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...
82 ######################################################################
83 # Current performance is summarized in following table. Numbers are
84 # CPU clock cycles spent to process single byte (less is better).
91 # Core2 7.3 6.0/+22% -
92 # Atom 12.5 9.3(*)/+35% -
93 # Westmere 7.3 5.5/+33% -
94 # Sandy Bridge 8.8 6.2/+40% 5.1(**)/+73%
95 # Ivy Bridge 7.2 4.8/+51% 4.7(**)/+53%
96 # Haswell 6.5 4.3/+51% 4.1(**)/+58%
97 # Bulldozer 11.6 6.0/+92%
98 # VIA Nano 10.6 7.5/+41%
100 # (*) Loop is 1056 instructions long and expected result is ~8.25.
101 # It remains mystery [to me] why ILP is limited to 1.7.
103 # (**) As per above comment, the result is for AVX *plus* sh[rl]d.
105 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
106 push(@INC,"${dir}","${dir}../../perlasm");
109 &asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386");
112 for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
115 `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
116 =~ /GNU assembler version ([2-9]\.[0-9]+)/ &&
117 $1>=2.19); # first version supporting AVX
119 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32n" &&
120 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ &&
121 $1>=2.03); # first version supporting AVX
123 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32" &&
124 `ml 2>&1` =~ /Version ([0-9]+)\./ &&
125 $1>=10); # first version supporting AVX
127 &external_label("OPENSSL_ia32cap_P") if ($xmm);
138 @V=($A,$B,$C,$D,$E,$T);
140 $alt=0; # 1 denotes alternative IALU implementation, which performs
141 # 8% *worse* on P4, same on Westmere and Atom, 2% better on
146 local($n,$a,$b,$c,$d,$e,$f)=@_;
148 &comment("00_15 $n");
150 &mov($f,$c); # f to hold F_00_19(b,c,d)
151 if ($n==0) { &mov($tmp1,$a); }
152 else { &mov($a,$tmp1); }
153 &rotl($tmp1,5); # tmp1=ROTATE(a,5)
155 &add($tmp1,$e); # tmp1+=e;
156 &mov($e,&swtmp($n%16)); # e becomes volatile and is loaded
157 # with xi, also note that e becomes
160 &rotr($b,2); # b=ROTATE(b,30)
161 &xor($f,$d); # f holds F_00_19(b,c,d)
162 &lea($tmp1,&DWP(0x5a827999,$tmp1,$e)); # tmp1+=K_00_19+xi
164 if ($n==15) { &mov($e,&swtmp(($n+1)%16));# pre-fetch f for next round
165 &add($f,$tmp1); } # f+=tmp1
166 else { &add($tmp1,$f); } # f becomes a in next round
167 &mov($tmp1,$a) if ($alt && $n==15);
172 local($n,$a,$b,$c,$d,$e,$f)=@_;
174 &comment("16_19 $n");
178 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
179 &and($tmp1,$c); # tmp1 to hold F_00_19(b,c,d), b&=c^d
180 &xor($f,&swtmp(($n+8)%16));
181 &xor($tmp1,$d); # tmp1=F_00_19(b,c,d)
182 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
183 &rotl($f,1); # f=ROTATE(f,1)
184 &add($e,$tmp1); # e+=F_00_19(b,c,d)
185 &xor($c,$d); # restore $c
186 &mov($tmp1,$a); # b in next round
187 &rotr($b,$n==16?2:7); # b=ROTATE(b,30)
188 &mov(&swtmp($n%16),$f); # xi=f
189 &rotl($a,5); # ROTATE(a,5)
190 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
191 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
192 &add($f,$a); # f+=ROTATE(a,5)
194 &mov($tmp1,$c); # tmp1 to hold F_00_19(b,c,d)
195 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
197 &xor($f,&swtmp(($n+8)%16));
199 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
200 &rotl($f,1); # f=ROTATE(f,1)
201 &xor($tmp1,$d); # tmp1=F_00_19(b,c,d)
202 &add($e,$tmp1); # e+=F_00_19(b,c,d)
204 &rotr($b,2); # b=ROTATE(b,30)
205 &mov(&swtmp($n%16),$f); # xi=f
206 &rotl($tmp1,5); # ROTATE(a,5)
207 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
208 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
209 &add($f,$tmp1); # f+=ROTATE(a,5)
215 local($n,$a,$b,$c,$d,$e,$f)=@_;
216 local $K=($n<40)?0x6ed9eba1:0xca62c1d6;
218 &comment("20_39 $n");
221 &xor($tmp1,$c); # tmp1 to hold F_20_39(b,c,d), b^=c
222 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
223 &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d)
224 &xor($f,&swtmp(($n+8)%16));
225 &add($e,$tmp1); # e+=F_20_39(b,c,d)
226 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
227 &rotl($f,1); # f=ROTATE(f,1)
228 &mov($tmp1,$a); # b in next round
229 &rotr($b,7); # b=ROTATE(b,30)
230 &mov(&swtmp($n%16),$f) if($n<77);# xi=f
231 &rotl($a,5); # ROTATE(a,5)
