3 # ====================================================================
4 # [Re]written by Andy Polyakov <appro@openssl.org> 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 # "[Re]written" was achieved in two major overhauls. In 2004 BODY_*
11 # functions were re-implemented to address P4 performance issue [see
12 # commentary below], and in 2006 the rest was rewritten in order to
13 # gain freedom to liberate licensing terms.
15 # January, September 2004.
17 # It was noted that Intel IA-32 C compiler generates code which
18 # performs ~30% *faster* on P4 CPU than original *hand-coded*
19 # SHA1 assembler implementation. To address this problem (and
20 # prove that humans are still better than machines:-), the
21 # original code was overhauled, which resulted in following
22 # performance changes:
24 # compared with original compared with Intel cc
25 # assembler impl. generated code
30 # As you can see Pentium came out as looser:-( Yet I reckoned that
31 # improvement on P4 outweights the loss and incorporate this
32 # re-tuned code to 0.9.7 and later.
33 # ----------------------------------------------------------------
34 # <appro@fy.chalmers.se>
38 # George Spelvin has tipped that F_40_59(b,c,d) can be rewritten as
39 # '(c&d) + (b&(c^d))', which allows to accumulate partial results
40 # and lighten "pressure" on scratch registers. This resulted in
41 # >12% performance improvement on contemporary AMD cores (with no
42 # degradation on other CPUs:-). Also, the code was revised to maximize
43 # "distance" between instructions producing input to 'lea' instruction
44 # and the 'lea' instruction itself, which is essential for Intel Atom
45 # core and resulted in ~15% improvement.
49 # Add SSSE3, Supplemental[!] SSE3, implementation. The idea behind it
50 # is to offload message schedule denoted by Wt in NIST specification,
51 # or Xupdate in OpenSSL source, to SIMD unit. The idea is not novel,
52 # and in SSE2 context was first explored by Dean Gaudet in 2004, see
53 # http://arctic.org/~dean/crypto/sha1.html. Since then several things
54 # have changed that made it interesting again:
56 # a) XMM units became faster and wider;
57 # b) instruction set became more versatile;
58 # c) an important observation was made by Max Locktykhin, which made
59 # it possible to reduce amount of instructions required to perform
60 # the operation in question, for further details see
61 # http://software.intel.com/en-us/articles/improving-the-performance-of-the-secure-hash-algorithm-1/.
65 # Add AVX code path, probably most controversial... The thing is that
66 # switch to AVX alone improves performance by as little as 4% in
67 # comparison to SSSE3 code path. But below result doesn't look like
68 # 4% improvement... Trouble is that Sandy Bridge decodes 'ro[rl]' as
69 # pair of µ-ops, and it's the additional µ-ops, two per round, that
70 # make it run slower than Core2 and Westmere. But 'sh[rl]d' is decoded
71 # as single µ-op by Sandy Bridge and it's replacing 'ro[rl]' with
72 # equivalent 'sh[rl]d' that is responsible for the impressive 5.1
73 # cycles per processed byte. But 'sh[rl]d' is not something that used
74 # to be fast, nor does it appear to be fast in upcoming Bulldozer
75 # [according to its optimization manual]. Which is why AVX code path
76 # is guarded by *both* AVX and synthetic bit denoting Intel CPUs.
77 # One can argue that it's unfair to AMD, but without 'sh[rl]d' it
78 # makes no sense to keep the AVX code path. If somebody feels that
79 # strongly, it's probably more appropriate to discuss possibility of
80 # using vector rotate XOP on AMD...
84 # Add support for Intel SHA Extensions.
86 ######################################################################
87 # Current performance is summarized in following table. Numbers are
88 # CPU clock cycles spent to process single byte (less is better).
95 # Core2 7.3 6.0/+22% -
96 # Westmere 7.3 5.5/+33% -
97 # Sandy Bridge 8.8 6.2/+40% 5.1(**)/+73%
98 # Ivy Bridge 7.2 4.8/+51% 4.7(**)/+53%
99 # Haswell 6.5 4.3/+51% 4.1(**)/+58%
100 # Bulldozer 11.6 6.0/+92%
101 # VIA Nano 10.6 7.5/+41%
102 # Atom 12.5 9.3(*)/+35%
103 # Silvermont 14.5 9.9(*)/+46%
105 # (*) Loop is 1056 instructions long and expected result is ~8.25.
106 # The discrepancy is because of front-end limitations, so
107 # called MS-ROM penalties, and on Silvermont even rotate's
108 # limited parallelism.
110 # (**) As per above comment, the result is for AVX *plus* sh[rl]d.
