2 # Copyright 1998-2016 The OpenSSL Project Authors. All Rights Reserved.
4 # Licensed under the OpenSSL license (the "License"). You may not use
5 # this file except in compliance with the License. You can obtain a copy
6 # in the file LICENSE in the source distribution or at
7 # https://www.openssl.org/source/license.html
10 # ====================================================================
11 # [Re]written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12 # project. The module is, however, dual licensed under OpenSSL and
13 # CRYPTOGAMS licenses depending on where you obtain it. For further
14 # details see http://www.openssl.org/~appro/cryptogams/.
15 # ====================================================================
17 # "[Re]written" was achieved in two major overhauls. In 2004 BODY_*
18 # functions were re-implemented to address P4 performance issue [see
19 # commentary below], and in 2006 the rest was rewritten in order to
20 # gain freedom to liberate licensing terms.
22 # January, September 2004.
24 # It was noted that Intel IA-32 C compiler generates code which
25 # performs ~30% *faster* on P4 CPU than original *hand-coded*
26 # SHA1 assembler implementation. To address this problem (and
27 # prove that humans are still better than machines:-), the
28 # original code was overhauled, which resulted in following
29 # performance changes:
31 # compared with original compared with Intel cc
32 # assembler impl. generated code
37 # As you can see Pentium came out as looser:-( Yet I reckoned that
38 # improvement on P4 outweights the loss and incorporate this
39 # re-tuned code to 0.9.7 and later.
40 # ----------------------------------------------------------------
41 # <appro@fy.chalmers.se>
45 # George Spelvin has tipped that F_40_59(b,c,d) can be rewritten as
46 # '(c&d) + (b&(c^d))', which allows to accumulate partial results
47 # and lighten "pressure" on scratch registers. This resulted in
48 # >12% performance improvement on contemporary AMD cores (with no
49 # degradation on other CPUs:-). Also, the code was revised to maximize
50 # "distance" between instructions producing input to 'lea' instruction
51 # and the 'lea' instruction itself, which is essential for Intel Atom
52 # core and resulted in ~15% improvement.
56 # Add SSSE3, Supplemental[!] SSE3, implementation. The idea behind it
57 # is to offload message schedule denoted by Wt in NIST specification,
58 # or Xupdate in OpenSSL source, to SIMD unit. The idea is not novel,
59 # and in SSE2 context was first explored by Dean Gaudet in 2004, see
60 # http://arctic.org/~dean/crypto/sha1.html. Since then several things
61 # have changed that made it interesting again:
63 # a) XMM units became faster and wider;
64 # b) instruction set became more versatile;
65 # c) an important observation was made by Max Locktykhin, which made
66 # it possible to reduce amount of instructions required to perform
67 # the operation in question, for further details see
68 # http://software.intel.com/en-us/articles/improving-the-performance-of-the-secure-hash-algorithm-1/.
72 # Add AVX code path, probably most controversial... The thing is that
73 # switch to AVX alone improves performance by as little as 4% in
74 # comparison to SSSE3 code path. But below result doesn't look like
75 # 4% improvement... Trouble is that Sandy Bridge decodes 'ro[rl]' as
76 # pair of µ-ops, and it's the additional µ-ops, two per round, that
77 # make it run slower than Core2 and Westmere. But 'sh[rl]d' is decoded
78 # as single µ-op by Sandy Bridge and it's replacing 'ro[rl]' with
79 # equivalent 'sh[rl]d' that is responsible for the impressive 5.1
80 # cycles per processed byte. But 'sh[rl]d' is not something that used
81 # to be fast, nor does it appear to be fast in upcoming Bulldozer
82 # [according to its optimization manual]. Which is why AVX code path
83 # is guarded by *both* AVX and synthetic bit denoting Intel CPUs.
84 # One can argue that it's unfair to AMD, but without 'sh[rl]d' it
85 # makes no sense to keep the AVX code path. If somebody feels that
86 # strongly, it's probably more appropriate to discuss possibility of
87 # using vector rotate XOP on AMD...
91 # Add support for Intel SHA Extensions.
93 ######################################################################
94 # Current performance is summarized in following table. Numbers are
95 # CPU clock cycles spent to process single byte (less is better).
102 # Core2 7.3 6.0/+22% -
103 # Westmere 7.3 5.5/+33% -
104 # Sandy Bridge 8.8 6.2/+40% 5.1(**)/+73%
105 # Ivy Bridge 7.2 4.8/+51% 4.7(**)/+53%
106 # Haswell 6.5 4.3/+51% 4.1(**)/+58%
107 # Bulldozer 11.6 6.0/+92%
108 # VIA Nano 10.6 7.5/+41%
109 # Atom 12.5 9.3(*)/+35%
110 # Silvermont 14.5 9.9(*)/+46%
112 # (*) Loop is 1056 instructions long and expected result is ~8.25.
113 # The discrepancy is because of front-end limitations, so
114 # called MS-ROM penalties, and on Silvermont even rotate's
115 # limited parallelism.
117 # (**) As per above comment, the result is for AVX *plus* sh[rl]d.