232 &xor($b,$c) if($n==39);# warm up for BODY_40_59
233 &and($tmp1,$b) if($n==39);
234 &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY
235 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
236 &add($f,$a); # f+=ROTATE(a,5)
237 &rotr($a,5) if ($n==79);
239 &mov($tmp1,$b); # tmp1 to hold F_20_39(b,c,d)
240 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
242 &xor($f,&swtmp(($n+8)%16));
243 &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d)
244 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
245 &rotl($f,1); # f=ROTATE(f,1)
246 &add($e,$tmp1); # e+=F_20_39(b,c,d)
247 &rotr($b,2); # b=ROTATE(b,30)
249 &rotl($tmp1,5); # ROTATE(a,5)
250 &mov(&swtmp($n%16),$f) if($n<77);# xi=f
251 &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY
252 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
253 &add($f,$tmp1); # f+=ROTATE(a,5)
259 local($n,$a,$b,$c,$d,$e,$f)=@_;
261 &comment("40_59 $n");
264 &add($e,$tmp1); # e+=b&(c^d)
265 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
267 &xor($f,&swtmp(($n+8)%16));
268 &xor($c,$d); # restore $c
269 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
270 &rotl($f,1); # f=ROTATE(f,1)
272 &rotr($b,7); # b=ROTATE(b,30)
273 &add($e,$tmp1); # e+=c&d
274 &mov($tmp1,$a); # b in next round
275 &mov(&swtmp($n%16),$f); # xi=f
276 &rotl($a,5); # ROTATE(a,5)
277 &xor($b,$c) if ($n<59);
278 &and($tmp1,$b) if ($n<59);# tmp1 to hold F_40_59(b,c,d)
279 &lea($f,&DWP(0x8f1bbcdc,$f,$e));# f+=K_40_59+e+(b&(c^d))
280 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
281 &add($f,$a); # f+=ROTATE(a,5)
283 &mov($tmp1,$c); # tmp1 to hold F_40_59(b,c,d)
284 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
286 &xor($f,&swtmp(($n+8)%16));
288 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
289 &rotl($f,1); # f=ROTATE(f,1)
290 &add($tmp1,$e); # b&(c^d)+=e
291 &rotr($b,2); # b=ROTATE(b,30)
292 &mov($e,$a); # e becomes volatile
293 &rotl($e,5); # ROTATE(a,5)
294 &mov(&swtmp($n%16),$f); # xi=f
295 &lea($f,&DWP(0x8f1bbcdc,$f,$tmp1));# f+=K_40_59+e+(b&(c^d))
297 &add($f,$e); # f+=ROTATE(a,5)
299 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
300 &add($f,$tmp1); # f+=c&d
304 &function_begin("sha1_block_data_order");
306 &static_label("ssse3_shortcut");
307 &static_label("avx_shortcut") if ($ymm);
308 &static_label("K_XX_XX");
310 &call (&label("pic_point")); # make it PIC!
311 &set_label("pic_point");
313 &picmeup($T,"OPENSSL_ia32cap_P",$tmp1,&label("pic_point"));
314 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
316 &mov ($A,&DWP(0,$T));
317 &mov ($D,&DWP(4,$T));
318 &test ($D,1<<9); # check SSSE3 bit
320 &test ($A,1<<24); # check FXSR bit
323 &and ($D,1<<28); # mask AVX bit
324 &and ($A,1<<30); # mask "Intel CPU" bit
326 &cmp ($A,1<<28|1<<30);
327 &je (&label("avx_shortcut"));
329 &jmp (&label("ssse3_shortcut"));
330 &set_label("x86",16);
332 &mov($tmp1,&wparam(0)); # SHA_CTX *c
333 &mov($T,&wparam(1)); # const void *input
334 &mov($A,&wparam(2)); # size_t num
335 &stack_push(16+3); # allocate X[16]
338 &mov(&wparam(2),$A); # pointer beyond the end of input
339 &mov($E,&DWP(16,$tmp1));# pre-load E
340 &jmp(&label("loop"));
342 &set_label("loop",16);
344 # copy input chunk to X, but reversing byte order!
345 for ($i=0; $i<16; $i+=4)
347 &mov($A,&DWP(4*($i+0),$T));
348 &mov($B,&DWP(4*($i+1),$T));
349 &mov($C,&DWP(4*($i+2),$T));
350 &mov($D,&DWP(4*($i+3),$T));
355 &mov(&swtmp($i+0),$A);
356 &mov(&swtmp($i+1),$B);
357 &mov(&swtmp($i+2),$C);
358 &mov(&swtmp($i+3),$D);
360 &mov(&wparam(1),$T); # redundant in 1st spin
362 &mov($A,&DWP(0,$tmp1)); # load SHA_CTX
363 &mov($B,&DWP(4,$tmp1));
364 &mov($C,&DWP(8,$tmp1));
365 &mov($D,&DWP(12,$tmp1));
368 for($i=0;$i<16;$i++) { &BODY_00_15($i,@V); unshift(@V,pop(@V)); }
369 for(;$i<20;$i++) { &BODY_16_19($i,@V); unshift(@V,pop(@V)); }
370 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
371 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
372 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
374 (($V[5] eq $D) and ($V[0] eq $E)) or die; # double-check
376 &mov($tmp1,&wparam(0)); # re-load SHA_CTX*
377 &mov($D,&wparam(1)); # D is last "T" and is discarded
379 &add($E,&DWP(0,$tmp1)); # E is last "A"...