112 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
113 push(@INC,"${dir}","${dir}../../perlasm");
117 open STDOUT,">$output";
119 &asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386");
122 for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
125 `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
126 =~ /GNU assembler version ([2-9]\.[0-9]+)/ &&
127 $1>=2.19); # first version supporting AVX
129 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32n" &&
130 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ &&
131 $1>=2.03); # first version supporting AVX
133 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32" &&
134 `ml 2>&1` =~ /Version ([0-9]+)\./ &&
135 $1>=10); # first version supporting AVX
137 $ymm=1 if ($xmm && !$ymm && `$ENV{CC} -v 2>&1` =~ /(^clang version|based on LLVM) ([3-9]\.[0-9]+)/ &&
138 $2>=3.0); # first version supporting AVX
140 $shaext=$xmm; ### set to zero if compiling for 1.0.1
142 &external_label("OPENSSL_ia32cap_P") if ($xmm);
153 @V=($A,$B,$C,$D,$E,$T);
155 $alt=0; # 1 denotes alternative IALU implementation, which performs
156 # 8% *worse* on P4, same on Westmere and Atom, 2% better on
161 local($n,$a,$b,$c,$d,$e,$f)=@_;
163 &comment("00_15 $n");
165 &mov($f,$c); # f to hold F_00_19(b,c,d)
166 if ($n==0) { &mov($tmp1,$a); }
167 else { &mov($a,$tmp1); }
168 &rotl($tmp1,5); # tmp1=ROTATE(a,5)
170 &add($tmp1,$e); # tmp1+=e;
171 &mov($e,&swtmp($n%16)); # e becomes volatile and is loaded
172 # with xi, also note that e becomes
175 &rotr($b,2); # b=ROTATE(b,30)
176 &xor($f,$d); # f holds F_00_19(b,c,d)
177 &lea($tmp1,&DWP(0x5a827999,$tmp1,$e)); # tmp1+=K_00_19+xi
179 if ($n==15) { &mov($e,&swtmp(($n+1)%16));# pre-fetch f for next round
180 &add($f,$tmp1); } # f+=tmp1
181 else { &add($tmp1,$f); } # f becomes a in next round
182 &mov($tmp1,$a) if ($alt && $n==15);
187 local($n,$a,$b,$c,$d,$e,$f)=@_;
189 &comment("16_19 $n");
193 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
194 &and($tmp1,$c); # tmp1 to hold F_00_19(b,c,d), b&=c^d
195 &xor($f,&swtmp(($n+8)%16));
196 &xor($tmp1,$d); # tmp1=F_00_19(b,c,d)
197 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
198 &rotl($f,1); # f=ROTATE(f,1)
199 &add($e,$tmp1); # e+=F_00_19(b,c,d)
200 &xor($c,$d); # restore $c
201 &mov($tmp1,$a); # b in next round
202 &rotr($b,$n==16?2:7); # b=ROTATE(b,30)
203 &mov(&swtmp($n%16),$f); # xi=f
204 &rotl($a,5); # ROTATE(a,5)
205 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
206 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
207 &add($f,$a); # f+=ROTATE(a,5)
209 &mov($tmp1,$c); # tmp1 to hold F_00_19(b,c,d)
210 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
212 &xor($f,&swtmp(($n+8)%16));
214 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
215 &rotl($f,1); # f=ROTATE(f,1)
216 &xor($tmp1,$d); # tmp1=F_00_19(b,c,d)
217 &add($e,$tmp1); # e+=F_00_19(b,c,d)
219 &rotr($b,2); # b=ROTATE(b,30)
220 &mov(&swtmp($n%16),$f); # xi=f
221 &rotl($tmp1,5); # ROTATE(a,5)
222 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
223 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
224 &add($f,$tmp1); # f+=ROTATE(a,5)
230 local($n,$a,$b,$c,$d,$e,$f)=@_;
231 local $K=($n<40)?0x6ed9eba1:0xca62c1d6;
233 &comment("20_39 $n");
236 &xor($tmp1,$c); # tmp1 to hold F_20_39(b,c,d), b^=c
237 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
238 &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d)
239 &xor($f,&swtmp(($n+8)%16));
240 &add($e,$tmp1); # e+=F_20_39(b,c,d)
241 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
242 &rotl($f,1); # f=ROTATE(f,1)
243 &mov($tmp1,$a); # b in next round
244 &rotr($b,7); # b=ROTATE(b,30)
245 &mov(&swtmp($n%16),$f) if($n<77);# xi=f
246 &rotl($a,5); # ROTATE(a,5)
247 &xor($b,$c) if($n==39);# warm up for BODY_40_59
248 &and($tmp1,$b) if($n==39);
249 &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY
250 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
251 &add($f,$a); # f+=ROTATE(a,5)
252 &rotr($a,5) if ($n==79);
254 &mov($tmp1,$b); # tmp1 to hold F_20_39(b,c,d)
255 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
257 &xor($f,&swtmp(($n+8)%16));
258 &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d)
259 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
260 &rotl($f,1); # f=ROTATE(f,1)
261 &add($e,$tmp1); # e+=F_20_39(b,c,d)
262 &rotr($b,2); # b=ROTATE(b,30)
264 &rotl($tmp1,5); # ROTATE(a,5)
265 &mov(&swtmp($n%16),$f) if($n<77);# xi=f
266 &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY
267 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
268 &add($f,$tmp1); # f+=ROTATE(a,5)
274 local($n,$a,$b,$c,$d,$e,$f)=@_;
276 &comment("40_59 $n");
279 &add($e,$tmp1); # e+=b&(c^d)
280 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
282 &xor($f,&swtmp(($n+8)%16));
283 &xor($c,$d); # restore $c
284 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
285 &rotl($f,1); # f=ROTATE(f,1)
287 &rotr($b,7); # b=ROTATE(b,30)
288 &add($e,$tmp1); # e+=c&d
289 &mov($tmp1,$a); # b in next round
290 &mov(&swtmp($n%16),$f); # xi=f
291 &rotl($a,5); # ROTATE(a,5)
292 &xor($b,$c) if ($n<59);
293 &and($tmp1,$b) if ($n<59);# tmp1 to hold F_40_59(b,c,d)
294 &lea($f,&DWP(0x8f1bbcdc,$f,$e));# f+=K_40_59+e+(b&(c^d))
295 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
296 &add($f,$a); # f+=ROTATE(a,5)
298 &mov($tmp1,$c); # tmp1 to hold F_40_59(b,c,d)
299 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
301 &xor($f,&swtmp(($n+8)%16));
303 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
304 &rotl($f,1); # f=ROTATE(f,1)
305 &add($tmp1,$e); # b&(c^d)+=e
306 &rotr($b,2); # b=ROTATE(b,30)
307 &mov($e,$a); # e becomes volatile
308 &rotl($e,5); # ROTATE(a,5)
309 &mov(&swtmp($n%16),$f); # xi=f
310 &lea($f,&DWP(0x8f1bbcdc,$f,$tmp1));# f+=K_40_59+e+(b&(c^d))
312 &add($f,$e); # f+=ROTATE(a,5)
314 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
315 &add($f,$tmp1); # f+=c&d
319 &function_begin("sha1_block_data_order");
321 &static_label("shaext_shortcut") if ($shaext);
322 &static_label("ssse3_shortcut");
323 &static_label("avx_shortcut") if ($ymm);
324 &static_label("K_XX_XX");
326 &call (&label("pic_point")); # make it PIC!
327 &set_label("pic_point");
329 &picmeup($T,"OPENSSL_ia32cap_P",$tmp1,&label("pic_point"));
330 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
332 &mov ($A,&DWP(0,$T));
333 &mov ($D,&DWP(4,$T));
334 &test ($D,1<<9); # check SSSE3 bit
336 &mov ($C,&DWP(8,$T));
337 &test ($A,1<<24); # check FXSR bit
340 &test ($C,1<<29); # check SHA bit
341 &jnz (&label("shaext_shortcut"));
344 &and ($D,1<<28); # mask AVX bit
345 &and ($A,1<<30); # mask "Intel CPU" bit
347 &cmp ($A,1<<28|1<<30);
348 &je (&label("avx_shortcut"));
350 &jmp (&label("ssse3_shortcut"));
351 &set_label("x86",16);
353 &mov($tmp1,&wparam(0)); # SHA_CTX *c
354 &mov($T,&wparam(1)); # const void *input
355 &mov($A,&wparam(2)); # size_t num
356 &stack_push(16+3); # allocate X[16]
359 &mov(&wparam(2),$A); # pointer beyond the end of input
360 &mov($E,&DWP(16,$tmp1));# pre-load E
361 &jmp(&label("loop"));
363 &set_label("loop",16);
365 # copy input chunk to X, but reversing byte order!