119 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
120 push(@INC,"${dir}","${dir}../../perlasm");
124 open STDOUT,">$output";
126 &asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386");
129 for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
132 `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
133 =~ /GNU assembler version ([2-9]\.[0-9]+)/ &&
134 $1>=2.19); # first version supporting AVX
136 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32n" &&
137 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ &&
138 $1>=2.03); # first version supporting AVX
140 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32" &&
141 `ml 2>&1` =~ /Version ([0-9]+)\./ &&
142 $1>=10); # first version supporting AVX
144 $ymm=1 if ($xmm && !$ymm && `$ENV{CC} -v 2>&1` =~ /(^clang version|based on LLVM) ([3-9]\.[0-9]+)/ &&
145 $2>=3.0); # first version supporting AVX
147 $shaext=$xmm; ### set to zero if compiling for 1.0.1
149 &external_label("OPENSSL_ia32cap_P") if ($xmm);
160 @V=($A,$B,$C,$D,$E,$T);
162 $alt=0; # 1 denotes alternative IALU implementation, which performs
163 # 8% *worse* on P4, same on Westmere and Atom, 2% better on
168 local($n,$a,$b,$c,$d,$e,$f)=@_;
170 &comment("00_15 $n");
172 &mov($f,$c); # f to hold F_00_19(b,c,d)
173 if ($n==0) { &mov($tmp1,$a); }
174 else { &mov($a,$tmp1); }
175 &rotl($tmp1,5); # tmp1=ROTATE(a,5)
177 &add($tmp1,$e); # tmp1+=e;
178 &mov($e,&swtmp($n%16)); # e becomes volatile and is loaded
179 # with xi, also note that e becomes
182 &rotr($b,2); # b=ROTATE(b,30)
183 &xor($f,$d); # f holds F_00_19(b,c,d)
184 &lea($tmp1,&DWP(0x5a827999,$tmp1,$e)); # tmp1+=K_00_19+xi
186 if ($n==15) { &mov($e,&swtmp(($n+1)%16));# pre-fetch f for next round
187 &add($f,$tmp1); } # f+=tmp1
188 else { &add($tmp1,$f); } # f becomes a in next round
189 &mov($tmp1,$a) if ($alt && $n==15);
194 local($n,$a,$b,$c,$d,$e,$f)=@_;
196 &comment("16_19 $n");
200 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
201 &and($tmp1,$c); # tmp1 to hold F_00_19(b,c,d), b&=c^d
202 &xor($f,&swtmp(($n+8)%16));
203 &xor($tmp1,$d); # tmp1=F_00_19(b,c,d)
204 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
205 &rotl($f,1); # f=ROTATE(f,1)
206 &add($e,$tmp1); # e+=F_00_19(b,c,d)
207 &xor($c,$d); # restore $c
208 &mov($tmp1,$a); # b in next round
209 &rotr($b,$n==16?2:7); # b=ROTATE(b,30)
210 &mov(&swtmp($n%16),$f); # xi=f
211 &rotl($a,5); # ROTATE(a,5)
212 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
213 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
214 &add($f,$a); # f+=ROTATE(a,5)
216 &mov($tmp1,$c); # tmp1 to hold F_00_19(b,c,d)
217 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
219 &xor($f,&swtmp(($n+8)%16));
221 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
222 &rotl($f,1); # f=ROTATE(f,1)
223 &xor($tmp1,$d); # tmp1=F_00_19(b,c,d)
224 &add($e,$tmp1); # e+=F_00_19(b,c,d)
226 &rotr($b,2); # b=ROTATE(b,30)
227 &mov(&swtmp($n%16),$f); # xi=f
228 &rotl($tmp1,5); # ROTATE(a,5)
229 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
230 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
231 &add($f,$tmp1); # f+=ROTATE(a,5)
237 local($n,$a,$b,$c,$d,$e,$f)=@_;
238 local $K=($n<40)?0x6ed9eba1:0xca62c1d6;
240 &comment("20_39 $n");
243 &xor($tmp1,$c); # tmp1 to hold F_20_39(b,c,d), b^=c
244 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
245 &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d)
246 &xor($f,&swtmp(($n+8)%16));
247 &add($e,$tmp1); # e+=F_20_39(b,c,d)
248 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
249 &rotl($f,1); # f=ROTATE(f,1)
250 &mov($tmp1,$a); # b in next round
251 &rotr($b,7); # b=ROTATE(b,30)
252 &mov(&swtmp($n%16),$f) if($n<77);# xi=f
253 &rotl($a,5); # ROTATE(a,5)
254 &xor($b,$c) if($n==39);# warm up for BODY_40_59
255 &and($tmp1,$b) if($n==39);
256 &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY
257 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
258 &add($f,$a); # f+=ROTATE(a,5)
259 &rotr($a,5) if ($n==79);
261 &mov($tmp1,$b); # tmp1 to hold F_20_39(b,c,d)
262 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
264 &xor($f,&swtmp(($n+8)%16));
265 &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d)
266 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
267 &rotl($f,1); # f=ROTATE(f,1)
268 &add($e,$tmp1); # e+=F_20_39(b,c,d)
269 &rotr($b,2); # b=ROTATE(b,30)
271 &rotl($tmp1,5); # ROTATE(a,5)
272 &mov(&swtmp($n%16),$f) if($n<77);# xi=f
273 &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY
274 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
275 &add($f,$tmp1); # f+=ROTATE(a,5)
281 local($n,$a,$b,$c,$d,$e,$f)=@_;
283 &comment("40_59 $n");
286 &add($e,$tmp1); # e+=b&(c^d)
287 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
289 &xor($f,&swtmp(($n+8)%16));
290 &xor($c,$d); # restore $c
291 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
292 &rotl($f,1); # f=ROTATE(f,1)
294 &rotr($b,7); # b=ROTATE(b,30)
295 &add($e,$tmp1); # e+=c&d
296 &mov($tmp1,$a); # b in next round
297 &mov(&swtmp($n%16),$f); # xi=f
298 &rotl($a,5); # ROTATE(a,5)
299 &xor($b,$c) if ($n<59);
300 &and($tmp1,$b) if ($n<59);# tmp1 to hold F_40_59(b,c,d)
301 &lea($f,&DWP(0x8f1bbcdc,$f,$e));# f+=K_40_59+e+(b&(c^d))
302 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
303 &add($f,$a); # f+=ROTATE(a,5)
305 &mov($tmp1,$c); # tmp1 to hold F_40_59(b,c,d)
306 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
308 &xor($f,&swtmp(($n+8)%16));
310 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
311 &rotl($f,1); # f=ROTATE(f,1)
312 &add($tmp1,$e); # b&(c^d)+=e
313 &rotr($b,2); # b=ROTATE(b,30)
314 &mov($e,$a); # e becomes volatile
315 &rotl($e,5); # ROTATE(a,5)
316 &mov(&swtmp($n%16),$f); # xi=f
317 &lea($f,&DWP(0x8f1bbcdc,$f,$tmp1));# f+=K_40_59+e+(b&(c^d))
319 &add($f,$e); # f+=ROTATE(a,5)
321 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
322 &add($f,$tmp1); # f+=c&d
326 &function_begin("sha1_block_data_order");
328 &static_label("shaext_shortcut") if ($shaext);
329 &static_label("ssse3_shortcut");
330 &static_label("avx_shortcut") if ($ymm);
331 &static_label("K_XX_XX");
333 &call (&label("pic_point")); # make it PIC!