380 &add($T,&DWP(4,$tmp1));
381 &add($A,&DWP(8,$tmp1));
382 &add($B,&DWP(12,$tmp1));
383 &add($C,&DWP(16,$tmp1));
385 &mov(&DWP(0,$tmp1),$E); # update SHA_CTX
386 &add($D,64); # advance input pointer
387 &mov(&DWP(4,$tmp1),$T);
388 &cmp($D,&wparam(2)); # have we reached the end yet?
389 &mov(&DWP(8,$tmp1),$A);
390 &mov($E,$C); # C is last "E" which needs to be "pre-loaded"
391 &mov(&DWP(12,$tmp1),$B);
392 &mov($T,$D); # input pointer
393 &mov(&DWP(16,$tmp1),$C);
397 &function_end("sha1_block_data_order");
400 ######################################################################
401 # The SSSE3 implementation.
403 # %xmm[0-7] are used as ring @X[] buffer containing quadruples of last
404 # 32 elements of the message schedule or Xupdate outputs. First 4
405 # quadruples are simply byte-swapped input, next 4 are calculated
406 # according to method originally suggested by Dean Gaudet (modulo
407 # being implemented in SSSE3). Once 8 quadruples or 32 elements are
408 # collected, it switches to routine proposed by Max Locktyukhin.
410 # Calculations inevitably require temporary reqisters, and there are
411 # no %xmm registers left to spare. For this reason part of the ring
412 # buffer, X[2..4] to be specific, is offloaded to 3 quadriples ring
413 # buffer on the stack. Keep in mind that X[2] is alias X[-6], X[3] -
414 # X[-5], and X[4] - X[-4]...
416 # Another notable optimization is aggressive stack frame compression
417 # aiming to minimize amount of 9-byte instructions...
419 # Yet another notable optimization is "jumping" $B variable. It means
420 # that there is no register permanently allocated for $B value. This
421 # allowed to eliminate one instruction from body_20_39...
423 my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded
424 my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4
425 my @V=($A,$B,$C,$D,$E);
426 my $j=0; # hash round
431 my $_rol=sub { &rol(@_) };
432 my $_ror=sub { &ror(@_) };
434 &function_begin("_sha1_block_data_order_ssse3");
435 &call (&label("pic_point")); # make it PIC!
436 &set_label("pic_point");
438 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
439 &set_label("ssse3_shortcut");
441 &movdqa (@X[3],&QWP(0,$tmp1)); # K_00_19
442 &movdqa (@X[4],&QWP(16,$tmp1)); # K_20_39
443 &movdqa (@X[5],&QWP(32,$tmp1)); # K_40_59
444 &movdqa (@X[6],&QWP(48,$tmp1)); # K_60_79
445 &movdqa (@X[2],&QWP(64,$tmp1)); # pbswap mask
447 &mov ($E,&wparam(0)); # load argument block
448 &mov ($inp=@T[1],&wparam(1));
449 &mov ($D,&wparam(2));
454 # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area
455 # X[4]+K X[5]+K X[6]+K X[7]+K
456 # X[8]+K X[9]+K X[10]+K X[11]+K
457 # X[12]+K X[13]+K X[14]+K X[15]+K
459 # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area
460 # X[4] X[5] X[6] X[7]
461 # X[8] X[9] X[10] X[11] # even borrowed for K_00_19
463 # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants
464 # K_40_59 K_40_59 K_40_59 K_40_59
465 # K_60_79 K_60_79 K_60_79 K_60_79
466 # K_00_19 K_00_19 K_00_19 K_00_19
469 # +192 ctx # argument block
476 &movdqa (&QWP(112+0,"esp"),@X[4]); # copy constants
477 &movdqa (&QWP(112+16,"esp"),@X[5]);
478 &movdqa (&QWP(112+32,"esp"),@X[6]);
479 &shl ($D,6); # len*64
480 &movdqa (&QWP(112+48,"esp"),@X[3]);
481 &add ($D,$inp); # end of input
482 &movdqa (&QWP(112+64,"esp"),@X[2]);
484 &mov (&DWP(192+0,"esp"),$E); # save argument block
485 &mov (&DWP(192+4,"esp"),$inp);
486 &mov (&DWP(192+8,"esp"),$D);
487 &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp
489 &mov ($A,&DWP(0,$E)); # load context
490 &mov ($B,&DWP(4,$E));
491 &mov ($C,&DWP(8,$E));
492 &mov ($D,&DWP(12,$E));
493 &mov ($E,&DWP(16,$E));
494 &mov (@T[0],$B); # magic seed
496 &movdqu (@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3]
497 &movdqu (@X[-3&7],&QWP(-48,$inp));
498 &movdqu (@X[-2&7],&QWP(-32,$inp));
499 &movdqu (@X[-1&7],&QWP(-16,$inp));
500 &pshufb (@X[-4&7],@X[2]); # byte swap
501 &pshufb (@X[-3&7],@X[2]);
502 &pshufb (@X[-2&7],@X[2]);
503 &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
504 &pshufb (@X[-1&7],@X[2]);
505 &paddd (@X[-4&7],@X[3]); # add K_00_19
506 &paddd (@X[-3&7],@X[3]);
507 &paddd (@X[-2&7],@X[3]);
508 &movdqa (&QWP(0,"esp"),@X[-4&7]); # X[]+K xfer to IALU
509 &psubd (@X[-4&7],@X[3]); # restore X[]
510 &movdqa (&QWP(0+16,"esp"),@X[-3&7]);
511 &psubd (@X[-3&7],@X[3]);
512 &movdqa (&QWP(0+32,"esp"),@X[-2&7]);
514 &psubd (@X[-2&7],@X[3]);
516 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
518 &jmp (&label("loop"));
520 ######################################################################
521 # SSE instruction sequence is first broken to groups of indepentent
522 # instructions, independent in respect to their inputs and shifter
523 # (not all architectures have more than one). Then IALU instructions
524 # are "knitted in" between the SSE groups. Distance is maintained for
525 # SSE latency of 2 in hope that it fits better upcoming AMD Bulldozer
526 # [which allegedly also implements SSSE3]...