366 for ($i=0; $i<16; $i+=4)
368 &mov($A,&DWP(4*($i+0),$T));
369 &mov($B,&DWP(4*($i+1),$T));
370 &mov($C,&DWP(4*($i+2),$T));
371 &mov($D,&DWP(4*($i+3),$T));
376 &mov(&swtmp($i+0),$A);
377 &mov(&swtmp($i+1),$B);
378 &mov(&swtmp($i+2),$C);
379 &mov(&swtmp($i+3),$D);
381 &mov(&wparam(1),$T); # redundant in 1st spin
383 &mov($A,&DWP(0,$tmp1)); # load SHA_CTX
384 &mov($B,&DWP(4,$tmp1));
385 &mov($C,&DWP(8,$tmp1));
386 &mov($D,&DWP(12,$tmp1));
389 for($i=0;$i<16;$i++) { &BODY_00_15($i,@V); unshift(@V,pop(@V)); }
390 for(;$i<20;$i++) { &BODY_16_19($i,@V); unshift(@V,pop(@V)); }
391 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
392 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
393 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
395 (($V[5] eq $D) and ($V[0] eq $E)) or die; # double-check
397 &mov($tmp1,&wparam(0)); # re-load SHA_CTX*
398 &mov($D,&wparam(1)); # D is last "T" and is discarded
400 &add($E,&DWP(0,$tmp1)); # E is last "A"...
401 &add($T,&DWP(4,$tmp1));
402 &add($A,&DWP(8,$tmp1));
403 &add($B,&DWP(12,$tmp1));
404 &add($C,&DWP(16,$tmp1));
406 &mov(&DWP(0,$tmp1),$E); # update SHA_CTX
407 &add($D,64); # advance input pointer
408 &mov(&DWP(4,$tmp1),$T);
409 &cmp($D,&wparam(2)); # have we reached the end yet?
410 &mov(&DWP(8,$tmp1),$A);
411 &mov($E,$C); # C is last "E" which needs to be "pre-loaded"
412 &mov(&DWP(12,$tmp1),$B);
413 &mov($T,$D); # input pointer
414 &mov(&DWP(16,$tmp1),$C);
418 &function_end("sha1_block_data_order");
422 ######################################################################
423 # Intel SHA Extensions implementation of SHA1 update function.
425 my ($ctx,$inp,$num)=("edi","esi","ecx");
426 my ($ABCD,$E,$E_,$BSWAP)=map("xmm$_",(0..3));
427 my @MSG=map("xmm$_",(4..7));
430 my ($dst,$src,$imm)=@_;
431 if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
432 { &data_byte(0x0f,0x3a,0xcc,0xc0|($1<<3)|$2,$imm); }
435 my ($opcodelet,$dst,$src)=@_;
436 if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
437 { &data_byte(0x0f,0x38,$opcodelet,0xc0|($1<<3)|$2); }
439 sub sha1nexte { sha1op38(0xc8,@_); }
440 sub sha1msg1 { sha1op38(0xc9,@_); }
441 sub sha1msg2 { sha1op38(0xca,@_); }
443 &function_begin("_sha1_block_data_order_shaext");
444 &call (&label("pic_point")); # make it PIC!
445 &set_label("pic_point");
447 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
448 &set_label("shaext_shortcut");
449 &mov ($ctx,&wparam(0));
451 &mov ($inp,&wparam(1));
452 &mov ($num,&wparam(2));
455 &movdqu ($ABCD,&QWP(0,$ctx));
456 &movd ($E,&DWP(16,$ctx));
458 &movdqa ($BSWAP,&QWP(0x50,$tmp1)); # byte-n-word swap
460 &movdqu (@MSG[0],&QWP(0,$inp));
461 &pshufd ($ABCD,$ABCD,0b00011011); # flip word order
462 &movdqu (@MSG[1],&QWP(0x10,$inp));
463 &pshufd ($E,$E,0b00011011); # flip word order
464 &movdqu (@MSG[2],&QWP(0x20,$inp));
465 &pshufb (@MSG[0],$BSWAP);
466 &movdqu (@MSG[3],&QWP(0x30,$inp));
467 &pshufb (@MSG[1],$BSWAP);
468 &pshufb (@MSG[2],$BSWAP);
469 &pshufb (@MSG[3],$BSWAP);
470 &jmp (&label("loop_shaext"));
472 &set_label("loop_shaext",16);
474 &lea ("eax",&DWP(0x40,$inp));
475 &movdqa (&QWP(0,"esp"),$E); # offload $E
477 &cmovne ($inp,"eax");
478 &movdqa (&QWP(16,"esp"),$ABCD); # offload $ABCD
480 for($i=0;$i<20-4;$i+=2) {
481 &sha1msg1 (@MSG[0],@MSG[1]);
483 &sha1rnds4 ($ABCD,$E,int($i/5)); # 0-3...
484 &sha1nexte ($E_,@MSG[1]);
485 &pxor (@MSG[0],@MSG[2]);
486 &sha1msg1 (@MSG[1],@MSG[2]);
487 &sha1msg2 (@MSG[0],@MSG[3]);
490 &sha1rnds4 ($ABCD,$E_,int(($i+1)/5));
491 &sha1nexte ($E,@MSG[2]);
492 &pxor (@MSG[1],@MSG[3]);
493 &sha1msg2 (@MSG[1],@MSG[0]);
495 push(@MSG,shift(@MSG)); push(@MSG,shift(@MSG));
497 &movdqu (@MSG[0],&QWP(0,$inp));
499 &sha1rnds4 ($ABCD,$E,3); # 64-67
500 &sha1nexte ($E_,@MSG[1]);
501 &movdqu (@MSG[1],&QWP(0x10,$inp));
502 &pshufb (@MSG[0],$BSWAP);
505 &sha1rnds4 ($ABCD,$E_,3); # 68-71
506 &sha1nexte ($E,@MSG[2]);
507 &movdqu (@MSG[2],&QWP(0x20,$inp));
508 &pshufb (@MSG[1],$BSWAP);
511 &sha1rnds4 ($ABCD,$E,3); # 72-75
512 &sha1nexte ($E_,@MSG[3]);
513 &movdqu (@MSG[3],&QWP(0x30,$inp));
514 &pshufb (@MSG[2],$BSWAP);
517 &sha1rnds4 ($ABCD,$E_,3); # 76-79
518 &movdqa ($E_,&QWP(0,"esp"));
519 &pshufb (@MSG[3],$BSWAP);
521 &paddd ($ABCD,&QWP(16,"esp"));
523 &jnz (&label("loop_shaext"));
525 &pshufd ($ABCD,$ABCD,0b00011011);
526 &pshufd ($E,$E,0b00011011);
527 &movdqu (&QWP(0,$ctx),$ABCD)
528 &movd (&DWP(16,$ctx),$E);
530 &function_end("_sha1_block_data_order_shaext");
532 ######################################################################
533 # The SSSE3 implementation.