334 &set_label("pic_point");
336 &picmeup($T,"OPENSSL_ia32cap_P",$tmp1,&label("pic_point"));
337 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
339 &mov ($A,&DWP(0,$T));
340 &mov ($D,&DWP(4,$T));
341 &test ($D,1<<9); # check SSSE3 bit
343 &mov ($C,&DWP(8,$T));
344 &test ($A,1<<24); # check FXSR bit
347 &test ($C,1<<29); # check SHA bit
348 &jnz (&label("shaext_shortcut"));
351 &and ($D,1<<28); # mask AVX bit
352 &and ($A,1<<30); # mask "Intel CPU" bit
354 &cmp ($A,1<<28|1<<30);
355 &je (&label("avx_shortcut"));
357 &jmp (&label("ssse3_shortcut"));
358 &set_label("x86",16);
360 &mov($tmp1,&wparam(0)); # SHA_CTX *c
361 &mov($T,&wparam(1)); # const void *input
362 &mov($A,&wparam(2)); # size_t num
363 &stack_push(16+3); # allocate X[16]
366 &mov(&wparam(2),$A); # pointer beyond the end of input
367 &mov($E,&DWP(16,$tmp1));# pre-load E
368 &jmp(&label("loop"));
370 &set_label("loop",16);
372 # copy input chunk to X, but reversing byte order!
373 for ($i=0; $i<16; $i+=4)
375 &mov($A,&DWP(4*($i+0),$T));
376 &mov($B,&DWP(4*($i+1),$T));
377 &mov($C,&DWP(4*($i+2),$T));
378 &mov($D,&DWP(4*($i+3),$T));
383 &mov(&swtmp($i+0),$A);
384 &mov(&swtmp($i+1),$B);
385 &mov(&swtmp($i+2),$C);
386 &mov(&swtmp($i+3),$D);
388 &mov(&wparam(1),$T); # redundant in 1st spin
390 &mov($A,&DWP(0,$tmp1)); # load SHA_CTX
391 &mov($B,&DWP(4,$tmp1));
392 &mov($C,&DWP(8,$tmp1));
393 &mov($D,&DWP(12,$tmp1));
396 for($i=0;$i<16;$i++) { &BODY_00_15($i,@V); unshift(@V,pop(@V)); }
397 for(;$i<20;$i++) { &BODY_16_19($i,@V); unshift(@V,pop(@V)); }
398 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
399 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
400 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
402 (($V[5] eq $D) and ($V[0] eq $E)) or die; # double-check
404 &mov($tmp1,&wparam(0)); # re-load SHA_CTX*
405 &mov($D,&wparam(1)); # D is last "T" and is discarded
407 &add($E,&DWP(0,$tmp1)); # E is last "A"...
408 &add($T,&DWP(4,$tmp1));
409 &add($A,&DWP(8,$tmp1));
410 &add($B,&DWP(12,$tmp1));
411 &add($C,&DWP(16,$tmp1));
413 &mov(&DWP(0,$tmp1),$E); # update SHA_CTX
414 &add($D,64); # advance input pointer
415 &mov(&DWP(4,$tmp1),$T);
416 &cmp($D,&wparam(2)); # have we reached the end yet?
417 &mov(&DWP(8,$tmp1),$A);
418 &mov($E,$C); # C is last "E" which needs to be "pre-loaded"
419 &mov(&DWP(12,$tmp1),$B);
420 &mov($T,$D); # input pointer
421 &mov(&DWP(16,$tmp1),$C);
425 &function_end("sha1_block_data_order");
429 ######################################################################
430 # Intel SHA Extensions implementation of SHA1 update function.
432 my ($ctx,$inp,$num)=("edi","esi","ecx");
433 my ($ABCD,$E,$E_,$BSWAP)=map("xmm$_",(0..3));
434 my @MSG=map("xmm$_",(4..7));
437 my ($dst,$src,$imm)=@_;
438 if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
439 { &data_byte(0x0f,0x3a,0xcc,0xc0|($1<<3)|$2,$imm); }
442 my ($opcodelet,$dst,$src)=@_;
443 if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
444 { &data_byte(0x0f,0x38,$opcodelet,0xc0|($1<<3)|$2); }
446 sub sha1nexte { sha1op38(0xc8,@_); }
447 sub sha1msg1 { sha1op38(0xc9,@_); }
448 sub sha1msg2 { sha1op38(0xca,@_); }
450 &function_begin("_sha1_block_data_order_shaext");
451 &call (&label("pic_point")); # make it PIC!
452 &set_label("pic_point");
454 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
455 &set_label("shaext_shortcut");
456 &mov ($ctx,&wparam(0));
458 &mov ($inp,&wparam(1));
459 &mov ($num,&wparam(2));
462 &movdqu ($ABCD,&QWP(0,$ctx));
463 &movd ($E,&DWP(16,$ctx));
465 &movdqa ($BSWAP,&QWP(0x50,$tmp1)); # byte-n-word swap
467 &movdqu (@MSG[0],&QWP(0,$inp));
468 &pshufd ($ABCD,$ABCD,0b00011011); # flip word order
469 &movdqu (@MSG[1],&QWP(0x10,$inp));
470 &pshufd ($E,$E,0b00011011); # flip word order
471 &movdqu (@MSG[2],&QWP(0x20,$inp));
472 &pshufb (@MSG[0],$BSWAP);
473 &movdqu (@MSG[3],&QWP(0x30,$inp));
474 &pshufb (@MSG[1],$BSWAP);
475 &pshufb (@MSG[2],$BSWAP);
476 &pshufb (@MSG[3],$BSWAP);
477 &jmp (&label("loop_shaext"));
479 &set_label("loop_shaext",16);
481 &lea ("eax",&DWP(0x40,$inp));
482 &movdqa (&QWP(0,"esp"),$E); # offload $E
484 &cmovne ($inp,"eax");
485 &movdqa (&QWP(16,"esp"),$ABCD); # offload $ABCD
487 for($i=0;$i<20-4;$i+=2) {
488 &sha1msg1 (@MSG[0],@MSG[1]);
490 &sha1rnds4 ($ABCD,$E,int($i/5)); # 0-3...