528 # Temporary registers usage. X[2] is volatile at the entry and at the
529 # end is restored from backtrace ring buffer. X[3] is expected to
530 # contain current K_XX_XX constant and is used to caclulate X[-1]+K
531 # from previous round, it becomes volatile the moment the value is
532 # saved to stack for transfer to IALU. X[4] becomes volatile whenever
533 # X[-4] is accumulated and offloaded to backtrace ring buffer, at the
534 # end it is loaded with next K_XX_XX [which becomes X[3] in next
537 sub Xupdate_ssse3_16_31() # recall that $Xi starts wtih 4
540 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
543 eval(shift(@insns)); # ror
546 &punpcklqdq(@X[0],@X[-3&7]); # compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8);
547 &movdqa (@X[2],@X[-1&7]);
551 &paddd (@X[3],@X[-1&7]);
552 &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
553 eval(shift(@insns)); # rol
555 &psrldq (@X[2],4); # "X[-3]", 3 dwords
558 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
560 eval(shift(@insns)); # ror
562 &pxor (@X[2],@X[-2&7]); # "X[-3]"^"X[-8]"
567 &pxor (@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]"
569 eval(shift(@insns)); # rol
570 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
574 &movdqa (@X[4],@X[0]);
577 eval(shift(@insns)); # ror
578 &movdqa (@X[2],@X[0]);
581 &pslldq (@X[4],12); # "X[0]"<<96, extract one dword
582 &paddd (@X[0],@X[0]);
588 eval(shift(@insns)); # rol
589 &movdqa (@X[3],@X[4]);
596 eval(shift(@insns)); # ror
597 &por (@X[0],@X[2]); # "X[0]"<<<=1
599 &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer
605 eval(shift(@insns)); # rol
607 &movdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX
611 &pxor (@X[0],@X[3]); # "X[0]"^=("X[0]"<<96)<<<2
612 &pshufd (@X[1],@X[-3&7],0xee) if ($Xi<7); # was &movdqa (@X[1],@X[-2&7])
613 &pshufd (@X[3],@X[-1&7],0xee) if ($Xi==7);
617 foreach (@insns) { eval; } # remaining instructions [if any]
619 $Xi++; push(@X,shift(@X)); # "rotate" X[]
622 sub Xupdate_ssse3_32_79()
625 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
628 eval(shift(@insns)); # body_20_39
629 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
630 &punpcklqdq(@X[2],@X[-1&7]); # compose "X[-6]", was &palignr(@X[2],@X[-2&7],8)
633 eval(shift(@insns)); # rol
635 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
636 &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer
639 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
641 &movdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX...
642 } else { # ... or load next one
643 &movdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp"));
645 eval(shift(@insns)); # ror
646 &paddd (@X[3],@X[-1&7]);
649 &pxor (@X[0],@X[2]); # "X[0]"^="X[-6]"
650 eval(shift(@insns)); # body_20_39
653 eval(shift(@insns)); # rol
655 &movdqa (@X[2],@X[0]);
656 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
659 eval(shift(@insns)); # ror
661 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
664 eval(shift(@insns)); # body_20_39
668 eval(shift(@insns)); # rol
671 eval(shift(@insns)); # ror
673 eval(shift(@insns)) if (@insns[1] =~ /_rol/);
674 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
676 &por (@X[0],@X[2]); # "X[0]"<<<=2
677 eval(shift(@insns)); # body_20_39
679 &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer
681 eval(shift(@insns)); # rol
684 eval(shift(@insns)); # ror
685 &pshufd (@X[3],@X[-1],0xee) if ($Xi<19); # was &movdqa (@X[3],@X[0])
688 foreach (@insns) { eval; } # remaining instructions
690 $Xi++; push(@X,shift(@X)); # "rotate" X[]
693 sub Xuplast_ssse3_80()
696 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
706 &paddd (@X[3],@X[-1&7]);
712 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU
714 foreach (@insns) { eval; } # remaining instructions
716 &mov ($inp=@T[1],&DWP(192+4,"esp"));
717 &cmp ($inp,&DWP(192+8,"esp"));
718 &je (&label("done"));
720 &movdqa (@X[3],&QWP(112+48,"esp")); # K_00_19
721 &movdqa (@X[2],&QWP(112+64,"esp")); # pbswap mask
722 &movdqu (@X[-4&7],&QWP(0,$inp)); # load input
723 &movdqu (@X[-3&7],&QWP(16,$inp));
724 &movdqu (@X[-2&7],&QWP(32,$inp));
725 &movdqu (@X[-1&7],&QWP(48,$inp));
727 &pshufb (@X[-4&7],@X[2]); # byte swap
728 &mov (&DWP(192+4,"esp"),$inp);
729 &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
737 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
747 &pshufb (@X[($Xi-3)&7],@X[2]);
752 &paddd (@X[($Xi-4)&7],@X[3]);
757 &movdqa (&QWP(0+16*$Xi,"esp"),@X[($Xi-4)&7]); # X[]+K xfer to IALU
762 &psubd (@X[($Xi-4)&7],@X[3]);
764 foreach (@insns) { eval; }
771 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
774 foreach (@insns) { eval; }
777 sub body_00_19 () { # ((c^d)&b)^d
778 # on start @T[0]=(c^d)&b
779 return &body_20_39() if ($rx==19); $rx++;
781 '($a,$b,$c,$d,$e)=@V;'.
