]> git.ipfire.org Git - thirdparty/openssl.git/blob - crypto/sha/asm/sha1-586.pl
projects / thirdparty / openssl.git / blob
? search:


59e115557a2790d09e1d47993d3665319289f4e7
[thirdparty/openssl.git] / crypto / sha / asm / sha1-586.pl
1 #! /usr/bin/env perl
2 # Copyright 1998-2016 The OpenSSL Project Authors. All Rights Reserved.
3 #
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
8
9
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 # ====================================================================
16
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.
21
22 # January, September 2004.
23 #
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:
30 #
31 # compared with original compared with Intel cc
32 # assembler impl. generated code
33 # Pentium -16% +48%
34 # PIII/AMD +8% +16%
35 # P4 +85%(!) +45%
36 #
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>
42
43 # August 2009.
44 #
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.
53
54 # October 2010.
55 #
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:
62 #
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/.
69
70 # April 2011.
71 #
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...
88
89 # March 2014.
90 #
91 # Add support for Intel SHA Extensions.
92
93 ######################################################################
94 # Current performance is summarized in following table. Numbers are
95 # CPU clock cycles spent to process single byte (less is better).
96 #
97 # x86 SSSE3 AVX
98 # Pentium 15.7 -
99 # PIII 11.5 -
100 # P4 10.6 -
101 # AMD K8 7.1 -
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 # Skylake 6.4 4.1/+55% 4.1(**)/+55%
108 # Bulldozer 11.6 6.0/+92%
109 # VIA Nano 10.6 7.5/+41%
110 # Atom 12.5 9.3(*)/+35%
111 # Silvermont 14.5 9.9(*)/+46%
112 # Goldmont 8.8 6.7/+30% 1.7(***)/+415%
113 #
114 # (*) Loop is 1056 instructions long and expected result is ~8.25.
115 # The discrepancy is because of front-end limitations, so
116 # called MS-ROM penalties, and on Silvermont even rotate's
117 # limited parallelism.
118 #
119 # (**) As per above comment, the result is for AVX *plus* sh[rl]d.
120 #
121 # (***) SHAEXT result
122
123 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
124 push(@INC,"${dir}","${dir}../../perlasm");
125 require "x86asm.pl";
126
127 $output=pop;
128 open STDOUT,">$output";
129
130 &asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386");
131
132 $xmm=$ymm=0;
133 for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
134
135 $ymm=1 if ($xmm &&
136 `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
137 =~ /GNU assembler version ([2-9]\.[0-9]+)/ &&
138 $1>=2.19); # first version supporting AVX
139
140 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32n" &&
141 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ &&
142 $1>=2.03); # first version supporting AVX
143
144 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32" &&
145 `ml 2>&1` =~ /Version ([0-9]+)\./ &&
146 $1>=10); # first version supporting AVX
147
148 $ymm=1 if ($xmm && !$ymm && `$ENV{CC} -v 2>&1` =~ /(^clang version|based on LLVM) ([3-9]\.[0-9]+)/ &&
149 $2>=3.0); # first version supporting AVX
150
151 $shaext=$xmm; ### set to zero if compiling for 1.0.1
152
153 &external_label("OPENSSL_ia32cap_P") if ($xmm);
154
155
156 $A="eax";
157 $B="ebx";
158 $C="ecx";
159 $D="edx";
160 $E="edi";
161 $T="esi";
162 $tmp1="ebp";
163
164 @V=($A,$B,$C,$D,$E,$T);
165
166 $alt=0; # 1 denotes alternative IALU implementation, which performs
167 # 8% *worse* on P4, same on Westmere and Atom, 2% better on
168 # Sandy Bridge...
169
170 sub BODY_00_15
171 {
172 local($n,$a,$b,$c,$d,$e,$f)=@_;
173
174 &comment("00_15 $n");
175
176 &mov($f,$c); # f to hold F_00_19(b,c,d)
177 if ($n==0) { &mov($tmp1,$a); }
178 else { &mov($a,$tmp1); }
179 &rotl($tmp1,5); # tmp1=ROTATE(a,5)
180 &xor($f,$d);
181 &add($tmp1,$e); # tmp1+=e;
182 &mov($e,&swtmp($n%16)); # e becomes volatile and is loaded
183 # with xi, also note that e becomes
184 # f in next round...
185 &and($f,$b);
186 &rotr($b,2); # b=ROTATE(b,30)
187 &xor($f,$d); # f holds F_00_19(b,c,d)
188 &lea($tmp1,&DWP(0x5a827999,$tmp1,$e)); # tmp1+=K_00_19+xi
189
190 if ($n==15) { &mov($e,&swtmp(($n+1)%16));# pre-fetch f for next round
191 &add($f,$tmp1); } # f+=tmp1
192 else { &add($tmp1,$f); } # f becomes a in next round
193 &mov($tmp1,$a) if ($alt && $n==15);
194 }
195
196 sub BODY_16_19
197 {
198 local($n,$a,$b,$c,$d,$e,$f)=@_;
199
200 &comment("16_19 $n");
201
202 if ($alt) {
203 &xor($c,$d);
204 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
205 &and($tmp1,$c); # tmp1 to hold F_00_19(b,c,d), b&=c^d
206 &xor($f,&swtmp(($n+8)%16));
207 &xor($tmp1,$d); # tmp1=F_00_19(b,c,d)
208 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
209 &rotl($f,1); # f=ROTATE(f,1)
210 &add($e,$tmp1); # e+=F_00_19(b,c,d)
211 &xor($c,$d); # restore $c
212 &mov($tmp1,$a); # b in next round
213 &rotr($b,$n==16?2:7); # b=ROTATE(b,30)
214 &mov(&swtmp($n%16),$f); # xi=f
215 &rotl($a,5); # ROTATE(a,5)
216 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
217 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
218 &add($f,$a); # f+=ROTATE(a,5)