491 &sha1nexte ($E_,@MSG[1]);
492 &pxor (@MSG[0],@MSG[2]);
493 &sha1msg1 (@MSG[1],@MSG[2]);
494 &sha1msg2 (@MSG[0],@MSG[3]);
497 &sha1rnds4 ($ABCD,$E_,int(($i+1)/5));
498 &sha1nexte ($E,@MSG[2]);
499 &pxor (@MSG[1],@MSG[3]);
500 &sha1msg2 (@MSG[1],@MSG[0]);
502 push(@MSG,shift(@MSG)); push(@MSG,shift(@MSG));
504 &movdqu (@MSG[0],&QWP(0,$inp));
506 &sha1rnds4 ($ABCD,$E,3); # 64-67
507 &sha1nexte ($E_,@MSG[1]);
508 &movdqu (@MSG[1],&QWP(0x10,$inp));
509 &pshufb (@MSG[0],$BSWAP);
512 &sha1rnds4 ($ABCD,$E_,3); # 68-71
513 &sha1nexte ($E,@MSG[2]);
514 &movdqu (@MSG[2],&QWP(0x20,$inp));
515 &pshufb (@MSG[1],$BSWAP);
518 &sha1rnds4 ($ABCD,$E,3); # 72-75
519 &sha1nexte ($E_,@MSG[3]);
520 &movdqu (@MSG[3],&QWP(0x30,$inp));
521 &pshufb (@MSG[2],$BSWAP);
524 &sha1rnds4 ($ABCD,$E_,3); # 76-79
525 &movdqa ($E_,&QWP(0,"esp"));
526 &pshufb (@MSG[3],$BSWAP);
528 &paddd ($ABCD,&QWP(16,"esp"));
530 &jnz (&label("loop_shaext"));
532 &pshufd ($ABCD,$ABCD,0b00011011);
533 &pshufd ($E,$E,0b00011011);
534 &movdqu (&QWP(0,$ctx),$ABCD)
535 &movd (&DWP(16,$ctx),$E);
537 &function_end("_sha1_block_data_order_shaext");
539 ######################################################################
540 # The SSSE3 implementation.
542 # %xmm[0-7] are used as ring @X[] buffer containing quadruples of last
543 # 32 elements of the message schedule or Xupdate outputs. First 4
544 # quadruples are simply byte-swapped input, next 4 are calculated
545 # according to method originally suggested by Dean Gaudet (modulo
546 # being implemented in SSSE3). Once 8 quadruples or 32 elements are
547 # collected, it switches to routine proposed by Max Locktyukhin.
549 # Calculations inevitably require temporary reqisters, and there are
550 # no %xmm registers left to spare. For this reason part of the ring
551 # buffer, X[2..4] to be specific, is offloaded to 3 quadriples ring
552 # buffer on the stack. Keep in mind that X[2] is alias X[-6], X[3] -
553 # X[-5], and X[4] - X[-4]...
555 # Another notable optimization is aggressive stack frame compression
556 # aiming to minimize amount of 9-byte instructions...
558 # Yet another notable optimization is "jumping" $B variable. It means
559 # that there is no register permanently allocated for $B value. This
560 # allowed to eliminate one instruction from body_20_39...
562 my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded
563 my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4
564 my @V=($A,$B,$C,$D,$E);
565 my $j=0; # hash round
570 my $_rol=sub { &rol(@_) };
571 my $_ror=sub { &ror(@_) };
573 &function_begin("_sha1_block_data_order_ssse3");
574 &call (&label("pic_point")); # make it PIC!
575 &set_label("pic_point");
577 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
578 &set_label("ssse3_shortcut");
580 &movdqa (@X[3],&QWP(0,$tmp1)); # K_00_19
581 &movdqa (@X[4],&QWP(16,$tmp1)); # K_20_39
582 &movdqa (@X[5],&QWP(32,$tmp1)); # K_40_59
583 &movdqa (@X[6],&QWP(48,$tmp1)); # K_60_79
584 &movdqa (@X[2],&QWP(64,$tmp1)); # pbswap mask
586 &mov ($E,&wparam(0)); # load argument block
587 &mov ($inp=@T[1],&wparam(1));
588 &mov ($D,&wparam(2));
593 # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area
594 # X[4]+K X[5]+K X[6]+K X[7]+K
595 # X[8]+K X[9]+K X[10]+K X[11]+K
596 # X[12]+K X[13]+K X[14]+K X[15]+K
598 # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area
599 # X[4] X[5] X[6] X[7]
600 # X[8] X[9] X[10] X[11] # even borrowed for K_00_19
602 # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants
603 # K_40_59 K_40_59 K_40_59 K_40_59
604 # K_60_79 K_60_79 K_60_79 K_60_79
605 # K_00_19 K_00_19 K_00_19 K_00_19
608 # +192 ctx # argument block
615 &movdqa (&QWP(112+0,"esp"),@X[4]); # copy constants
616 &movdqa (&QWP(112+16,"esp"),@X[5]);
617 &movdqa (&QWP(112+32,"esp"),@X[6]);
618 &shl ($D,6); # len*64
619 &movdqa (&QWP(112+48,"esp"),@X[3]);
620 &add ($D,$inp); # end of input
621 &movdqa (&QWP(112+64,"esp"),@X[2]);
623 &mov (&DWP(192+0,"esp"),$E); # save argument block
624 &mov (&DWP(192+4,"esp"),$inp);
625 &mov (&DWP(192+8,"esp"),$D);
626 &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp
628 &mov ($A,&DWP(0,$E)); # load context
629 &mov ($B,&DWP(4,$E));
630 &mov ($C,&DWP(8,$E));
631 &mov ($D,&DWP(12,$E));
632 &mov ($E,&DWP(16,$E));
633 &mov (@T[0],$B); # magic seed
635 &movdqu (@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3]
636 &movdqu (@X[-3&7],&QWP(-48,$inp));
637 &movdqu (@X[-2&7],&QWP(-32,$inp));
638 &movdqu (@X[-1&7],&QWP(-16,$inp));
639 &pshufb (@X[-4&7],@X[2]); # byte swap
640 &pshufb (@X[-3&7],@X[2]);
641 &pshufb (@X[-2&7],@X[2]);
642 &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
643 &pshufb (@X[-1&7],@X[2]);
644 &paddd (@X[-4&7],@X[3]); # add K_00_19
645 &paddd (@X[-3&7],@X[3]);
646 &paddd (@X[-2&7],@X[3]);
647 &movdqa (&QWP(0,"esp"),@X[-4&7]); # X[]+K xfer to IALU
648 &psubd (@X[-4&7],@X[3]); # restore X[]
649 &movdqa (&QWP(0+16,"esp"),@X[-3&7]);
650 &psubd (@X[-3&7],@X[3]);
651 &movdqa (&QWP(0+32,"esp"),@X[-2&7]);
653 &psubd (@X[-2&7],@X[3]);
655 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
657 &jmp (&label("loop"));
659 ######################################################################
660 # SSE instruction sequence is first broken to groups of indepentent
661 # instructions, independent in respect to their inputs and shifter
662 # (not all architectures have more than one). Then IALU instructions
663 # are "knitted in" between the SSE groups. Distance is maintained for
664 # SSE latency of 2 in hope that it fits better upcoming AMD Bulldozer
665 # [which allegedly also implements SSSE3]...