782 '&$_ror ($b,$j?7:2);', # $b>>>2
784 '&mov (@T[1],$a);', # $b in next round
786 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer
787 '&xor ($b,$c);', # $c^$d for next round
791 '&and (@T[1],$b);', # ($b&($c^$d)) for next round
793 '&xor ($b,$c);', # restore $b
794 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
798 sub body_20_39 () { # b^d^c
800 return &body_40_59() if ($rx==39); $rx++;
802 '($a,$b,$c,$d,$e)=@V;'.
803 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer
804 '&xor (@T[0],$d) if($j==19);'.
805 '&xor (@T[0],$c) if($j> 19);', # ($b^$d^$c)
806 '&mov (@T[1],$a);', # $b in next round
810 '&xor (@T[1],$c) if ($j< 79);', # $b^$d for next round
812 '&$_ror ($b,7);', # $b>>>2
813 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
817 sub body_40_59 () { # ((b^c)&(c^d))^c
818 # on entry @T[0]=(b^c), (c^=d)
821 '($a,$b,$c,$d,$e)=@V;'.
822 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer
823 '&and (@T[0],$c) if ($j>=40);', # (b^c)&(c^d)
824 '&xor ($c,$d) if ($j>=40);', # restore $c
826 '&$_ror ($b,7);', # $b>>>2
827 '&mov (@T[1],$a);', # $b for next round
832 '&xor (@T[1],$c) if ($j==59);'.
833 '&xor (@T[1],$b) if ($j< 59);', # b^c for next round
835 '&xor ($b,$c) if ($j< 59);', # c^d for next round
836 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
840 sub bodyx_00_19 () { # ((c^d)&b)^d
841 # on start @T[0]=(b&c)^(~b&d), $e+=X[]+K
842 return &bodyx_20_39() if ($rx==19); $rx++;
844 '($a,$b,$c,$d,$e)=@V;'.
846 '&rorx ($b,$b,2) if ($j==0);'. # $b>>>2
847 '&rorx ($b,@T[1],7) if ($j!=0);', # $b>>>2
848 '&lea ($e,&DWP(0,$e,@T[0]));',
849 '&rorx (@T[0],$a,5);',
851 '&andn (@T[1],$a,$c);',
853 '&add ($d,&DWP(4*(($j+1)&15),"esp"));', # X[]+K xfer
856 '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
860 sub bodyx_20_39 () { # b^d^c
862 return &bodyx_40_59() if ($rx==39); $rx++;
864 '($a,$b,$c,$d,$e)=@V;'.
866 '&add ($e,($j==19?@T[0]:$b))',
867 '&rorx ($b,@T[1],7);', # $b>>>2
868 '&rorx (@T[0],$a,5);',
870 '&xor ($a,$b) if ($j<79);',
871 '&add ($d,&DWP(4*(($j+1)&15),"esp")) if ($j<79);', # X[]+K xfer
872 '&xor ($a,$c) if ($j<79);',
873 '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
877 sub bodyx_40_59 () { # ((b^c)&(c^d))^c
878 # on start $b=((b^c)&(c^d))^c
879 return &bodyx_20_39() if ($rx==59); $rx++;
881 '($a,$b,$c,$d,$e)=@V;'.