535 # %xmm[0-7] are used as ring @X[] buffer containing quadruples of last
536 # 32 elements of the message schedule or Xupdate outputs. First 4
537 # quadruples are simply byte-swapped input, next 4 are calculated
538 # according to method originally suggested by Dean Gaudet (modulo
539 # being implemented in SSSE3). Once 8 quadruples or 32 elements are
540 # collected, it switches to routine proposed by Max Locktyukhin.
542 # Calculations inevitably require temporary reqisters, and there are
543 # no %xmm registers left to spare. For this reason part of the ring
544 # buffer, X[2..4] to be specific, is offloaded to 3 quadriples ring
545 # buffer on the stack. Keep in mind that X[2] is alias X[-6], X[3] -
546 # X[-5], and X[4] - X[-4]...
548 # Another notable optimization is aggressive stack frame compression
549 # aiming to minimize amount of 9-byte instructions...
551 # Yet another notable optimization is "jumping" $B variable. It means
552 # that there is no register permanently allocated for $B value. This
553 # allowed to eliminate one instruction from body_20_39...
555 my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded
556 my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4
557 my @V=($A,$B,$C,$D,$E);
558 my $j=0; # hash round
563 my $_rol=sub { &rol(@_) };
564 my $_ror=sub { &ror(@_) };
566 &function_begin("_sha1_block_data_order_ssse3");
567 &call (&label("pic_point")); # make it PIC!
568 &set_label("pic_point");
570 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
571 &set_label("ssse3_shortcut");
573 &movdqa (@X[3],&QWP(0,$tmp1)); # K_00_19
574 &movdqa (@X[4],&QWP(16,$tmp1)); # K_20_39
575 &movdqa (@X[5],&QWP(32,$tmp1)); # K_40_59
576 &movdqa (@X[6],&QWP(48,$tmp1)); # K_60_79
577 &movdqa (@X[2],&QWP(64,$tmp1)); # pbswap mask
579 &mov ($E,&wparam(0)); # load argument block
580 &mov ($inp=@T[1],&wparam(1));
581 &mov ($D,&wparam(2));
586 # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area
587 # X[4]+K X[5]+K X[6]+K X[7]+K
588 # X[8]+K X[9]+K X[10]+K X[11]+K
589 # X[12]+K X[13]+K X[14]+K X[15]+K
591 # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area
592 # X[4] X[5] X[6] X[7]
593 # X[8] X[9] X[10] X[11] # even borrowed for K_00_19
595 # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants
596 # K_40_59 K_40_59 K_40_59 K_40_59
597 # K_60_79 K_60_79 K_60_79 K_60_79
598 # K_00_19 K_00_19 K_00_19 K_00_19
601 # +192 ctx # argument block
608 &movdqa (&QWP(112+0,"esp"),@X[4]); # copy constants
609 &movdqa (&QWP(112+16,"esp"),@X[5]);
610 &movdqa (&QWP(112+32,"esp"),@X[6]);
611 &shl ($D,6); # len*64
612 &movdqa (&QWP(112+48,"esp"),@X[3]);
613 &add ($D,$inp); # end of input
614 &movdqa (&QWP(112+64,"esp"),@X[2]);
616 &mov (&DWP(192+0,"esp"),$E); # save argument block
617 &mov (&DWP(192+4,"esp"),$inp);
618 &mov (&DWP(192+8,"esp"),$D);
619 &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp
621 &mov ($A,&DWP(0,$E)); # load context
622 &mov ($B,&DWP(4,$E));
623 &mov ($C,&DWP(8,$E));
624 &mov ($D,&DWP(12,$E));
625 &mov ($E,&DWP(16,$E));
626 &mov (@T[0],$B); # magic seed
628 &movdqu (@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3]
629 &movdqu (@X[-3&7],&QWP(-48,$inp));
630 &movdqu (@X[-2&7],&QWP(-32,$inp));
631 &movdqu (@X[-1&7],&QWP(-16,$inp));
632 &pshufb (@X[-4&7],@X[2]); # byte swap
633 &pshufb (@X[-3&7],@X[2]);
634 &pshufb (@X[-2&7],@X[2]);
635 &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
636 &pshufb (@X[-1&7],@X[2]);
637 &paddd (@X[-4&7],@X[3]); # add K_00_19
638 &paddd (@X[-3&7],@X[3]);
639 &paddd (@X[-2&7],@X[3]);
640 &movdqa (&QWP(0,"esp"),@X[-4&7]); # X[]+K xfer to IALU
641 &psubd (@X[-4&7],@X[3]); # restore X[]
642 &movdqa (&QWP(0+16,"esp"),@X[-3&7]);
643 &psubd (@X[-3&7],@X[3]);
644 &movdqa (&QWP(0+32,"esp"),@X[-2&7]);
646 &psubd (@X[-2&7],@X[3]);
648 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
650 &jmp (&label("loop"));
652 ######################################################################
653 # SSE instruction sequence is first broken to groups of indepentent
654 # instructions, independent in respect to their inputs and shifter
655 # (not all architectures have more than one). Then IALU instructions
656 # are "knitted in" between the SSE groups. Distance is maintained for
657 # SSE latency of 2 in hope that it fits better upcoming AMD Bulldozer
658 # [which allegedly also implements SSSE3]...