219 } else {
220 &mov($tmp1,$c); # tmp1 to hold F_00_19(b,c,d)
221 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
222 &xor($tmp1,$d);
223 &xor($f,&swtmp(($n+8)%16));
224 &and($tmp1,$b);
225 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
226 &rotl($f,1); # f=ROTATE(f,1)
227 &xor($tmp1,$d); # tmp1=F_00_19(b,c,d)
228 &add($e,$tmp1); # e+=F_00_19(b,c,d)
229 &mov($tmp1,$a);
230 &rotr($b,2); # b=ROTATE(b,30)
231 &mov(&swtmp($n%16),$f); # xi=f
232 &rotl($tmp1,5); # ROTATE(a,5)
233 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
234 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
235 &add($f,$tmp1); # f+=ROTATE(a,5)
236 }
237 }
238
239 sub BODY_20_39
240 {
241 local($n,$a,$b,$c,$d,$e,$f)=@_;
242 local $K=($n<40)?0x6ed9eba1:0xca62c1d6;
243
244 &comment("20_39 $n");
245
246 if ($alt) {
247 &xor($tmp1,$c); # tmp1 to hold F_20_39(b,c,d), b^=c
248 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
249 &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d)
250 &xor($f,&swtmp(($n+8)%16));
251 &add($e,$tmp1); # e+=F_20_39(b,c,d)
252 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
253 &rotl($f,1); # f=ROTATE(f,1)
254 &mov($tmp1,$a); # b in next round
255 &rotr($b,7); # b=ROTATE(b,30)
256 &mov(&swtmp($n%16),$f) if($n<77);# xi=f
257 &rotl($a,5); # ROTATE(a,5)
258 &xor($b,$c) if($n==39);# warm up for BODY_40_59
259 &and($tmp1,$b) if($n==39);
260 &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY
261 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
262 &add($f,$a); # f+=ROTATE(a,5)
263 &rotr($a,5) if ($n==79);
264 } else {
265 &mov($tmp1,$b); # tmp1 to hold F_20_39(b,c,d)
266 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
267 &xor($tmp1,$c);
268 &xor($f,&swtmp(($n+8)%16));
269 &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d)
270 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
271 &rotl($f,1); # f=ROTATE(f,1)
272 &add($e,$tmp1); # e+=F_20_39(b,c,d)
273 &rotr($b,2); # b=ROTATE(b,30)
274 &mov($tmp1,$a);
275 &rotl($tmp1,5); # ROTATE(a,5)
276 &mov(&swtmp($n%16),$f) if($n<77);# xi=f
277 &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY
278 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
279 &add($f,$tmp1); # f+=ROTATE(a,5)
280 }
281 }
282
283 sub BODY_40_59
284 {
285 local($n,$a,$b,$c,$d,$e,$f)=@_;
286
287 &comment("40_59 $n");
288
289 if ($alt) {
290 &add($e,$tmp1); # e+=b&(c^d)
291 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
292 &mov($tmp1,$d);
293 &xor($f,&swtmp(($n+8)%16));
294 &xor($c,$d); # restore $c
295 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
296 &rotl($f,1); # f=ROTATE(f,1)
297 &and($tmp1,$c);
298 &rotr($b,7); # b=ROTATE(b,30)
299 &add($e,$tmp1); # e+=c&d
300 &mov($tmp1,$a); # b in next round
301 &mov(&swtmp($n%16),$f); # xi=f
302 &rotl($a,5); # ROTATE(a,5)
303 &xor($b,$c) if ($n<59);
304 &and($tmp1,$b) if ($n<59);# tmp1 to hold F_40_59(b,c,d)
305 &lea($f,&DWP(0x8f1bbcdc,$f,$e));# f+=K_40_59+e+(b&(c^d))
306 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
307 &add($f,$a); # f+=ROTATE(a,5)
308 } else {
309 &mov($tmp1,$c); # tmp1 to hold F_40_59(b,c,d)
310 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
311 &xor($tmp1,$d);
312 &xor($f,&swtmp(($n+8)%16));
313 &and($tmp1,$b);
314 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
315 &rotl($f,1); # f=ROTATE(f,1)
316 &add($tmp1,$e); # b&(c^d)+=e
317 &rotr($b,2); # b=ROTATE(b,30)
318 &mov($e,$a); # e becomes volatile
319 &rotl($e,5); # ROTATE(a,5)
320 &mov(&swtmp($n%16),$f); # xi=f
321 &lea($f,&DWP(0x8f1bbcdc,$f,$tmp1));# f+=K_40_59+e+(b&(c^d))
322 &mov($tmp1,$c);
323 &add($f,$e); # f+=ROTATE(a,5)
324 &and($tmp1,$d);
325 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
326 &add($f,$tmp1); # f+=c&d
327 }
328 }
329
330 &function_begin("sha1_block_data_order");
331 if ($xmm) {
332 &static_label("shaext_shortcut") if ($shaext);
333 &static_label("ssse3_shortcut");
334 &static_label("avx_shortcut") if ($ymm);
335 &static_label("K_XX_XX");
336
337 &call (&label("pic_point")); # make it PIC!
338 &set_label("pic_point");
339 &blindpop($tmp1);
340 &picmeup($T,"OPENSSL_ia32cap_P",$tmp1,&label("pic_point"));
341 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
342
343 &mov ($A,&DWP(0,$T));
344 &mov ($D,&DWP(4,$T));
345 &test ($D,1<<9); # check SSSE3 bit
346 &jz (&label("x86"));
347 &mov ($C,&DWP(8,$T));
348 &test ($A,1<<24); # check FXSR bit
349 &jz (&label("x86"));
350 if ($shaext) {
351 &test ($C,1<<29); # check SHA bit
352 &jnz (&label("shaext_shortcut"));
353 }
354 if ($ymm) {
355 &and ($D,1<<28); # mask AVX bit
356 &and ($A,1<<30); # mask "Intel CPU" bit
357 &or ($A,$D);
358 &cmp ($A,1<<28|1<<30);
359 &je (&label("avx_shortcut"));
360 }
361 &jmp (&label("ssse3_shortcut"));
362 &set_label("x86",16);
363 }
364 &mov($tmp1,&wparam(0)); # SHA_CTX *c
365 &mov($T,&wparam(1)); # const void *input
366 &mov($A,&wparam(2)); # size_t num
367 &stack_push(16+3); # allocate X[16]
368 &shl($A,6);
369 &add($A,$T);
370 &mov(&wparam(2),$A); # pointer beyond the end of input
371 &mov($E,&DWP(16,$tmp1));# pre-load E
372 &jmp(&label("loop"));
373
374 &set_label("loop",16);
375
376 # copy input chunk to X, but reversing byte order!