667 # Temporary registers usage. X[2] is volatile at the entry and at the
668 # end is restored from backtrace ring buffer. X[3] is expected to
669 # contain current K_XX_XX constant and is used to caclulate X[-1]+K
670 # from previous round, it becomes volatile the moment the value is
671 # saved to stack for transfer to IALU. X[4] becomes volatile whenever
672 # X[-4] is accumulated and offloaded to backtrace ring buffer, at the
673 # end it is loaded with next K_XX_XX [which becomes X[3] in next
676 sub Xupdate_ssse3_16_31() # recall that $Xi starts wtih 4
679 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
682 eval(shift(@insns)); # ror
685 &punpcklqdq(@X[0],@X[-3&7]); # compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8);
686 &movdqa (@X[2],@X[-1&7]);
690 &paddd (@X[3],@X[-1&7]);
691 &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
692 eval(shift(@insns)); # rol
694 &psrldq (@X[2],4); # "X[-3]", 3 dwords
697 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
699 eval(shift(@insns)); # ror
701 &pxor (@X[2],@X[-2&7]); # "X[-3]"^"X[-8]"
706 &pxor (@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]"
708 eval(shift(@insns)); # rol
709 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
713 &movdqa (@X[4],@X[0]);
716 eval(shift(@insns)); # ror
717 &movdqa (@X[2],@X[0]);
720 &pslldq (@X[4],12); # "X[0]"<<96, extract one dword
721 &paddd (@X[0],@X[0]);
727 eval(shift(@insns)); # rol
728 &movdqa (@X[3],@X[4]);
735 eval(shift(@insns)); # ror
736 &por (@X[0],@X[2]); # "X[0]"<<<=1
738 &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer
744 eval(shift(@insns)); # rol
746 &movdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX
750 &pxor (@X[0],@X[3]); # "X[0]"^=("X[0]"<<96)<<<2
751 &pshufd (@X[1],@X[-3&7],0xee) if ($Xi<7); # was &movdqa (@X[1],@X[-2&7])
752 &pshufd (@X[3],@X[-1&7],0xee) if ($Xi==7);
756 foreach (@insns) { eval; } # remaining instructions [if any]
758 $Xi++; push(@X,shift(@X)); # "rotate" X[]
761 sub Xupdate_ssse3_32_79()
764 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
767 eval(shift(@insns)); # body_20_39
768 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
769 &punpcklqdq(@X[2],@X[-1&7]); # compose "X[-6]", was &palignr(@X[2],@X[-2&7],8)
772 eval(shift(@insns)); # rol
774 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
775 &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer
778 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
780 &movdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX...
781 } else { # ... or load next one
782 &movdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp"));
784 eval(shift(@insns)); # ror
785 &paddd (@X[3],@X[-1&7]);
788 &pxor (@X[0],@X[2]); # "X[0]"^="X[-6]"
789 eval(shift(@insns)); # body_20_39
792 eval(shift(@insns)); # rol
794 &movdqa (@X[2],@X[0]);
795 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
798 eval(shift(@insns)); # ror
800 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
803 eval(shift(@insns)); # body_20_39
807 eval(shift(@insns)); # rol
810 eval(shift(@insns)); # ror
812 eval(shift(@insns)) if (@insns[1] =~ /_rol/);
813 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
815 &por (@X[0],@X[2]); # "X[0]"<<<=2
816 eval(shift(@insns)); # body_20_39
818 &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer
820 eval(shift(@insns)); # rol
823 eval(shift(@insns)); # ror
824 &pshufd (@X[3],@X[-1],0xee) if ($Xi<19); # was &movdqa (@X[3],@X[0])
827 foreach (@insns) { eval; } # remaining instructions
829 $Xi++; push(@X,shift(@X)); # "rotate" X[]
832 sub Xuplast_ssse3_80()
835 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
845 &paddd (@X[3],@X[-1&7]);
851 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU
853 foreach (@insns) { eval; } # remaining instructions
855 &mov ($inp=@T[1],&DWP(192+4,"esp"));
856 &cmp ($inp,&DWP(192+8,"esp"));
857 &je (&label("done"));
859 &movdqa (@X[3],&QWP(112+48,"esp")); # K_00_19
860 &movdqa (@X[2],&QWP(112+64,"esp")); # pbswap mask
861 &movdqu (@X[-4&7],&QWP(0,$inp)); # load input
862 &movdqu (@X[-3&7],&QWP(16,$inp));
863 &movdqu (@X[-2&7],&QWP(32,$inp));
864 &movdqu (@X[-1&7],&QWP(48,$inp));
866 &pshufb (@X[-4&7],@X[2]); # byte swap
867 &mov (&DWP(192+4,"esp"),$inp);
868 &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
876 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
886 &pshufb (@X[($Xi-3)&7],@X[2]);
891 &paddd (@X[($Xi-4)&7],@X[3]);
896 &movdqa (&QWP(0+16*$Xi,"esp"),@X[($Xi-4)&7]); # X[]+K xfer to IALU
901 &psubd (@X[($Xi-4)&7],@X[3]);
903 foreach (@insns) { eval; }
910 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
913 foreach (@insns) { eval; }
916 sub body_00_19 () { # ((c^d)&b)^d
917 # on start @T[0]=(c^d)&b
918 return &body_20_39() if ($rx==19); $rx++;
920 '($a,$b,$c,$d,$e)=@V;'.
921 '&$_ror ($b,$j?7:2);', # $b>>>2
923 '&mov (@T[1],$a);', # $b in next round
925 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer
926 '&xor ($b,$c);', # $c^$d for next round
930 '&and (@T[1],$b);', # ($b&($c^$d)) for next round
932 '&xor ($b,$c);', # restore $b
933 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
937 sub body_20_39 () { # b^d^c
939 return &body_40_59() if ($rx==39); $rx++;
941 '($a,$b,$c,$d,$e)=@V;'.
942 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer
943 '&xor (@T[0],$d) if($j==19);'.
944 '&xor (@T[0],$c) if($j> 19);', # ($b^$d^$c)
945 '&mov (@T[1],$a);', # $b in next round
949 '&xor (@T[1],$c) if ($j< 79);', # $b^$d for next round
951 '&$_ror ($b,7);', # $b>>>2
952 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
956 sub body_40_59 () { # ((b^c)&(c^d))^c
957 # on entry @T[0]=(b^c), (c^=d)
960 '($a,$b,$c,$d,$e)=@V;'.