660 # Temporary registers usage. X[2] is volatile at the entry and at the
661 # end is restored from backtrace ring buffer. X[3] is expected to
662 # contain current K_XX_XX constant and is used to caclulate X[-1]+K
663 # from previous round, it becomes volatile the moment the value is
664 # saved to stack for transfer to IALU. X[4] becomes volatile whenever
665 # X[-4] is accumulated and offloaded to backtrace ring buffer, at the
666 # end it is loaded with next K_XX_XX [which becomes X[3] in next
669 sub Xupdate_ssse3_16_31() # recall that $Xi starts wtih 4
672 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
675 eval(shift(@insns)); # ror
678 &punpcklqdq(@X[0],@X[-3&7]); # compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8);
679 &movdqa (@X[2],@X[-1&7]);
683 &paddd (@X[3],@X[-1&7]);
684 &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
685 eval(shift(@insns)); # rol
687 &psrldq (@X[2],4); # "X[-3]", 3 dwords
690 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
692 eval(shift(@insns)); # ror
694 &pxor (@X[2],@X[-2&7]); # "X[-3]"^"X[-8]"
699 &pxor (@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]"
701 eval(shift(@insns)); # rol
702 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
706 &movdqa (@X[4],@X[0]);
709 eval(shift(@insns)); # ror
710 &movdqa (@X[2],@X[0]);
713 &pslldq (@X[4],12); # "X[0]"<<96, extract one dword
714 &paddd (@X[0],@X[0]);
720 eval(shift(@insns)); # rol
721 &movdqa (@X[3],@X[4]);
728 eval(shift(@insns)); # ror
729 &por (@X[0],@X[2]); # "X[0]"<<<=1
731 &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer
737 eval(shift(@insns)); # rol
739 &movdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX
743 &pxor (@X[0],@X[3]); # "X[0]"^=("X[0]"<<96)<<<2
744 &pshufd (@X[1],@X[-3&7],0xee) if ($Xi<7); # was &movdqa (@X[1],@X[-2&7])
745 &pshufd (@X[3],@X[-1&7],0xee) if ($Xi==7);
749 foreach (@insns) { eval; } # remaining instructions [if any]
751 $Xi++; push(@X,shift(@X)); # "rotate" X[]
754 sub Xupdate_ssse3_32_79()
757 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
760 eval(shift(@insns)); # body_20_39
761 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
762 &punpcklqdq(@X[2],@X[-1&7]); # compose "X[-6]", was &palignr(@X[2],@X[-2&7],8)
765 eval(shift(@insns)); # rol
767 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
768 &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer
771 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
773 &movdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX...
774 } else { # ... or load next one
775 &movdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp"));
777 eval(shift(@insns)); # ror
778 &paddd (@X[3],@X[-1&7]);
781 &pxor (@X[0],@X[2]); # "X[0]"^="X[-6]"
782 eval(shift(@insns)); # body_20_39
785 eval(shift(@insns)); # rol
787 &movdqa (@X[2],@X[0]);
788 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
791 eval(shift(@insns)); # ror
793 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
796 eval(shift(@insns)); # body_20_39
800 eval(shift(@insns)); # rol
803 eval(shift(@insns)); # ror
805 eval(shift(@insns)) if (@insns[1] =~ /_rol/);
806 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
808 &por (@X[0],@X[2]); # "X[0]"<<<=2
809 eval(shift(@insns)); # body_20_39
811 &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer
813 eval(shift(@insns)); # rol
816 eval(shift(@insns)); # ror
817 &pshufd (@X[3],@X[-1],0xee) if ($Xi<19); # was &movdqa (@X[3],@X[0])
820 foreach (@insns) { eval; } # remaining instructions
822 $Xi++; push(@X,shift(@X)); # "rotate" X[]
825 sub Xuplast_ssse3_80()
828 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
838 &paddd (@X[3],@X[-1&7]);
844 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU
846 foreach (@insns) { eval; } # remaining instructions
848 &mov ($inp=@T[1],&DWP(192+4,"esp"));
849 &cmp ($inp,&DWP(192+8,"esp"));
850 &je (&label("done"));
852 &movdqa (@X[3],&QWP(112+48,"esp")); # K_00_19
853 &movdqa (@X[2],&QWP(112+64,"esp")); # pbswap mask
854 &movdqu (@X[-4&7],&QWP(0,$inp)); # load input
855 &movdqu (@X[-3&7],&QWP(16,$inp));
856 &movdqu (@X[-2&7],&QWP(32,$inp));
857 &movdqu (@X[-1&7],&QWP(48,$inp));
859 &pshufb (@X[-4&7],@X[2]); # byte swap
860 &mov (&DWP(192+4,"esp"),$inp);
861 &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
869 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
879 &pshufb (@X[($Xi-3)&7],@X[2]);
884 &paddd (@X[($Xi-4)&7],@X[3]);
889 &movdqa (&QWP(0+16*$Xi,"esp"),@X[($Xi-4)&7]); # X[]+K xfer to IALU
894 &psubd (@X[($Xi-4)&7],@X[3]);
896 foreach (@insns) { eval; }
903 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
906 foreach (@insns) { eval; }
909 sub body_00_19 () { # ((c^d)&b)^d
910 # on start @T[0]=(c^d)&b
911 return &body_20_39() if ($rx==19); $rx++;
913 '($a,$b,$c,$d,$e)=@V;'.
914 '&$_ror ($b,$j?7:2);', # $b>>>2
916 '&mov (@T[1],$a);', # $b in next round
918 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer
919 '&xor ($b,$c);', # $c^$d for next round
923 '&and (@T[1],$b);', # ($b&($c^$d)) for next round
925 '&xor ($b,$c);', # restore $b
926 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
930 sub body_20_39 () { # b^d^c
932 return &body_40_59() if ($rx==39); $rx++;
934 '($a,$b,$c,$d,$e)=@V;'.
935 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer
936 '&xor (@T[0],$d) if($j==19);'.
937 '&xor (@T[0],$c) if($j> 19);', # ($b^$d^$c)
938 '&mov (@T[1],$a);', # $b in next round
942 '&xor (@T[1],$c) if ($j< 79);', # $b^$d for next round
944 '&$_ror ($b,7);', # $b>>>2
945 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
949 sub body_40_59 () { # ((b^c)&(c^d))^c
950 # on entry @T[0]=(b^c), (c^=d)
953 '($a,$b,$c,$d,$e)=@V;'.
954 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer
955 '&and (@T[0],$c) if ($j>=40);', # (b^c)&(c^d)
956 '&xor ($c,$d) if ($j>=40);', # restore $c
958 '&$_ror ($b,7);', # $b>>>2
959 '&mov (@T[1],$a);', # $b for next round
964 '&xor (@T[1],$c) if ($j==59);'.