377 for ($i=0; $i<16; $i+=4)
378 {
379 &mov($A,&DWP(4*($i+0),$T));
380 &mov($B,&DWP(4*($i+1),$T));
381 &mov($C,&DWP(4*($i+2),$T));
382 &mov($D,&DWP(4*($i+3),$T));
383 &bswap($A);
384 &bswap($B);
385 &bswap($C);
386 &bswap($D);
387 &mov(&swtmp($i+0),$A);
388 &mov(&swtmp($i+1),$B);
389 &mov(&swtmp($i+2),$C);
390 &mov(&swtmp($i+3),$D);
391 }
392 &mov(&wparam(1),$T); # redundant in 1st spin
393
394 &mov($A,&DWP(0,$tmp1)); # load SHA_CTX
395 &mov($B,&DWP(4,$tmp1));
396 &mov($C,&DWP(8,$tmp1));
397 &mov($D,&DWP(12,$tmp1));
398 # E is pre-loaded
399
400 for($i=0;$i<16;$i++) { &BODY_00_15($i,@V); unshift(@V,pop(@V)); }
401 for(;$i<20;$i++) { &BODY_16_19($i,@V); unshift(@V,pop(@V)); }
402 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
403 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
404 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
405
406 (($V[5] eq $D) and ($V[0] eq $E)) or die; # double-check
407
408 &mov($tmp1,&wparam(0)); # re-load SHA_CTX*
409 &mov($D,&wparam(1)); # D is last "T" and is discarded
410
411 &add($E,&DWP(0,$tmp1)); # E is last "A"...
412 &add($T,&DWP(4,$tmp1));
413 &add($A,&DWP(8,$tmp1));
414 &add($B,&DWP(12,$tmp1));
415 &add($C,&DWP(16,$tmp1));
416
417 &mov(&DWP(0,$tmp1),$E); # update SHA_CTX
418 &add($D,64); # advance input pointer
419 &mov(&DWP(4,$tmp1),$T);
420 &cmp($D,&wparam(2)); # have we reached the end yet?
421 &mov(&DWP(8,$tmp1),$A);
422 &mov($E,$C); # C is last "E" which needs to be "pre-loaded"
423 &mov(&DWP(12,$tmp1),$B);
424 &mov($T,$D); # input pointer
425 &mov(&DWP(16,$tmp1),$C);
426 &jb(&label("loop"));
427
428 &stack_pop(16+3);
429 &function_end("sha1_block_data_order");
430
431 if ($xmm) {
432 if ($shaext) {
433 ######################################################################
434 # Intel SHA Extensions implementation of SHA1 update function.
435 #
436 my ($ctx,$inp,$num)=("edi","esi","ecx");
437 my ($ABCD,$E,$E_,$BSWAP)=map("xmm$_",(0..3));
438 my @MSG=map("xmm$_",(4..7));
439
440 sub sha1rnds4 {
441 my ($dst,$src,$imm)=@_;
442 if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
443 { &data_byte(0x0f,0x3a,0xcc,0xc0|($1<<3)|$2,$imm); }
444 }
445 sub sha1op38 {
446 my ($opcodelet,$dst,$src)=@_;
447 if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
448 { &data_byte(0x0f,0x38,$opcodelet,0xc0|($1<<3)|$2); }
449 }
450 sub sha1nexte { sha1op38(0xc8,@_); }
451 sub sha1msg1 { sha1op38(0xc9,@_); }
452 sub sha1msg2 { sha1op38(0xca,@_); }
453
454 &function_begin("_sha1_block_data_order_shaext");
455 &call (&label("pic_point")); # make it PIC!
456 &set_label("pic_point");
457 &blindpop($tmp1);
458 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
459 &set_label("shaext_shortcut");
460 &mov ($ctx,&wparam(0));
461 &mov ("ebx","esp");
462 &mov ($inp,&wparam(1));
463 &mov ($num,&wparam(2));
464 &sub ("esp",32);
465
466 &movdqu ($ABCD,&QWP(0,$ctx));
467 &movd ($E,&DWP(16,$ctx));
468 &and ("esp",-32);
469 &movdqa ($BSWAP,&QWP(0x50,$tmp1)); # byte-n-word swap
470
471 &movdqu (@MSG[0],&QWP(0,$inp));
472 &pshufd ($ABCD,$ABCD,0b00011011); # flip word order
473 &movdqu (@MSG[1],&QWP(0x10,$inp));
474 &pshufd ($E,$E,0b00011011); # flip word order
475 &movdqu (@MSG[2],&QWP(0x20,$inp));
476 &pshufb (@MSG[0],$BSWAP);
477 &movdqu (@MSG[3],&QWP(0x30,$inp));
478 &pshufb (@MSG[1],$BSWAP);
479 &pshufb (@MSG[2],$BSWAP);
480 &pshufb (@MSG[3],$BSWAP);
481 &jmp (&label("loop_shaext"));
482
483 &set_label("loop_shaext",16);
484 &dec ($num);
485 &lea ("eax",&DWP(0x40,$inp));
486 &movdqa (&QWP(0,"esp"),$E); # offload $E
487 &paddd ($E,@MSG[0]);
488 &cmovne ($inp,"eax");
489 &movdqa (&QWP(16,"esp"),$ABCD); # offload $ABCD
490
491 for($i=0;$i<20-4;$i+=2) {
492 &sha1msg1 (@MSG[0],@MSG[1]);
493 &movdqa ($E_,$ABCD);
494 &sha1rnds4 ($ABCD,$E,int($i/5)); # 0-3...