883 '&rorx (@T[0],$a,5)',
884 '&lea ($e,&DWP(0,$e,$b))',
885 '&rorx ($b,@T[1],7)', # $b>>>2
886 '&add ($d,&DWP(4*(($j+1)&15),"esp"))', # X[]+K xfer
889 '&xor ($a,$b)', # b^c for next round
890 '&xor (@T[1],$b)', # c^d for next round
894 '&xor ($a,$b)' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
898 &set_label("loop",16);
899 &Xupdate_ssse3_16_31(\&body_00_19);
900 &Xupdate_ssse3_16_31(\&body_00_19);
901 &Xupdate_ssse3_16_31(\&body_00_19);
902 &Xupdate_ssse3_16_31(\&body_00_19);
903 &Xupdate_ssse3_32_79(\&body_00_19);
904 &Xupdate_ssse3_32_79(\&body_20_39);
905 &Xupdate_ssse3_32_79(\&body_20_39);
906 &Xupdate_ssse3_32_79(\&body_20_39);
907 &Xupdate_ssse3_32_79(\&body_20_39);
908 &Xupdate_ssse3_32_79(\&body_20_39);
909 &Xupdate_ssse3_32_79(\&body_40_59);
910 &Xupdate_ssse3_32_79(\&body_40_59);
911 &Xupdate_ssse3_32_79(\&body_40_59);
912 &Xupdate_ssse3_32_79(\&body_40_59);
913 &Xupdate_ssse3_32_79(\&body_40_59);
914 &Xupdate_ssse3_32_79(\&body_20_39);
915 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done"
917 $saved_j=$j; @saved_V=@V;
919 &Xloop_ssse3(\&body_20_39);
920 &Xloop_ssse3(\&body_20_39);
921 &Xloop_ssse3(\&body_20_39);
923 &mov (@T[1],&DWP(192,"esp")); # update context
924 &add ($A,&DWP(0,@T[1]));
925 &add (@T[0],&DWP(4,@T[1])); # $b
926 &add ($C,&DWP(8,@T[1]));
927 &mov (&DWP(0,@T[1]),$A);
928 &add ($D,&DWP(12,@T[1]));
929 &mov (&DWP(4,@T[1]),@T[0]);
930 &add ($E,&DWP(16,@T[1]));
931 &mov (&DWP(8,@T[1]),$C);
933 &mov (&DWP(12,@T[1]),$D);
935 &mov (&DWP(16,@T[1]),$E);
937 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
941 &jmp (&label("loop"));
943 &set_label("done",16); $j=$saved_j; @V=@saved_V;
945 &Xtail_ssse3(\&body_20_39);
946 &Xtail_ssse3(\&body_20_39);
947 &Xtail_ssse3(\&body_20_39);
949 &mov (@T[1],&DWP(192,"esp")); # update context
950 &add ($A,&DWP(0,@T[1]));
951 &mov ("esp",&DWP(192+12,"esp")); # restore %esp
952 &add (@T[0],&DWP(4,@T[1])); # $b
953 &add ($C,&DWP(8,@T[1]));
954 &mov (&DWP(0,@T[1]),$A);
955 &add ($D,&DWP(12,@T[1]));
956 &mov (&DWP(4,@T[1]),@T[0]);
957 &add ($E,&DWP(16,@T[1]));
958 &mov (&DWP(8,@T[1]),$C);
959 &mov (&DWP(12,@T[1]),$D);
960 &mov (&DWP(16,@T[1]),$E);
962 &function_end("_sha1_block_data_order_ssse3");
967 my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded
968 my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4
969 my @V=($A,$B,$C,$D,$E);
970 my $j=0; # hash round
974 my $_rol=sub { &shld(@_[0],@_) };
975 my $_ror=sub { &shrd(@_[0],@_) };
977 &function_begin("_sha1_block_data_order_avx");
978 &call (&label("pic_point")); # make it PIC!
979 &set_label("pic_point");
981 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
982 &set_label("avx_shortcut");
985 &vmovdqa(@X[3],&QWP(0,$tmp1)); # K_00_19
986 &vmovdqa(@X[4],&QWP(16,$tmp1)); # K_20_39
987 &vmovdqa(@X[5],&QWP(32,$tmp1)); # K_40_59
988 &vmovdqa(@X[6],&QWP(48,$tmp1)); # K_60_79
989 &vmovdqa(@X[2],&QWP(64,$tmp1)); # pbswap mask
991 &mov ($E,&wparam(0)); # load argument block
992 &mov ($inp=@T[1],&wparam(1));
993 &mov ($D,&wparam(2));
998 # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area
999 # X[4]+K X[5]+K X[6]+K X[7]+K
1000 # X[8]+K X[9]+K X[10]+K X[11]+K
1001 # X[12]+K X[13]+K X[14]+K X[15]+K
1003 # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area
1004 # X[4] X[5] X[6] X[7]
1005 # X[8] X[9] X[10] X[11] # even borrowed for K_00_19
1007 # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants
1008 # K_40_59 K_40_59 K_40_59 K_40_59
1009 # K_60_79 K_60_79 K_60_79 K_60_79
1010 # K_00_19 K_00_19 K_00_19 K_00_19
1013 # +192 ctx # argument block
1020 &vmovdqa(&QWP(112+0,"esp"),@X[4]); # copy constants
1021 &vmovdqa(&QWP(112+16,"esp"),@X[5]);
1022 &vmovdqa(&QWP(112+32,"esp"),@X[6]);
1023 &shl ($D,6); # len*64
1024 &vmovdqa(&QWP(112+48,"esp"),@X[3]);
1025 &add ($D,$inp); # end of input
1026 &vmovdqa(&QWP(112+64,"esp"),@X[2]);
1028 &mov (&DWP(192+0,"esp"),$E); # save argument block