961 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer
962 '&and (@T[0],$c) if ($j>=40);', # (b^c)&(c^d)
963 '&xor ($c,$d) if ($j>=40);', # restore $c
965 '&$_ror ($b,7);', # $b>>>2
966 '&mov (@T[1],$a);', # $b for next round
971 '&xor (@T[1],$c) if ($j==59);'.
972 '&xor (@T[1],$b) if ($j< 59);', # b^c for next round
974 '&xor ($b,$c) if ($j< 59);', # c^d for next round
975 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
979 sub bodyx_00_19 () { # ((c^d)&b)^d
980 # on start @T[0]=(b&c)^(~b&d), $e+=X[]+K
981 return &bodyx_20_39() if ($rx==19); $rx++;
983 '($a,$b,$c,$d,$e)=@V;'.
985 '&rorx ($b,$b,2) if ($j==0);'. # $b>>>2
986 '&rorx ($b,@T[1],7) if ($j!=0);', # $b>>>2
987 '&lea ($e,&DWP(0,$e,@T[0]));',
988 '&rorx (@T[0],$a,5);',
990 '&andn (@T[1],$a,$c);',
992 '&add ($d,&DWP(4*(($j+1)&15),"esp"));', # X[]+K xfer
995 '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
999 sub bodyx_20_39 () { # b^d^c
1001 return &bodyx_40_59() if ($rx==39); $rx++;
1003 '($a,$b,$c,$d,$e)=@V;'.
1005 '&add ($e,($j==19?@T[0]:$b))',
1006 '&rorx ($b,@T[1],7);', # $b>>>2
1007 '&rorx (@T[0],$a,5);',
1009 '&xor ($a,$b) if ($j<79);',
1010 '&add ($d,&DWP(4*(($j+1)&15),"esp")) if ($j<79);', # X[]+K xfer
1011 '&xor ($a,$c) if ($j<79);',
1012 '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
1016 sub bodyx_40_59 () { # ((b^c)&(c^d))^c
1017 # on start $b=((b^c)&(c^d))^c
1018 return &bodyx_20_39() if ($rx==59); $rx++;
1020 '($a,$b,$c,$d,$e)=@V;'.
1022 '&rorx (@T[0],$a,5)',
1023 '&lea ($e,&DWP(0,$e,$b))',
1024 '&rorx ($b,@T[1],7)', # $b>>>2
1025 '&add ($d,&DWP(4*(($j+1)&15),"esp"))', # X[]+K xfer
1028 '&xor ($a,$b)', # b^c for next round
1029 '&xor (@T[1],$b)', # c^d for next round
1033 '&xor ($a,$b)' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
1037 &set_label("loop",16);
1038 &Xupdate_ssse3_16_31(\&body_00_19);
1039 &Xupdate_ssse3_16_31(\&body_00_19);
1040 &Xupdate_ssse3_16_31(\&body_00_19);
1041 &Xupdate_ssse3_16_31(\&body_00_19);
1042 &Xupdate_ssse3_32_79(\&body_00_19);
1043 &Xupdate_ssse3_32_79(\&body_20_39);
1044 &Xupdate_ssse3_32_79(\&body_20_39);
1045 &Xupdate_ssse3_32_79(\&body_20_39);
1046 &Xupdate_ssse3_32_79(\&body_20_39);
1047 &Xupdate_ssse3_32_79(\&body_20_39);
1048 &Xupdate_ssse3_32_79(\&body_40_59);
1049 &Xupdate_ssse3_32_79(\&body_40_59);
1050 &Xupdate_ssse3_32_79(\&body_40_59);
1051 &Xupdate_ssse3_32_79(\&body_40_59);
1052 &Xupdate_ssse3_32_79(\&body_40_59);
1053 &Xupdate_ssse3_32_79(\&body_20_39);
1054 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done"
1056 $saved_j=$j; @saved_V=@V;
1058 &Xloop_ssse3(\&body_20_39);
1059 &Xloop_ssse3(\&body_20_39);
1060 &Xloop_ssse3(\&body_20_39);
1062 &mov (@T[1],&DWP(192,"esp")); # update context
1063 &add ($A,&DWP(0,@T[1]));
1064 &add (@T[0],&DWP(4,@T[1])); # $b
1065 &add ($C,&DWP(8,@T[1]));
1066 &mov (&DWP(0,@T[1]),$A);
1067 &add ($D,&DWP(12,@T[1]));
1068 &mov (&DWP(4,@T[1]),@T[0]);
1069 &add ($E,&DWP(16,@T[1]));
1070 &mov (&DWP(8,@T[1]),$C);
1072 &mov (&DWP(12,@T[1]),$D);
1074 &mov (&DWP(16,@T[1]),$E);
1076 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
1080 &jmp (&label("loop"));
1082 &set_label("done",16); $j=$saved_j; @V=@saved_V;
1084 &Xtail_ssse3(\&body_20_39);
1085 &Xtail_ssse3(\&body_20_39);
1086 &Xtail_ssse3(\&body_20_39);
1088 &mov (@T[1],&DWP(192,"esp")); # update context
1089 &add ($A,&DWP(0,@T[1]));
1090 &mov ("esp",&DWP(192+12,"esp")); # restore %esp
1091 &add (@T[0],&DWP(4,@T[1])); # $b
1092 &add ($C,&DWP(8,@T[1]));
1093 &mov (&DWP(0,@T[1]),$A);
1094 &add ($D,&DWP(12,@T[1]));
1095 &mov (&DWP(4,@T[1]),@T[0]);
1096 &add ($E,&DWP(16,@T[1]));
1097 &mov (&DWP(8,@T[1]),$C);
1098 &mov (&DWP(12,@T[1]),$D);
1099 &mov (&DWP(16,@T[1]),$E);
1101 &function_end("_sha1_block_data_order_ssse3");
1106 my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded
1107 my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4
1108 my @V=($A,$B,$C,$D,$E);
1109 my $j=0; # hash round
1113 my $_rol=sub { &shld(@_[0],@_) };
1114 my $_ror=sub { &shrd(@_[0],@_) };
1116 &function_begin("_sha1_block_data_order_avx");
1117 &call (&label("pic_point")); # make it PIC!