495 &sha1nexte ($E_,@MSG[1]);
496 &pxor (@MSG[0],@MSG[2]);
497 &sha1msg1 (@MSG[1],@MSG[2]);
498 &sha1msg2 (@MSG[0],@MSG[3]);
499
500 &movdqa ($E,$ABCD);
501 &sha1rnds4 ($ABCD,$E_,int(($i+1)/5));
502 &sha1nexte ($E,@MSG[2]);
503 &pxor (@MSG[1],@MSG[3]);
504 &sha1msg2 (@MSG[1],@MSG[0]);
505
506 push(@MSG,shift(@MSG)); push(@MSG,shift(@MSG));
507 }
508 &movdqu (@MSG[0],&QWP(0,$inp));
509 &movdqa ($E_,$ABCD);
510 &sha1rnds4 ($ABCD,$E,3); # 64-67
511 &sha1nexte ($E_,@MSG[1]);
512 &movdqu (@MSG[1],&QWP(0x10,$inp));
513 &pshufb (@MSG[0],$BSWAP);
514
515 &movdqa ($E,$ABCD);
516 &sha1rnds4 ($ABCD,$E_,3); # 68-71
517 &sha1nexte ($E,@MSG[2]);
518 &movdqu (@MSG[2],&QWP(0x20,$inp));
519 &pshufb (@MSG[1],$BSWAP);
520
521 &movdqa ($E_,$ABCD);
522 &sha1rnds4 ($ABCD,$E,3); # 72-75
523 &sha1nexte ($E_,@MSG[3]);
524 &movdqu (@MSG[3],&QWP(0x30,$inp));
525 &pshufb (@MSG[2],$BSWAP);
526
527 &movdqa ($E,$ABCD);
528 &sha1rnds4 ($ABCD,$E_,3); # 76-79
529 &movdqa ($E_,&QWP(0,"esp"));
530 &pshufb (@MSG[3],$BSWAP);
531 &sha1nexte ($E,$E_);
532 &paddd ($ABCD,&QWP(16,"esp"));
533
534 &jnz (&label("loop_shaext"));
535
536 &pshufd ($ABCD,$ABCD,0b00011011);
537 &pshufd ($E,$E,0b00011011);
538 &movdqu (&QWP(0,$ctx),$ABCD)
539 &movd (&DWP(16,$ctx),$E);
540 &mov ("esp","ebx");
541 &function_end("_sha1_block_data_order_shaext");
542 }
543 ######################################################################
544 # The SSSE3 implementation.
545 #
546 # %xmm[0-7] are used as ring @X[] buffer containing quadruples of last
547 # 32 elements of the message schedule or Xupdate outputs. First 4
548 # quadruples are simply byte-swapped input, next 4 are calculated
549 # according to method originally suggested by Dean Gaudet (modulo
550 # being implemented in SSSE3). Once 8 quadruples or 32 elements are
551 # collected, it switches to routine proposed by Max Locktyukhin.
552 #
553 # Calculations inevitably require temporary reqisters, and there are
554 # no %xmm registers left to spare. For this reason part of the ring
555 # buffer, X[2..4] to be specific, is offloaded to 3 quadriples ring
556 # buffer on the stack. Keep in mind that X[2] is alias X[-6], X[3] -
557 # X[-5], and X[4] - X[-4]...
558 #
559 # Another notable optimization is aggressive stack frame compression
560 # aiming to minimize amount of 9-byte instructions...
561 #
562 # Yet another notable optimization is "jumping" $B variable. It means
563 # that there is no register permanently allocated for $B value. This
564 # allowed to eliminate one instruction from body_20_39...
565 #
566 my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded
567 my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4
568 my @V=($A,$B,$C,$D,$E);
569 my $j=0; # hash round
570 my $rx=0;
571 my @T=($T,$tmp1);
572 my $inp;
573
574 my $_rol=sub { &rol(@_) };
575 my $_ror=sub { &ror(@_) };
576
577 &function_begin("_sha1_block_data_order_ssse3");
578 &call (&label("pic_point")); # make it PIC!
579 &set_label("pic_point");
580 &blindpop($tmp1);
581 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
582 &set_label("ssse3_shortcut");
583
584 &movdqa (@X[3],&QWP(0,$tmp1)); # K_00_19
585 &movdqa (@X[4],&QWP(16,$tmp1)); # K_20_39
586 &movdqa (@X[5],&QWP(32,$tmp1)); # K_40_59
587 &movdqa (@X[6],&QWP(48,$tmp1)); # K_60_79
588 &movdqa (@X[2],&QWP(64,$tmp1)); # pbswap mask
589
590 &mov ($E,&wparam(0)); # load argument block
591 &mov ($inp=@T[1],&wparam(1));
592 &mov ($D,&wparam(2));
593 &mov (@T[0],"esp");
594
595 # stack frame layout
596 #
597 # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area
598 # X[4]+K X[5]+K X[6]+K X[7]+K
599 # X[8]+K X[9]+K X[10]+K X[11]+K
600 # X[12]+K X[13]+K X[14]+K X[15]+K
601 #
602 # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area
603 # X[4] X[5] X[6] X[7]
604 # X[8] X[9] X[10] X[11] # even borrowed for K_00_19
605 #
606 # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants
607 # K_40_59 K_40_59 K_40_59 K_40_59
608 # K_60_79 K_60_79 K_60_79 K_60_79
609 # K_00_19 K_00_19 K_00_19 K_00_19
610 # pbswap mask
611 #
612 # +192 ctx # argument block
613 # +196 inp
614 # +200 end
615 # +204 esp
616 &sub ("esp",208);
617 &and ("esp",-64);
618
619 &movdqa (&QWP(112+0,"esp"),@X[4]); # copy constants
620 &movdqa (&QWP(112+16,"esp"),@X[5]);
621 &movdqa (&QWP(112+32,"esp"),@X[6]);
622 &shl ($D,6); # len*64
623 &movdqa (&QWP(112+48,"esp"),@X[3]);
624 &add ($D,$inp); # end of input
625 &movdqa (&QWP(112+64,"esp"),@X[2]);
626 &add ($inp,64);
627 &mov (&DWP(192+0,"esp"),$E); # save argument block
628 &mov (&DWP(192+4,"esp"),$inp);
629 &mov (&DWP(192+8,"esp"),$D);
630 &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp
631
632 &mov ($A,&DWP(0,$E)); # load context
633 &mov ($B,&DWP(4,$E));
634 &mov ($C,&DWP(8,$E));
635 &mov ($D,&DWP(12,$E));
636 &mov ($E,&DWP(16,$E));
637 &mov (@T[0],$B); # magic seed
638
639 &movdqu (@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3]
640 &movdqu (@X[-3&7],&QWP(-48,$inp));
641 &movdqu (@X[-2&7],&QWP(-32,$inp));
642 &movdqu (@X[-1&7],&QWP(-16,$inp));
643 &pshufb (@X[-4&7],@X[2]); # byte swap
644 &pshufb (@X[-3&7],@X[2]);
645 &pshufb (@X[-2&7],@X[2]);
646 &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
647 &pshufb (@X[-1&7],@X[2]);
648 &paddd (@X[-4&7],@X[3]); # add K_00_19
649 &paddd (@X[-3&7],@X[3]);
650 &paddd (@X[-2&7],@X[3]);
651 &movdqa (&QWP(0,"esp"),@X[-4&7]); # X[]+K xfer to IALU
652 &psubd (@X[-4&7],@X[3]); # restore X[]
653 &movdqa (&QWP(0+16,"esp"),@X[-3&7]);
654 &psubd (@X[-3&7],@X[3]);
655 &movdqa (&QWP(0+32,"esp"),@X[-2&7]);
656 &mov (@T[1],$C);
657 &psubd (@X[-2&7],@X[3]);
658 &xor (@T[1],$D);
659 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
660 &and (@T[0],@T[1]);
661 &jmp (&label("loop"));
662
663 ######################################################################
664 # SSE instruction sequence is first broken to groups of independent
665 # instructions, independent in respect to their inputs and shifter
666 # (not all architectures have more than one). Then IALU instructions
667 # are "knitted in" between the SSE groups. Distance is maintained for
668 # SSE latency of 2 in hope that it fits better upcoming AMD Bulldozer
669 # [which allegedly also implements SSSE3]...