1029 &mov (&DWP(192+4,"esp"),$inp);
1030 &mov (&DWP(192+8,"esp"),$D);
1031 &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp
1033 &mov ($A,&DWP(0,$E)); # load context
1034 &mov ($B,&DWP(4,$E));
1035 &mov ($C,&DWP(8,$E));
1036 &mov ($D,&DWP(12,$E));
1037 &mov ($E,&DWP(16,$E));
1038 &mov (@T[0],$B); # magic seed
1040 &vmovdqu(@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3]
1041 &vmovdqu(@X[-3&7],&QWP(-48,$inp));
1042 &vmovdqu(@X[-2&7],&QWP(-32,$inp));
1043 &vmovdqu(@X[-1&7],&QWP(-16,$inp));
1044 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1045 &vpshufb(@X[-3&7],@X[-3&7],@X[2]);
1046 &vpshufb(@X[-2&7],@X[-2&7],@X[2]);
1047 &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
1048 &vpshufb(@X[-1&7],@X[-1&7],@X[2]);
1049 &vpaddd (@X[0],@X[-4&7],@X[3]); # add K_00_19
1050 &vpaddd (@X[1],@X[-3&7],@X[3]);
1051 &vpaddd (@X[2],@X[-2&7],@X[3]);
1052 &vmovdqa(&QWP(0,"esp"),@X[0]); # X[]+K xfer to IALU
1054 &vmovdqa(&QWP(0+16,"esp"),@X[1]);
1056 &vmovdqa(&QWP(0+32,"esp"),@X[2]);
1058 &jmp (&label("loop"));
1060 sub Xupdate_avx_16_31() # recall that $Xi starts wtih 4
1063 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
1064 my ($a,$b,$c,$d,$e);
1066 eval(shift(@insns));
1067 eval(shift(@insns));
1068 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1069 eval(shift(@insns));
1070 eval(shift(@insns));
1072 &vpaddd (@X[3],@X[3],@X[-1&7]);
1073 &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
1074 eval(shift(@insns));
1075 eval(shift(@insns));
1076 &vpsrldq(@X[2],@X[-1&7],4); # "X[-3]", 3 dwords
1077 eval(shift(@insns));
1078 eval(shift(@insns));
1079 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1080 eval(shift(@insns));
1081 eval(shift(@insns));
1083 &vpxor (@X[2],@X[2],@X[-2&7]); # "X[-3]"^"X[-8]"
1084 eval(shift(@insns));
1085 eval(shift(@insns));
1086 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
1087 eval(shift(@insns));
1088 eval(shift(@insns));
1090 &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]"
1091 eval(shift(@insns));
1092 eval(shift(@insns));
1093 eval(shift(@insns));
1094 eval(shift(@insns));
1096 &vpsrld (@X[2],@X[0],31);
1097 eval(shift(@insns));
1098 eval(shift(@insns));
1099 eval(shift(@insns));
1100 eval(shift(@insns));
1102 &vpslldq(@X[4],@X[0],12); # "X[0]"<<96, extract one dword
1103 &vpaddd (@X[0],@X[0],@X[0]);
1104 eval(shift(@insns));
1105 eval(shift(@insns));
1106 eval(shift(@insns));
1107 eval(shift(@insns));
1109 &vpsrld (@X[3],@X[4],30);
1110 &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=1
1111 eval(shift(@insns));
1112 eval(shift(@insns));
1113 eval(shift(@insns));
1114 eval(shift(@insns));
1116 &vpslld (@X[4],@X[4],2);
1117 &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer
1118 eval(shift(@insns));
1119 eval(shift(@insns));
1120 &vpxor (@X[0],@X[0],@X[3]);
1121 eval(shift(@insns));
1122 eval(shift(@insns));
1123 eval(shift(@insns));
1124 eval(shift(@insns));
1126 &vpxor (@X[0],@X[0],@X[4]); # "X[0]"^=("X[0]"<<96)<<<2
1127 eval(shift(@insns));
1128 eval(shift(@insns));
1129 &vmovdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX
1130 eval(shift(@insns));
1131 eval(shift(@insns));
1133 foreach (@insns) { eval; } # remaining instructions [if any]
1135 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1138 sub Xupdate_avx_32_79()
1141 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
1142 my ($a,$b,$c,$d,$e);
1144 &vpalignr(@X[2],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1145 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1146 eval(shift(@insns)); # body_20_39
1147 eval(shift(@insns));
1148 eval(shift(@insns));
1149 eval(shift(@insns)); # rol
1151 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1152 &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer
1153 eval(shift(@insns));
1154 eval(shift(@insns));
1156 &vmovdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX...