965 '&xor (@T[1],$b) if ($j< 59);', # b^c for next round
967 '&xor ($b,$c) if ($j< 59);', # c^d for next round
968 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
972 sub bodyx_00_19 () { # ((c^d)&b)^d
973 # on start @T[0]=(b&c)^(~b&d), $e+=X[]+K
974 return &bodyx_20_39() if ($rx==19); $rx++;
976 '($a,$b,$c,$d,$e)=@V;'.
978 '&rorx ($b,$b,2) if ($j==0);'. # $b>>>2
979 '&rorx ($b,@T[1],7) if ($j!=0);', # $b>>>2
980 '&lea ($e,&DWP(0,$e,@T[0]));',
981 '&rorx (@T[0],$a,5);',
983 '&andn (@T[1],$a,$c);',
985 '&add ($d,&DWP(4*(($j+1)&15),"esp"));', # X[]+K xfer
988 '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
992 sub bodyx_20_39 () { # b^d^c
994 return &bodyx_40_59() if ($rx==39); $rx++;
996 '($a,$b,$c,$d,$e)=@V;'.
998 '&add ($e,($j==19?@T[0]:$b))',
999 '&rorx ($b,@T[1],7);', # $b>>>2
1000 '&rorx (@T[0],$a,5);',
1002 '&xor ($a,$b) if ($j<79);',
1003 '&add ($d,&DWP(4*(($j+1)&15),"esp")) if ($j<79);', # X[]+K xfer
1004 '&xor ($a,$c) if ($j<79);',
1005 '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
1009 sub bodyx_40_59 () { # ((b^c)&(c^d))^c
1010 # on start $b=((b^c)&(c^d))^c
1011 return &bodyx_20_39() if ($rx==59); $rx++;
1013 '($a,$b,$c,$d,$e)=@V;'.
1015 '&rorx (@T[0],$a,5)',
1016 '&lea ($e,&DWP(0,$e,$b))',
1017 '&rorx ($b,@T[1],7)', # $b>>>2
1018 '&add ($d,&DWP(4*(($j+1)&15),"esp"))', # X[]+K xfer
1021 '&xor ($a,$b)', # b^c for next round
1022 '&xor (@T[1],$b)', # c^d for next round
1026 '&xor ($a,$b)' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
1030 &set_label("loop",16);
1031 &Xupdate_ssse3_16_31(\&body_00_19);
1032 &Xupdate_ssse3_16_31(\&body_00_19);
1033 &Xupdate_ssse3_16_31(\&body_00_19);
1034 &Xupdate_ssse3_16_31(\&body_00_19);
1035 &Xupdate_ssse3_32_79(\&body_00_19);
1036 &Xupdate_ssse3_32_79(\&body_20_39);
1037 &Xupdate_ssse3_32_79(\&body_20_39);
1038 &Xupdate_ssse3_32_79(\&body_20_39);
1039 &Xupdate_ssse3_32_79(\&body_20_39);
1040 &Xupdate_ssse3_32_79(\&body_20_39);
1041 &Xupdate_ssse3_32_79(\&body_40_59);
1042 &Xupdate_ssse3_32_79(\&body_40_59);
1043 &Xupdate_ssse3_32_79(\&body_40_59);
1044 &Xupdate_ssse3_32_79(\&body_40_59);
1045 &Xupdate_ssse3_32_79(\&body_40_59);
1046 &Xupdate_ssse3_32_79(\&body_20_39);
1047 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done"
1049 $saved_j=$j; @saved_V=@V;
1051 &Xloop_ssse3(\&body_20_39);
1052 &Xloop_ssse3(\&body_20_39);
1053 &Xloop_ssse3(\&body_20_39);
1055 &mov (@T[1],&DWP(192,"esp")); # update context
1056 &add ($A,&DWP(0,@T[1]));
1057 &add (@T[0],&DWP(4,@T[1])); # $b
1058 &add ($C,&DWP(8,@T[1]));
1059 &mov (&DWP(0,@T[1]),$A);
1060 &add ($D,&DWP(12,@T[1]));
1061 &mov (&DWP(4,@T[1]),@T[0]);
1062 &add ($E,&DWP(16,@T[1]));
1063 &mov (&DWP(8,@T[1]),$C);
1065 &mov (&DWP(12,@T[1]),$D);
1067 &mov (&DWP(16,@T[1]),$E);
1069 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
1073 &jmp (&label("loop"));
1075 &set_label("done",16); $j=$saved_j; @V=@saved_V;
1077 &Xtail_ssse3(\&body_20_39);
1078 &Xtail_ssse3(\&body_20_39);
1079 &Xtail_ssse3(\&body_20_39);
1081 &mov (@T[1],&DWP(192,"esp")); # update context
1082 &add ($A,&DWP(0,@T[1]));
1083 &mov ("esp",&DWP(192+12,"esp")); # restore %esp
1084 &add (@T[0],&DWP(4,@T[1])); # $b
1085 &add ($C,&DWP(8,@T[1]));
1086 &mov (&DWP(0,@T[1]),$A);
1087 &add ($D,&DWP(12,@T[1]));
1088 &mov (&DWP(4,@T[1]),@T[0]);
1089 &add ($E,&DWP(16,@T[1]));
1090 &mov (&DWP(8,@T[1]),$C);
1091 &mov (&DWP(12,@T[1]),$D);
1092 &mov (&DWP(16,@T[1]),$E);
1094 &function_end("_sha1_block_data_order_ssse3");
1099 my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded
1100 my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4
1101 my @V=($A,$B,$C,$D,$E);
1102 my $j=0; # hash round
1106 my $_rol=sub { &shld(@_[0],@_) };
1107 my $_ror=sub { &shrd(@_[0],@_) };
1109 &function_begin("_sha1_block_data_order_avx");
1110 &call (&label("pic_point")); # make it PIC!