670 #
671 # Temporary registers usage. X[2] is volatile at the entry and at the
672 # end is restored from backtrace ring buffer. X[3] is expected to
673 # contain current K_XX_XX constant and is used to calculate X[-1]+K
674 # from previous round, it becomes volatile the moment the value is
675 # saved to stack for transfer to IALU. X[4] becomes volatile whenever
676 # X[-4] is accumulated and offloaded to backtrace ring buffer, at the
677 # end it is loaded with next K_XX_XX [which becomes X[3] in next
678 # round]...
679 #
680 sub Xupdate_ssse3_16_31() # recall that $Xi starts with 4
681 { use integer;
682 my $body = shift;
683 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
684 my ($a,$b,$c,$d,$e);
685
686 eval(shift(@insns)); # ror
687 eval(shift(@insns));
688 eval(shift(@insns));
689 &punpcklqdq(@X[0],@X[-3&7]); # compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8);
690 &movdqa (@X[2],@X[-1&7]);
691 eval(shift(@insns));
692 eval(shift(@insns));
693
694 &paddd (@X[3],@X[-1&7]);
695 &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
696 eval(shift(@insns)); # rol
697 eval(shift(@insns));
698 &psrldq (@X[2],4); # "X[-3]", 3 dwords
699 eval(shift(@insns));
700 eval(shift(@insns));
701 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
702 eval(shift(@insns));
703 eval(shift(@insns)); # ror
704
705 &pxor (@X[2],@X[-2&7]); # "X[-3]"^"X[-8]"
706 eval(shift(@insns));
707 eval(shift(@insns));
708 eval(shift(@insns));
709
710 &pxor (@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]"
711 eval(shift(@insns));
712 eval(shift(@insns)); # rol
713 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
714 eval(shift(@insns));
715 eval(shift(@insns));
716
717 &movdqa (@X[4],@X[0]);
718 eval(shift(@insns));
719 eval(shift(@insns));
720 eval(shift(@insns)); # ror
721 &movdqa (@X[2],@X[0]);
722 eval(shift(@insns));
723
724 &pslldq (@X[4],12); # "X[0]"<<96, extract one dword
725 &paddd (@X[0],@X[0]);
726 eval(shift(@insns));
727 eval(shift(@insns));
728
729 &psrld (@X[2],31);
730 eval(shift(@insns));
731 eval(shift(@insns)); # rol
732 &movdqa (@X[3],@X[4]);
733 eval(shift(@insns));
734 eval(shift(@insns));
735 eval(shift(@insns));
736
737 &psrld (@X[4],30);
738 eval(shift(@insns));
739 eval(shift(@insns)); # ror
740 &por (@X[0],@X[2]); # "X[0]"<<<=1
741 eval(shift(@insns));
742 &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer
743 eval(shift(@insns));
744 eval(shift(@insns));
745
746 &pslld (@X[3],2);
747 eval(shift(@insns));
748 eval(shift(@insns)); # rol
749 &pxor (@X[0],@X[4]);
750 &movdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX
751 eval(shift(@insns));
752 eval(shift(@insns));
753
754 &pxor (@X[0],@X[3]); # "X[0]"^=("X[0]"<<96)<<<2
755 &pshufd (@X[1],@X[-3&7],0xee) if ($Xi<7); # was &movdqa (@X[1],@X[-2&7])
756 &pshufd (@X[3],@X[-1&7],0xee) if ($Xi==7);
757 eval(shift(@insns));
758 eval(shift(@insns));
759
760 foreach (@insns) { eval; } # remaining instructions [if any]
761
762 $Xi++; push(@X,shift(@X)); # "rotate" X[]
763 }
764
765 sub Xupdate_ssse3_32_79()
766 { use integer;
767 my $body = shift;
768 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
769 my ($a,$b,$c,$d,$e);
770
771 eval(shift(@insns)); # body_20_39
772 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
773 &punpcklqdq(@X[2],@X[-1&7]); # compose "X[-6]", was &palignr(@X[2],@X[-2&7],8)
774 eval(shift(@insns));
775 eval(shift(@insns));
776 eval(shift(@insns)); # rol
777
778 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
779 &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer
780 eval(shift(@insns));
781 eval(shift(@insns));
782 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
783 if ($Xi%5) {
784 &movdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX...