1118 &set_label("pic_point");
1120 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
1121 &set_label("avx_shortcut");
1124 &vmovdqa(@X[3],&QWP(0,$tmp1)); # K_00_19
1125 &vmovdqa(@X[4],&QWP(16,$tmp1)); # K_20_39
1126 &vmovdqa(@X[5],&QWP(32,$tmp1)); # K_40_59
1127 &vmovdqa(@X[6],&QWP(48,$tmp1)); # K_60_79
1128 &vmovdqa(@X[2],&QWP(64,$tmp1)); # pbswap mask
1130 &mov ($E,&wparam(0)); # load argument block
1131 &mov ($inp=@T[1],&wparam(1));
1132 &mov ($D,&wparam(2));
1135 # stack frame layout
1137 # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area
1138 # X[4]+K X[5]+K X[6]+K X[7]+K
1139 # X[8]+K X[9]+K X[10]+K X[11]+K
1140 # X[12]+K X[13]+K X[14]+K X[15]+K
1142 # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area
1143 # X[4] X[5] X[6] X[7]
1144 # X[8] X[9] X[10] X[11] # even borrowed for K_00_19
1146 # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants
1147 # K_40_59 K_40_59 K_40_59 K_40_59
1148 # K_60_79 K_60_79 K_60_79 K_60_79
1149 # K_00_19 K_00_19 K_00_19 K_00_19
1152 # +192 ctx # argument block
1159 &vmovdqa(&QWP(112+0,"esp"),@X[4]); # copy constants
1160 &vmovdqa(&QWP(112+16,"esp"),@X[5]);
1161 &vmovdqa(&QWP(112+32,"esp"),@X[6]);
1162 &shl ($D,6); # len*64
1163 &vmovdqa(&QWP(112+48,"esp"),@X[3]);
1164 &add ($D,$inp); # end of input
1165 &vmovdqa(&QWP(112+64,"esp"),@X[2]);
1167 &mov (&DWP(192+0,"esp"),$E); # save argument block
1168 &mov (&DWP(192+4,"esp"),$inp);
1169 &mov (&DWP(192+8,"esp"),$D);
1170 &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp
1172 &mov ($A,&DWP(0,$E)); # load context
1173 &mov ($B,&DWP(4,$E));
1174 &mov ($C,&DWP(8,$E));
1175 &mov ($D,&DWP(12,$E));
1176 &mov ($E,&DWP(16,$E));
1177 &mov (@T[0],$B); # magic seed
1179 &vmovdqu(@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3]
1180 &vmovdqu(@X[-3&7],&QWP(-48,$inp));
1181 &vmovdqu(@X[-2&7],&QWP(-32,$inp));
1182 &vmovdqu(@X[-1&7],&QWP(-16,$inp));
1183 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1184 &vpshufb(@X[-3&7],@X[-3&7],@X[2]);
1185 &vpshufb(@X[-2&7],@X[-2&7],@X[2]);
1186 &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
1187 &vpshufb(@X[-1&7],@X[-1&7],@X[2]);
1188 &vpaddd (@X[0],@X[-4&7],@X[3]); # add K_00_19
1189 &vpaddd (@X[1],@X[-3&7],@X[3]);
1190 &vpaddd (@X[2],@X[-2&7],@X[3]);
1191 &vmovdqa(&QWP(0,"esp"),@X[0]); # X[]+K xfer to IALU
1193 &vmovdqa(&QWP(0+16,"esp"),@X[1]);
1195 &vmovdqa(&QWP(0+32,"esp"),@X[2]);
1197 &jmp (&label("loop"));
1199 sub Xupdate_avx_16_31() # recall that $Xi starts wtih 4
1202 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
1203 my ($a,$b,$c,$d,$e);
1205 eval(shift(@insns));
1206 eval(shift(@insns));
1207 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1208 eval(shift(@insns));
1209 eval(shift(@insns));
1211 &vpaddd (@X[3],@X[3],@X[-1&7]);
1212 &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
1213 eval(shift(@insns));
1214 eval(shift(@insns));
1215 &vpsrldq(@X[2],@X[-1&7],4); # "X[-3]", 3 dwords
1216 eval(shift(@insns));
1217 eval(shift(@insns));
1218 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1219 eval(shift(@insns));
1220 eval(shift(@insns));
1222 &vpxor (@X[2],@X[2],@X[-2&7]); # "X[-3]"^"X[-8]"
1223 eval(shift(@insns));
1224 eval(shift(@insns));
1225 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
1226 eval(shift(@insns));
1227 eval(shift(@insns));
1229 &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]"
1230 eval(shift(@insns));
1231 eval(shift(@insns));
1232 eval(shift(@insns));
1233 eval(shift(@insns));
1235 &vpsrld (@X[2],@X[0],31);
1236 eval(shift(@insns));
1237 eval(shift(@insns));
1238 eval(shift(@insns));
1239 eval(shift(@insns));
1241 &vpslldq(@X[4],@X[0],12); # "X[0]"<<96, extract one dword
1242 &vpaddd (@X[0],@X[0],@X[0]);
1243 eval(shift(@insns));
1244 eval(shift(@insns));
1245 eval(shift(@insns));
1246 eval(shift(@insns));
1248 &vpsrld (@X[3],@X[4],30);
1249 &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=1
1250 eval(shift(@insns));
1251 eval(shift(@insns));
1252 eval(shift(@insns));
1253 eval(shift(@insns));
1255 &vpslld (@X[4],@X[4],2);
1256 &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer
1257 eval(shift(@insns));
1258 eval(shift(@insns));
1259 &vpxor (@X[0],@X[0],@X[3]);
1260 eval(shift(@insns));
1261 eval(shift(@insns));
1262 eval(shift(@insns));
1263 eval(shift(@insns));
1265 &vpxor (@X[0],@X[0],@X[4]); # "X[0]"^=("X[0]"<<96)<<<2
1266 eval(shift(@insns));
1267 eval(shift(@insns));
1268 &vmovdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX
1269 eval(shift(@insns));
1270 eval(shift(@insns));
1272 foreach (@insns) { eval; } # remaining instructions [if any]
1274 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1277 sub Xupdate_avx_32_79()
1280 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
1281 my ($a,$b,$c,$d,$e);
1283 &vpalignr(@X[2],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1284 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1285 eval(shift(@insns)); # body_20_39
1286 eval(shift(@insns));
1287 eval(shift(@insns));
1288 eval(shift(@insns)); # rol
1290 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1291 &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer
1292 eval(shift(@insns));
1293 eval(shift(@insns));
1295 &vmovdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX...