1157 } else { # ... or load next one
1158 &vmovdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp"));
1160 &vpaddd (@X[3],@X[3],@X[-1&7]);
1161 eval(shift(@insns)); # ror
1162 eval(shift(@insns));
1164 &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-6]"
1165 eval(shift(@insns)); # body_20_39
1166 eval(shift(@insns));
1167 eval(shift(@insns));
1168 eval(shift(@insns)); # rol
1170 &vpsrld (@X[2],@X[0],30);
1171 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
1172 eval(shift(@insns));
1173 eval(shift(@insns));
1174 eval(shift(@insns)); # ror
1175 eval(shift(@insns));
1177 &vpslld (@X[0],@X[0],2);
1178 eval(shift(@insns)); # body_20_39
1179 eval(shift(@insns));
1180 eval(shift(@insns));
1181 eval(shift(@insns)); # rol
1182 eval(shift(@insns));
1183 eval(shift(@insns));
1184 eval(shift(@insns)); # ror
1185 eval(shift(@insns));
1187 &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=2
1188 eval(shift(@insns)); # body_20_39
1189 eval(shift(@insns));
1190 &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer
1191 eval(shift(@insns));
1192 eval(shift(@insns)); # rol
1193 eval(shift(@insns));
1194 eval(shift(@insns));
1195 eval(shift(@insns)); # ror
1196 eval(shift(@insns));
1198 foreach (@insns) { eval; } # remaining instructions
1200 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1203 sub Xuplast_avx_80()
1206 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1207 my ($a,$b,$c,$d,$e);
1209 eval(shift(@insns));
1210 &vpaddd (@X[3],@X[3],@X[-1&7]);
1211 eval(shift(@insns));
1212 eval(shift(@insns));
1213 eval(shift(@insns));
1214 eval(shift(@insns));
1216 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU
1218 foreach (@insns) { eval; } # remaining instructions
1220 &mov ($inp=@T[1],&DWP(192+4,"esp"));
1221 &cmp ($inp,&DWP(192+8,"esp"));
1222 &je (&label("done"));
1224 &vmovdqa(@X[3],&QWP(112+48,"esp")); # K_00_19
1225 &vmovdqa(@X[2],&QWP(112+64,"esp")); # pbswap mask
1226 &vmovdqu(@X[-4&7],&QWP(0,$inp)); # load input
1227 &vmovdqu(@X[-3&7],&QWP(16,$inp));
1228 &vmovdqu(@X[-2&7],&QWP(32,$inp));
1229 &vmovdqu(@X[-1&7],&QWP(48,$inp));
1231 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1232 &mov (&DWP(192+4,"esp"),$inp);
1233 &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
1241 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1242 my ($a,$b,$c,$d,$e);
1244 eval(shift(@insns));
1245 eval(shift(@insns));
1246 &vpshufb (@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1247 eval(shift(@insns));
1248 eval(shift(@insns));
1249 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],@X[3]);
1250 eval(shift(@insns));
1251 eval(shift(@insns));
1252 eval(shift(@insns));
1253 eval(shift(@insns));
1254 &vmovdqa (&QWP(0+16*$Xi,"esp"),@X[$Xi&7]); # X[]+K xfer to IALU
1255 eval(shift(@insns));
1256 eval(shift(@insns));
1258 foreach (@insns) { eval; }
1265 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1266 my ($a,$b,$c,$d,$e);
1268 foreach (@insns) { eval; }
1271 &set_label("loop",16);
1272 &Xupdate_avx_16_31(\&body_00_19);
1273 &Xupdate_avx_16_31(\&body_00_19);
1274 &Xupdate_avx_16_31(\&body_00_19);
1275 &Xupdate_avx_16_31(\&body_00_19);
1276 &Xupdate_avx_32_79(\&body_00_19);
1277 &Xupdate_avx_32_79(\&body_20_39);
1278 &Xupdate_avx_32_79(\&body_20_39);
1279 &Xupdate_avx_32_79(\&body_20_39);
1280 &Xupdate_avx_32_79(\&body_20_39);
1281 &Xupdate_avx_32_79(\&body_20_39);
1282 &Xupdate_avx_32_79(\&body_40_59);
1283 &Xupdate_avx_32_79(\&body_40_59);
1284 &Xupdate_avx_32_79(\&body_40_59);
1285 &Xupdate_avx_32_79(\&body_40_59);
1286 &Xupdate_avx_32_79(\&body_40_59);
1287 &Xupdate_avx_32_79(\&body_20_39);
1288 &Xuplast_avx_80(\&body_20_39); # can jump to "done"
1290 $saved_j=$j; @saved_V=@V;
1292 &Xloop_avx(\&body_20_39);
1293 &Xloop_avx(\&body_20_39);
1294 &Xloop_avx(\&body_20_39);
1296 &mov (@T[1],&DWP(192,"esp")); # update context
1297 &add ($A,&DWP(0,@T[1]));
1298 &add (@T[0],&DWP(4,@T[1])); # $b
1299 &add ($C,&DWP(8,@T[1]));
1300 &mov (&DWP(0,@T[1]),$A);
1301 &add ($D,&DWP(12,@T[1]));
1302 &mov (&DWP(4,@T[1]),@T[0]);
1303 &add ($E,&DWP(16,@T[1]));
1305 &mov (&DWP(8,@T[1]),$C);
1307 &mov (&DWP(12,@T[1]),$D);
1308 &mov (&DWP(16,@T[1]),$E);
1313 &jmp (&label("loop"));
1315 &set_label("done",16); $j=$saved_j; @V=@saved_V;
1317 &Xtail_avx(\&body_20_39);
1318 &Xtail_avx(\&body_20_39);
1319 &Xtail_avx(\&body_20_39);
1323 &mov (@T[1],&DWP(192,"esp")); # update context
1324 &add ($A,&DWP(0,@T[1]));
1325 &mov ("esp",&DWP(192+12,"esp")); # restore %esp
1326 &add (@T[0],&DWP(4,@T[1])); # $b
1327 &add ($C,&DWP(8,@T[1]));
1328 &mov (&DWP(0,@T[1]),$A);
1329 &add ($D,&DWP(12,@T[1]));
1330 &mov (&DWP(4,@T[1]),@T[0]);
1331 &add ($E,&DWP(16,@T[1]));
1332 &mov (&DWP(8,@T[1]),$C);
1333 &mov (&DWP(12,@T[1]),$D);
1334 &mov (&DWP(16,@T[1]),$E);
1335 &function_end("_sha1_block_data_order_avx");
1337 &set_label("K_XX_XX",64);
1338 &data_word(0x5a827999,0x5a827999,0x5a827999,0x5a827999); # K_00_19
1339 &data_word(0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1); # K_20_39
1340 &data_word(0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc); # K_40_59
1341 &data_word(0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6); # K_60_79
1342 &data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f); # pbswap mask
1344 &asciz("SHA1 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");