785 } else { # ... or load next one
786 &movdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp"));
787 }
788 eval(shift(@insns)); # ror
789 &paddd (@X[3],@X[-1&7]);
790 eval(shift(@insns));
791
792 &pxor (@X[0],@X[2]); # "X[0]"^="X[-6]"
793 eval(shift(@insns)); # body_20_39
794 eval(shift(@insns));
795 eval(shift(@insns));
796 eval(shift(@insns)); # rol
797
798 &movdqa (@X[2],@X[0]);
799 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
800 eval(shift(@insns));
801 eval(shift(@insns));
802 eval(shift(@insns)); # ror
803 eval(shift(@insns));
804 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
805
806 &pslld (@X[0],2);
807 eval(shift(@insns)); # body_20_39
808 eval(shift(@insns));
809 &psrld (@X[2],30);
810 eval(shift(@insns));
811 eval(shift(@insns)); # rol
812 eval(shift(@insns));
813 eval(shift(@insns));
814 eval(shift(@insns)); # ror
815 eval(shift(@insns));
816 eval(shift(@insns)) if (@insns[1] =~ /_rol/);
817 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
818
819 &por (@X[0],@X[2]); # "X[0]"<<<=2
820 eval(shift(@insns)); # body_20_39
821 eval(shift(@insns));
822 &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer
823 eval(shift(@insns));
824 eval(shift(@insns)); # rol
825 eval(shift(@insns));
826 eval(shift(@insns));
827 eval(shift(@insns)); # ror
828 &pshufd (@X[3],@X[-1],0xee) if ($Xi<19); # was &movdqa (@X[3],@X[0])
829 eval(shift(@insns));
830
831 foreach (@insns) { eval; } # remaining instructions
832
833 $Xi++; push(@X,shift(@X)); # "rotate" X[]
834 }
835
836 sub Xuplast_ssse3_80()
837 { use integer;
838 my $body = shift;
839 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
840 my ($a,$b,$c,$d,$e);
841
842 eval(shift(@insns));
843 eval(shift(@insns));
844 eval(shift(@insns));
845 eval(shift(@insns));
846 eval(shift(@insns));
847 eval(shift(@insns));
848 eval(shift(@insns));
849 &paddd (@X[3],@X[-1&7]);
850 eval(shift(@insns));
851 eval(shift(@insns));
852 eval(shift(@insns));
853 eval(shift(@insns));
854
855 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU
856
857 foreach (@insns) { eval; } # remaining instructions
858
859 &mov ($inp=@T[1],&DWP(192+4,"esp"));
860 &cmp ($inp,&DWP(192+8,"esp"));
861 &je (&label("done"));
862
863 &movdqa (@X[3],&QWP(112+48,"esp")); # K_00_19
864 &movdqa (@X[2],&QWP(112+64,"esp")); # pbswap mask
865 &movdqu (@X[-4&7],&QWP(0,$inp)); # load input
866 &movdqu (@X[-3&7],&QWP(16,$inp));
867 &movdqu (@X[-2&7],&QWP(32,$inp));
868 &movdqu (@X[-1&7],&QWP(48,$inp));
869 &add ($inp,64);
870 &pshufb (@X[-4&7],@X[2]); # byte swap
871 &mov (&DWP(192+4,"esp"),$inp);
872 &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
873
874 $Xi=0;
875 }
876
877 sub Xloop_ssse3()
878 { use integer;
879 my $body = shift;
880 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
881 my ($a,$b,$c,$d,$e);
882
883 eval(shift(@insns));
884 eval(shift(@insns));
885 eval(shift(@insns));
886 eval(shift(@insns));
887 eval(shift(@insns));
888 eval(shift(@insns));
889 eval(shift(@insns));
890 &pshufb (@X[($Xi-3)&7],@X[2]);
891 eval(shift(@insns));
892 eval(shift(@insns));
893 eval(shift(@insns));
894 eval(shift(@insns));
895 &paddd (@X[($Xi-4)&7],@X[3]);
896 eval(shift(@insns));
897 eval(shift(@insns));
898 eval(shift(@insns));
899 eval(shift(@insns));
900 &movdqa (&QWP(0+16*$Xi,"esp"),@X[($Xi-4)&7]); # X[]+K xfer to IALU
901 eval(shift(@insns));
902 eval(shift(@insns));
903 eval(shift(@insns));
904 eval(shift(@insns));
905 &psubd (@X[($Xi-4)&7],@X[3]);
906
907 foreach (@insns) { eval; }
908 $Xi++;
909 }
910
911 sub Xtail_ssse3()
912 { use integer;
913 my $body = shift;
914 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
915 my ($a,$b,$c,$d,$e);
916
917 foreach (@insns) { eval; }
918 }
919
920 sub body_00_19 () { # ((c^d)&b)^d
921 # on start @T[0]=(c^d)&b
922 return &body_20_39() if ($rx==19); $rx++;
923 (
924 '($a,$b,$c,$d,$e)=@V;'.
925 '&$_ror ($b,$j?7:2);', # $b>>>2
926 '&xor (@T[0],$d);',
927 '&mov (@T[1],$a);', # $b in next round
928
929 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer
930 '&xor ($b,$c);', # $c^$d for next round
931
932 '&$_rol ($a,5);',
933 '&add ($e,@T[0]);',
934 '&and (@T[1],$b);', # ($b&($c^$d)) for next round
935
936 '&xor ($b,$c);', # restore $b
937 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
938 );
939 }
940
941 sub body_20_39 () { # b^d^c
942 # on entry @T[0]=b^d
943 return &body_40_59() if ($rx==39); $rx++;
944 (
945 '($a,$b,$c,$d,$e)=@V;'.
946 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer
947 '&xor (@T[0],$d) if($j==19);'.
948 '&xor (@T[0],$c) if($j> 19);', # ($b^$d^$c)
949 '&mov (@T[1],$a);', # $b in next round
950
951 '&$_rol ($a,5);',
952 '&add ($e,@T[0]);',
953 '&xor (@T[1],$c) if ($j< 79);', # $b^$d for next round
954
955 '&$_ror ($b,7);', # $b>>>2
956 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
957 );
958 }
959
960 sub body_40_59 () { # ((b^c)&(c^d))^c
961 # on entry @T[0]=(b^c), (c^=d)
962 $rx++;
963 (
964 '($a,$b,$c,$d,$e)=@V;'.
965 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer
966 '&and (@T[0],$c) if ($j>=40);', # (b^c)&(c^d)
967 '&xor ($c,$d) if ($j>=40);', # restore $c
968
969 '&$_ror ($b,7);', # $b>>>2
970 '&mov (@T[1],$a);', # $b for next round
971 '&xor (@T[0],$c);',
972
973 '&$_rol ($a,5);',
974 '&add ($e,@T[0]);',
975 '&xor (@T[1],$c) if ($j==59);'.
