0x47edaee6,0x81c2c92e); # u64
600 &data_word(0x1482353b,0x92722c85); # u64
601 &data_word(0x4cf10364,0xa2bfe8a1); # u64
602 &data_word(0xbc423001,0xa81a664b); # u64
603 &data_word(0xd0f89791,0xc24b8b70); # u64
604 &data_word(0x0654be30,0xc76c51a3); # u64
605 &data_word(0xd6ef5218,0xd192e819); # u64
606 &data_word(0x5565a910,0xd6990624); # u64
607 &data_word(0x5771202a,0xf40e3585); # u64
608 &data_word(0x32bbd1b8,0x106aa070); # u64
609 &data_word(0xb8d2d0c8,0x19a4c116); # u64
610 &data_word(0x5141ab53,0x1e376c08); # u64
611 &data_word(0xdf8eeb99,0x2748774c); # u64
612 &data_word(0xe19b48a8,0x34b0bcb5); # u64
613 &data_word(0xc5c95a63,0x391c0cb3); # u64
614 &data_word(0xe3418acb,0x4ed8aa4a); # u64
615 &data_word(0x7763e373,0x5b9cca4f); # u64
616 &data_word(0xd6b2b8a3,0x682e6ff3); # u64
617 &data_word(0x5defb2fc,0x748f82ee); # u64
618 &data_word(0x43172f60,0x78a5636f); # u64
619 &data_word(0xa1f0ab72,0x84c87814); # u64
620 &data_word(0x1a6439ec,0x8cc70208); # u64
621 &data_word(0x23631e28,0x90befffa); # u64
622 &data_word(0xde82bde9,0xa4506ceb); # u64
623 &data_word(0xb2c67915,0xbef9a3f7); # u64
624 &data_word(0xe372532b,0xc67178f2); # u64
625 &data_word(0xea26619c,0xca273ece); # u64
626 &data_word(0x21c0c207,0xd186b8c7); # u64
627 &data_word(0xcde0eb1e,0xeada7dd6); # u64
628 &data_word(0xee6ed178,0xf57d4f7f); # u64
629 &data_word(0x72176fba,0x06f067aa); # u64
630 &data_word(0xa2c898a6,0x0a637dc5); # u64
631 &data_word(0xbef90dae,0x113f9804); # u64
632 &data_word(0x131c471b,0x1b710b35); # u64
633 &data_word(0x23047d84,0x28db77f5); # u64
634 &data_word(0x40c72493,0x32caab7b); # u64
635 &data_word(0x15c9bebc,0x3c9ebe0a); # u64
636 &data_word(0x9c100d4c,0x431d67c4); # u64
637 &data_word(0xcb3e42b6,0x4cc5d4be); # u64
638 &data_word(0xfc657e2a,0x597f299c); # u64
639 &data_word(0x3ad6faec,0x5fcb6fab); # u64
640 &data_word(0x4a475817,0x6c44198c); # u64
641 &function_end_B("sha512_block_data_order");
642 &asciz("SHA512 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");