1111 &set_label("pic_point");
1113 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
1114 &set_label("avx_shortcut");
1117 &vmovdqa(@X[3],&QWP(0,$tmp1)); # K_00_19
1118 &vmovdqa(@X[4],&QWP(16,$tmp1)); # K_20_39
1119 &vmovdqa(@X[5],&QWP(32,$tmp1)); # K_40_59
1120 &vmovdqa(@X[6],&QWP(48,$tmp1)); # K_60_79
1121 &vmovdqa(@X[2],&QWP(64,$tmp1)); # pbswap mask
1123 &mov ($E,&wparam(0)); # load argument block
1124 &mov ($inp=@T[1],&wparam(1));
1125 &mov ($D,&wparam(2));
1128 # stack frame layout
1130 # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area
1131 # X[4]+K X[5]+K X[6]+K X[7]+K
1132 # X[8]+K X[9]+K X[10]+K X[11]+K
1133 # X[12]+K X[13]+K X[14]+K X[15]+K
1135 # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area
1136 # X[4] X[5] X[6] X[7]
1137 # X[8] X[9] X[10] X[11] # even borrowed for K_00_19
1139 # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants
1140 # K_40_59 K_40_59 K_40_59 K_40_59
1141 # K_60_79 K_60_79 K_60_79 K_60_79
1142 # K_00_19 K_00_19 K_00_19 K_00_19
1145 # +192 ctx # argument block
1152 &vmovdqa(&QWP(112+0,"esp"),@X[4]); # copy constants
1153 &vmovdqa(&QWP(112+16,"esp"),@X[5]);
1154 &vmovdqa(&QWP(112+32,"esp"),@X[6]);
1155 &shl ($D,6); # len*64
1156 &vmovdqa(&QWP(112+48,"esp"),@X[3]);
1157 &add ($D,$inp); # end of input
1158 &vmovdqa(&QWP(112+64,"esp"),@X[2]);
1160 &mov (&DWP(192+0,"esp"),$E); # save argument block
1161 &mov (&DWP(192+4,"esp"),$inp);
1162 &mov (&DWP(192+8,"esp"),$D);
1163 &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp
1165 &mov ($A,&DWP(0,$E)); # load context
1166 &mov ($B,&DWP(4,$E));
1167 &mov ($C,&DWP(8,$E));
1168 &mov ($D,&DWP(12,$E));
1169 &mov ($E,&DWP(16,$E));
1170 &mov (@T[0],$B); # magic seed
1172 &vmovdqu(@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3]
1173 &vmovdqu(@X[-3&7],&QWP(-48,$inp));
1174 &vmovdqu(@X[-2&7],&QWP(-32,$inp));
1175 &vmovdqu(@X[-1&7],&QWP(-16,$inp));
1176 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1177 &vpshufb(@X[-3&7],@X[-3&7],@X[2]);
1178 &vpshufb(@X[-2&7],@X[-2&7],@X[2]);
1179 &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
1180 &vpshufb(@X[-1&7],@X[-1&7],@X[2]);
1181 &vpaddd (@X[0],@X[-4&7],@X[3]); # add K_00_19
1182 &vpaddd (@X[1],@X[-3&7],@X[3]);
1183 &vpaddd (@X[2],@X[-2&7],@X[3]);
1184 &vmovdqa(&QWP(0,"esp"),@X[0]); # X[]+K xfer to IALU
1186 &vmovdqa(&QWP(0+16,"esp"),@X[1]);
1188 &vmovdqa(&QWP(0+32,"esp"),@X[2]);
1190 &jmp (&label("loop"));
1192 sub Xupdate_avx_16_31() # recall that $Xi starts wtih 4
1195 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
1196 my ($a,$b,$c,$d,$e);
1198 eval(shift(@insns));
1199 eval(shift(@insns));
1200 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1201 eval(shift(@insns));
1202 eval(shift(@insns));
1204 &vpaddd (@X[3],@X[3],@X[-1&7]);
1205 &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
1206 eval(shift(@insns));
1207 eval(shift(@insns));
1208 &vpsrldq(@X[2],@X[-1&7],4); # "X[-3]", 3 dwords
1209 eval(shift(@insns));
1210 eval(shift(@insns));
1211 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1212 eval(shift(@insns));
1213 eval(shift(@insns));
1215 &vpxor (@X[2],@X[2],@X[-2&7]); # "X[-3]"^"X[-8]"
1216 eval(shift(@insns));
1217 eval(shift(@insns));
1218 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
1219 eval(shift(@insns));
1220 eval(shift(@insns));
1222 &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]"
1223 eval(shift(@insns));
1224 eval(shift(@insns));
1225 eval(shift(@insns));
1226 eval(shift(@insns));
1228 &vpsrld (@X[2],@X[0],31);
1229 eval(shift(@insns));
1230 eval(shift(@insns));
1231 eval(shift(@insns));
1232 eval(shift(@insns));
1234 &vpslldq(@X[4],@X[0],12); # "X[0]"<<96, extract one dword
1235 &vpaddd (@X[0],@X[0],@X[0]);
1236 eval(shift(@insns));
1237 eval(shift(@insns));
1238 eval(shift(@insns));
1239 eval(shift(@insns));
1241 &vpsrld (@X[3],@X[4],30);
1242 &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=1
1243 eval(shift(@insns));
1244 eval(shift(@insns));
1245 eval(shift(@insns));
1246 eval(shift(@insns));
1248 &vpslld (@X[4],@X[4],2);
1249 &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer
1250 eval(shift(@insns));
1251 eval(shift(@insns));
1252 &vpxor (@X[0],@X[0],@X[3]);
1253 eval(shift(@insns));
1254 eval(shift(@insns));
1255 eval(shift(@insns));
1256 eval(shift(@insns));
1258 &vpxor (@X[0],@X[0],@X[4]); # "X[0]"^=("X[0]"<<96)<<<2
1259 eval(shift(@insns));
1260 eval(shift(@insns));
1261 &vmovdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX
1262 eval(shift(@insns));
1263 eval(shift(@insns));
1265 foreach (@insns) { eval; } # remaining instructions [if any]
1267 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1270 sub Xupdate_avx_32_79()
1273 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
1274 my ($a,$b,$c,$d,$e);
1276 &vpalignr(@X[2],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1277 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1278 eval(shift(@insns)); # body_20_39
1279 eval(shift(@insns));
1280 eval(shift(@insns));
1281 eval(shift(@insns)); # rol
1283 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1284 &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer
1285 eval(shift(@insns));
1286 eval(shift(@insns));
1288 &vmovdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX...