976 '&xor (@T[1],$b) if ($j< 59);', # b^c for next round
977
978 '&xor ($b,$c) if ($j< 59);', # c^d for next round
979 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
980 );
981 }
982 ######
983 sub bodyx_00_19 () { # ((c^d)&b)^d
984 # on start @T[0]=(b&c)^(~b&d), $e+=X[]+K
985 return &bodyx_20_39() if ($rx==19); $rx++;
986 (
987 '($a,$b,$c,$d,$e)=@V;'.
988
989 '&rorx ($b,$b,2) if ($j==0);'. # $b>>>2
990 '&rorx ($b,@T[1],7) if ($j!=0);', # $b>>>2
991 '&lea ($e,&DWP(0,$e,@T[0]));',
992 '&rorx (@T[0],$a,5);',
993
994 '&andn (@T[1],$a,$c);',
995 '&and ($a,$b)',
996 '&add ($d,&DWP(4*(($j+1)&15),"esp"));', # X[]+K xfer
997
998 '&xor (@T[1],$a)',
999 '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
1000 );
1001 }
1002
1003 sub bodyx_20_39 () { # b^d^c
1004 # on start $b=b^c^d
1005 return &bodyx_40_59() if ($rx==39); $rx++;
1006 (
1007 '($a,$b,$c,$d,$e)=@V;'.
1008
1009 '&add ($e,($j==19?@T[0]:$b))',
1010 '&rorx ($b,@T[1],7);', # $b>>>2
1011 '&rorx (@T[0],$a,5);',
1012
1013 '&xor ($a,$b) if ($j<79);',
1014 '&add ($d,&DWP(4*(($j+1)&15),"esp")) if ($j<79);', # X[]+K xfer
1015 '&xor ($a,$c) if ($j<79);',
1016 '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
1017 );
1018 }
1019
1020 sub bodyx_40_59 () { # ((b^c)&(c^d))^c
1021 # on start $b=((b^c)&(c^d))^c
1022 return &bodyx_20_39() if ($rx==59); $rx++;
1023 (
1024 '($a,$b,$c,$d,$e)=@V;'.
1025
1026 '&rorx (@T[0],$a,5)',
1027 '&lea ($e,&DWP(0,$e,$b))',
1028 '&rorx ($b,@T[1],7)', # $b>>>2
1029 '&add ($d,&DWP(4*(($j+1)&15),"esp"))', # X[]+K xfer
1030
1031 '&mov (@T[1],$c)',
1032 '&xor ($a,$b)', # b^c for next round
1033 '&xor (@T[1],$b)', # c^d for next round
1034
1035 '&and ($a,@T[1])',
1036 '&add ($e,@T[0])',
1037 '&xor ($a,$b)' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
1038 );
1039 }
1040
1041 &set_label("loop",16);
1042 &Xupdate_ssse3_16_31(\&body_00_19);
1043 &Xupdate_ssse3_16_31(\&body_00_19);
1044 &Xupdate_ssse3_16_31(\&body_00_19);
1045 &Xupdate_ssse3_16_31(\&body_00_19);
1046 &Xupdate_ssse3_32_79(\&body_00_19);
1047 &Xupdate_ssse3_32_79(\&body_20_39);
1048 &Xupdate_ssse3_32_79(\&body_20_39);
1049 &Xupdate_ssse3_32_79(\&body_20_39);
1050 &Xupdate_ssse3_32_79(\&body_20_39);
1051 &Xupdate_ssse3_32_79(\&body_20_39);
1052 &Xupdate_ssse3_32_79(\&body_40_59);
1053 &Xupdate_ssse3_32_79(\&body_40_59);
1054 &Xupdate_ssse3_32_79(\&body_40_59);
1055 &Xupdate_ssse3_32_79(\&body_40_59);
1056 &Xupdate_ssse3_32_79(\&body_40_59);
1057 &Xupdate_ssse3_32_79(\&body_20_39);
1058 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done"
1059
1060 $saved_j=$j; @saved_V=@V;
1061
1062 &Xloop_ssse3(\&body_20_39);
1063 &Xloop_ssse3(\&body_20_39);
1064 &Xloop_ssse3(\&body_20_39);
1065
1066 &mov (@T[1],&DWP(192,"esp")); # update context
1067 &add ($A,&DWP(0,@T[1]));
1068 &add (@T[0],&DWP(4,@T[1])); # $b
1069 &add ($C,&DWP(8,@T[1]));
1070 &mov (&DWP(0,@T[1]),$A);
1071 &add ($D,&DWP(12,@T[1]));
1072 &mov (&DWP(4,@T[1]),@T[0]);
1073 &add ($E,&DWP(16,@T[1]));
1074 &mov (&DWP(8,@T[1]),$C);
1075 &mov ($B,$C);
1076 &mov (&DWP(12,@T[1]),$D);
1077 &xor ($B,$D);
1078 &mov (&DWP(16,@T[1]),$E);
1079 &mov (@T[1],@T[0]);
1080 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
1081 &and (@T[0],$B);
1082 &mov ($B,$T[1]);
1083
1084 &jmp (&label("loop"));
1085
1086 &set_label("done",16); $j=$saved_j; @V=@saved_V;
1087
1088 &Xtail_ssse3(\&body_20_39);
1089 &Xtail_ssse3(\&body_20_39);
1090 &Xtail_ssse3(\&body_20_39);
1091
1092 &mov (@T[1],&DWP(192,"esp")); # update context
1093 &add ($A,&DWP(0,@T[1]));
1094 &mov ("esp",&DWP(192+12,"esp")); # restore %esp
1095 &add (@T[0],&DWP(4,@T[1])); # $b
1096 &add ($C,&DWP(8,@T[1]));
1097 &mov (&DWP(0,@T[1]),$A);
1098 &add ($D,&DWP(12,@T[1]));
1099 &mov (&DWP(4,@T[1]),@T[0]);
1100 &add ($E,&DWP(16,@T[1]));
1101 &mov (&DWP(8,@T[1]),$C);
1102 &mov (&DWP(12,@T[1]),$D);
1103 &mov (&DWP(16,@T[1]),$E);
1104
1105 &function_end("_sha1_block_data_order_ssse3");
1106
1107 $rx=0; # reset
1108
1109 if ($ymm) {
1110 my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded
1111 my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4
1112 my @V=($A,$B,$C,$D,$E);
1113 my $j=0; # hash round
1114 my @T=($T,$tmp1);
1115 my $inp;
1116
1117 my $_rol=sub { &shld(@_[0],@_) };
1118 my $_ror=sub { &shrd(@_[0],@_) };
1119
1120 &function_begin("_sha1_block_data_order_avx");
1121 &call (&label("pic_point")); # make it PIC!