1296 } else { # ... or load next one
1297 &vmovdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp"));
1299 &vpaddd (@X[3],@X[3],@X[-1&7]);
1300 eval(shift(@insns)); # ror
1301 eval(shift(@insns));
1303 &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-6]"
1304 eval(shift(@insns)); # body_20_39
1305 eval(shift(@insns));
1306 eval(shift(@insns));
1307 eval(shift(@insns)); # rol
1309 &vpsrld (@X[2],@X[0],30);
1310 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
1311 eval(shift(@insns));
1312 eval(shift(@insns));
1313 eval(shift(@insns)); # ror
1314 eval(shift(@insns));
1316 &vpslld (@X[0],@X[0],2);
1317 eval(shift(@insns)); # body_20_39
1318 eval(shift(@insns));
1319 eval(shift(@insns));
1320 eval(shift(@insns)); # rol
1321 eval(shift(@insns));
1322 eval(shift(@insns));
1323 eval(shift(@insns)); # ror
1324 eval(shift(@insns));
1326 &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=2
1327 eval(shift(@insns)); # body_20_39
1328 eval(shift(@insns));
1329 &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer
1330 eval(shift(@insns));
1331 eval(shift(@insns)); # rol
1332 eval(shift(@insns));
1333 eval(shift(@insns));
1334 eval(shift(@insns)); # ror
1335 eval(shift(@insns));
1337 foreach (@insns) { eval; } # remaining instructions
1339 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1342 sub Xuplast_avx_80()
1345 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1346 my ($a,$b,$c,$d,$e);
1348 eval(shift(@insns));
1349 &vpaddd (@X[3],@X[3],@X[-1&7]);
1350 eval(shift(@insns));
1351 eval(shift(@insns));
1352 eval(shift(@insns));
1353 eval(shift(@insns));
1355 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU
1357 foreach (@insns) { eval; } # remaining instructions
1359 &mov ($inp=@T[1],&DWP(192+4,"esp"));
1360 &cmp ($inp,&DWP(192+8,"esp"));
1361 &je (&label("done"));
1363 &vmovdqa(@X[3],&QWP(112+48,"esp")); # K_00_19
1364 &vmovdqa(@X[2],&QWP(112+64,"esp")); # pbswap mask
1365 &vmovdqu(@X[-4&7],&QWP(0,$inp)); # load input
1366 &vmovdqu(@X[-3&7],&QWP(16,$inp));
1367 &vmovdqu(@X[-2&7],&QWP(32,$inp));
1368 &vmovdqu(@X[-1&7],&QWP(48,$inp));
1370 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1371 &mov (&DWP(192+4,"esp"),$inp);
1372 &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
1380 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1381 my ($a,$b,$c,$d,$e);
1383 eval(shift(@insns));
1384 eval(shift(@insns));
1385 &vpshufb (@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1386 eval(shift(@insns));
1387 eval(shift(@insns));
1388 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],@X[3]);
1389 eval(shift(@insns));
1390 eval(shift(@insns));
1391 eval(shift(@insns));
1392 eval(shift(@insns));
1393 &vmovdqa (&QWP(0+16*$Xi,"esp"),@X[$Xi&7]); # X[]+K xfer to IALU
1394 eval(shift(@insns));
1395 eval(shift(@insns));
1397 foreach (@insns) { eval; }
1404 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1405 my ($a,$b,$c,$d,$e);
1407 foreach (@insns) { eval; }
1410 &set_label("loop",16);
1411 &Xupdate_avx_16_31(\&body_00_19);
1412 &Xupdate_avx_16_31(\&body_00_19);
1413 &Xupdate_avx_16_31(\&body_00_19);
1414 &Xupdate_avx_16_31(\&body_00_19);
1415 &Xupdate_avx_32_79(\&body_00_19);
1416 &Xupdate_avx_32_79(\&body_20_39);
1417 &Xupdate_avx_32_79(\&body_20_39);
1418 &Xupdate_avx_32_79(\&body_20_39);
1419 &Xupdate_avx_32_79(\&body_20_39);
1420 &Xupdate_avx_32_79(\&body_20_39);
1421 &Xupdate_avx_32_79(\&body_40_59);
1422 &Xupdate_avx_32_79(\&body_40_59);
1423 &Xupdate_avx_32_79(\&body_40_59);
1424 &Xupdate_avx_32_79(\&body_40_59);
1425 &Xupdate_avx_32_79(\&body_40_59);
1426 &Xupdate_avx_32_79(\&body_20_39);
1427 &Xuplast_avx_80(\&body_20_39); # can jump to "done"
1429 $saved_j=$j; @saved_V=@V;
1431 &Xloop_avx(\&body_20_39);
1432 &Xloop_avx(\&body_20_39);
1433 &Xloop_avx(\&body_20_39);
1435 &mov (@T[1],&DWP(192,"esp")); # update context
1436 &add ($A,&DWP(0,@T[1]));
1437 &add (@T[0],&DWP(4,@T[1])); # $b
1438 &add ($C,&DWP(8,@T[1]));
1439 &mov (&DWP(0,@T[1]),$A);
1440 &add ($D,&DWP(12,@T[1]));
1441 &mov (&DWP(4,@T[1]),@T[0]);
1442 &add ($E,&DWP(16,@T[1]));
1444 &mov (&DWP(8,@T[1]),$C);
1446 &mov (&DWP(12,@T[1]),$D);
1447 &mov (&DWP(16,@T[1]),$E);
1452 &jmp (&label("loop"));
1454 &set_label("done",16); $j=$saved_j; @V=@saved_V;
1456 &Xtail_avx(\&body_20_39);
1457 &Xtail_avx(\&body_20_39);
1458 &Xtail_avx(\&body_20_39);
1462 &mov (@T[1],&DWP(192,"esp")); # update context
1463 &add ($A,&DWP(0,@T[1]));
1464 &mov ("esp",&DWP(192+12,"esp")); # restore %esp
1465 &add (@T[0],&DWP(4,@T[1])); # $b
1466 &add ($C,&DWP(8,@T[1]));
1467 &mov (&DWP(0,@T[1]),$A);
1468 &add ($D,&DWP(12,@T[1]));
1469 &mov (&DWP(4,@T[1]),@T[0]);
1470 &add ($E,&DWP(16,@T[1]));
1471 &mov (&DWP(8,@T[1]),$C);
1472 &mov (&DWP(12,@T[1]),$D);
1473 &mov (&DWP(16,@T[1]),$E);
1474 &function_end("_sha1_block_data_order_avx");
1476 &set_label("K_XX_XX",64);
1477 &data_word(0x5a827999,0x5a827999,0x5a827999,0x5a827999); # K_00_19
1478 &data_word(0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1); # K_20_39
1479 &data_word(0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc); # K_40_59
1480 &data_word(0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6); # K_60_79
1481 &data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f); # pbswap mask
1482 &data_byte(0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0);
1484 &asciz("SHA1 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");