1289 } else { # ... or load next one
1290 &vmovdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp"));
1292 &vpaddd (@X[3],@X[3],@X[-1&7]);
1293 eval(shift(@insns)); # ror
1294 eval(shift(@insns));
1296 &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-6]"
1297 eval(shift(@insns)); # body_20_39
1298 eval(shift(@insns));
1299 eval(shift(@insns));
1300 eval(shift(@insns)); # rol
1302 &vpsrld (@X[2],@X[0],30);
1303 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
1304 eval(shift(@insns));
1305 eval(shift(@insns));
1306 eval(shift(@insns)); # ror
1307 eval(shift(@insns));
1309 &vpslld (@X[0],@X[0],2);
1310 eval(shift(@insns)); # body_20_39
1311 eval(shift(@insns));
1312 eval(shift(@insns));
1313 eval(shift(@insns)); # rol
1314 eval(shift(@insns));
1315 eval(shift(@insns));
1316 eval(shift(@insns)); # ror
1317 eval(shift(@insns));
1319 &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=2
1320 eval(shift(@insns)); # body_20_39
1321 eval(shift(@insns));
1322 &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer
1323 eval(shift(@insns));
1324 eval(shift(@insns)); # rol
1325 eval(shift(@insns));
1326 eval(shift(@insns));
1327 eval(shift(@insns)); # ror
1328 eval(shift(@insns));
1330 foreach (@insns) { eval; } # remaining instructions
1332 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1335 sub Xuplast_avx_80()
1338 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1339 my ($a,$b,$c,$d,$e);
1341 eval(shift(@insns));
1342 &vpaddd (@X[3],@X[3],@X[-1&7]);
1343 eval(shift(@insns));
1344 eval(shift(@insns));
1345 eval(shift(@insns));
1346 eval(shift(@insns));
1348 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU
1350 foreach (@insns) { eval; } # remaining instructions
1352 &mov ($inp=@T[1],&DWP(192+4,"esp"));
1353 &cmp ($inp,&DWP(192+8,"esp"));
1354 &je (&label("done"));
1356 &vmovdqa(@X[3],&QWP(112+48,"esp")); # K_00_19
1357 &vmovdqa(@X[2],&QWP(112+64,"esp")); # pbswap mask
1358 &vmovdqu(@X[-4&7],&QWP(0,$inp)); # load input
1359 &vmovdqu(@X[-3&7],&QWP(16,$inp));
1360 &vmovdqu(@X[-2&7],&QWP(32,$inp));
1361 &vmovdqu(@X[-1&7],&QWP(48,$inp));
1363 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1364 &mov (&DWP(192+4,"esp"),$inp);
1365 &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
1373 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1374 my ($a,$b,$c,$d,$e);
1376 eval(shift(@insns));
1377 eval(shift(@insns));
1378 &vpshufb (@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1379 eval(shift(@insns));
1380 eval(shift(@insns));
1381 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],@X[3]);
1382 eval(shift(@insns));
1383 eval(shift(@insns));
1384 eval(shift(@insns));
1385 eval(shift(@insns));
1386 &vmovdqa (&QWP(0+16*$Xi,"esp"),@X[$Xi&7]); # X[]+K xfer to IALU
1387 eval(shift(@insns));
1388 eval(shift(@insns));
1390 foreach (@insns) { eval; }
1397 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1398 my ($a,$b,$c,$d,$e);
1400 foreach (@insns) { eval; }
1403 &set_label("loop",16);
1404 &Xupdate_avx_16_31(\&body_00_19);
1405 &Xupdate_avx_16_31(\&body_00_19);
1406 &Xupdate_avx_16_31(\&body_00_19);
1407 &Xupdate_avx_16_31(\&body_00_19);
1408 &Xupdate_avx_32_79(\&body_00_19);
1409 &Xupdate_avx_32_79(\&body_20_39);
1410 &Xupdate_avx_32_79(\&body_20_39);
1411 &Xupdate_avx_32_79(\&body_20_39);
1412 &Xupdate_avx_32_79(\&body_20_39);
1413 &Xupdate_avx_32_79(\&body_20_39);
1414 &Xupdate_avx_32_79(\&body_40_59);
1415 &Xupdate_avx_32_79(\&body_40_59);
1416 &Xupdate_avx_32_79(\&body_40_59);
1417 &Xupdate_avx_32_79(\&body_40_59);
1418 &Xupdate_avx_32_79(\&body_40_59);
1419 &Xupdate_avx_32_79(\&body_20_39);
1420 &Xuplast_avx_80(\&body_20_39); # can jump to "done"
1422 $saved_j=$j; @saved_V=@V;
1424 &Xloop_avx(\&body_20_39);
1425 &Xloop_avx(\&body_20_39);
1426 &Xloop_avx(\&body_20_39);
1428 &mov (@T[1],&DWP(192,"esp")); # update context
1429 &add ($A,&DWP(0,@T[1]));
1430 &add (@T[0],&DWP(4,@T[1])); # $b
1431 &add ($C,&DWP(8,@T[1]));
1432 &mov (&DWP(0,@T[1]),$A);
1433 &add ($D,&DWP(12,@T[1]));
1434 &mov (&DWP(4,@T[1]),@T[0]);
1435 &add ($E,&DWP(16,@T[1]));
1437 &mov (&DWP(8,@T[1]),$C);
1439 &mov (&DWP(12,@T[1]),$D);
1440 &mov (&DWP(16,@T[1]),$E);
1445 &jmp (&label("loop"));
1447 &set_label("done",16); $j=$saved_j; @V=@saved_V;
1449 &Xtail_avx(\&body_20_39);
1450 &Xtail_avx(\&body_20_39);
1451 &Xtail_avx(\&body_20_39);
1455 &mov (@T[1],&DWP(192,"esp")); # update context
1456 &add ($A,&DWP(0,@T[1]));
1457 &mov ("esp",&DWP(192+12,"esp")); # restore %esp
1458 &add (@T[0],&DWP(4,@T[1])); # $b
1459 &add ($C,&DWP(8,@T[1]));
1460 &mov (&DWP(0,@T[1]),$A);
1461 &add ($D,&DWP(12,@T[1]));
1462 &mov (&DWP(4,@T[1]),@T[0]);
1463 &add ($E,&DWP(16,@T[1]));
1464 &mov (&DWP(8,@T[1]),$C);
1465 &mov (&DWP(12,@T[1]),$D);
1466 &mov (&DWP(16,@T[1]),$E);
1467 &function_end("_sha1_block_data_order_avx");
1469 &set_label("K_XX_XX",64);
1470 &data_word(0x5a827999,0x5a827999,0x5a827999,0x5a827999); # K_00_19
1471 &data_word(0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1); # K_20_39
1472 &data_word(0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc); # K_40_59
1473 &data_word(0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6); # K_60_79
1474 &data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f); # pbswap mask
1475 &data_byte(0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0);
1477 &asciz("SHA1 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");