1122 &set_label("pic_point");
1123 &blindpop($tmp1);
1124 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
1125 &set_label("avx_shortcut");
1126 &vzeroall();
1127
1128 &vmovdqa(@X[3],&QWP(0,$tmp1)); # K_00_19
1129 &vmovdqa(@X[4],&QWP(16,$tmp1)); # K_20_39
1130 &vmovdqa(@X[5],&QWP(32,$tmp1)); # K_40_59
1131 &vmovdqa(@X[6],&QWP(48,$tmp1)); # K_60_79
1132 &vmovdqa(@X[2],&QWP(64,$tmp1)); # pbswap mask
1133
1134 &mov ($E,&wparam(0)); # load argument block
1135 &mov ($inp=@T[1],&wparam(1));
1136 &mov ($D,&wparam(2));
1137 &mov (@T[0],"esp");
1138
1139 # stack frame layout
1140 #
1141 # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area
1142 # X[4]+K X[5]+K X[6]+K X[7]+K
1143 # X[8]+K X[9]+K X[10]+K X[11]+K
1144 # X[12]+K X[13]+K X[14]+K X[15]+K
1145 #
1146 # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area
1147 # X[4] X[5] X[6] X[7]
1148 # X[8] X[9] X[10] X[11] # even borrowed for K_00_19
1149 #
1150 # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants
1151 # K_40_59 K_40_59 K_40_59 K_40_59
1152 # K_60_79 K_60_79 K_60_79 K_60_79
1153 # K_00_19 K_00_19 K_00_19 K_00_19
1154 # pbswap mask
1155 #
1156 # +192 ctx # argument block
1157 # +196 inp
1158 # +200 end
1159 # +204 esp
1160 &sub ("esp",208);
1161 &and ("esp",-64);
1162
1163 &vmovdqa(&QWP(112+0,"esp"),@X[4]); # copy constants
1164 &vmovdqa(&QWP(112+16,"esp"),@X[5]);
1165 &vmovdqa(&QWP(112+32,"esp"),@X[6]);
1166 &shl ($D,6); # len*64
1167 &vmovdqa(&QWP(112+48,"esp"),@X[3]);
1168 &add ($D,$inp); # end of input
1169 &vmovdqa(&QWP(112+64,"esp"),@X[2]);
1170 &add ($inp,64);
1171 &mov (&DWP(192+0,"esp"),$E); # save argument block
1172 &mov (&DWP(192+4,"esp"),$inp);
1173 &mov (&DWP(192+8,"esp"),$D);
1174 &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp
1175
1176 &mov ($A,&DWP(0,$E)); # load context
1177 &mov ($B,&DWP(4,$E));
1178 &mov ($C,&DWP(8,$E));
1179 &mov ($D,&DWP(12,$E));
1180 &mov ($E,&DWP(16,$E));
1181 &mov (@T[0],$B); # magic seed
1182
1183 &vmovdqu(@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3]
1184 &vmovdqu(@X[-3&7],&QWP(-48,$inp));
1185 &vmovdqu(@X[-2&7],&QWP(-32,$inp));
1186 &vmovdqu(@X[-1&7],&QWP(-16,$inp));
1187 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1188 &vpshufb(@X[-3&7],@X[-3&7],@X[2]);
1189 &vpshufb(@X[-2&7],@X[-2&7],@X[2]);
1190 &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
1191 &vpshufb(@X[-1&7],@X[-1&7],@X[2]);
1192 &vpaddd (@X[0],@X[-4&7],@X[3]); # add K_00_19
1193 &vpaddd (@X[1],@X[-3&7],@X[3]);
1194 &vpaddd (@X[2],@X[-2&7],@X[3]);
1195 &vmovdqa(&QWP(0,"esp"),@X[0]); # X[]+K xfer to IALU
1196 &mov (@T[1],$C);
1197 &vmovdqa(&QWP(0+16,"esp"),@X[1]);
1198 &xor (@T[1],$D);
1199 &vmovdqa(&QWP(0+32,"esp"),@X[2]);
1200 &and (@T[0],@T[1]);
1201 &jmp (&label("loop"));
1202
1203 sub Xupdate_avx_16_31() # recall that $Xi starts with 4
1204 { use integer;
1205 my $body = shift;
1206 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
1207 my ($a,$b,$c,$d,$e);
1208
1209 eval(shift(@insns));
1210 eval(shift(@insns));
1211 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1212 eval(shift(@insns));
1213 eval(shift(@insns));
1214
1215 &vpaddd (@X[3],@X[3],@X[-1&7]);
1216 &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
1217 eval(shift(@insns));
1218 eval(shift(@insns));
1219 &vpsrldq(@X[2],@X[-1&7],4); # "X[-3]", 3 dwords
1220 eval(shift(@insns));
1221 eval(shift(@insns));
1222 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1223 eval(shift(@insns));
1224 eval(shift(@insns));
1225
1226 &vpxor (@X[2],@X[2],@X[-2&7]); # "X[-3]"^"X[-8]"
1227 eval(shift(@insns));
1228 eval(shift(@insns));
1229 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
1230 eval(shift(@insns));
1231 eval(shift(@insns));
1232
1233 &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]"
1234 eval(shift(@insns));
1235 eval(shift(@insns));
1236 eval(shift(@insns));
1237 eval(shift(@insns));
1238
1239 &vpsrld (@X[2],@X[0],31);
1240 eval(shift(@insns));
1241 eval(shift(@insns));
1242 eval(shift(@insns));
1243 eval(shift(@insns));
1244
1245 &vpslldq(@X[4],@X[0],12); # "X[0]"<<96, extract one dword
1246 &vpaddd (@X[0],@X[0],@X[0]);
1247 eval(shift(@insns));
1248 eval(shift(@insns));
1249 eval(shift(@insns));
1250 eval(shift(@insns));
1251
1252 &vpsrld (@X[3],@X[4],30);
1253 &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=1
1254 eval(shift(@insns));
1255 eval(shift(@insns));
1256 eval(shift(@insns));
1257 eval(shift(@insns));
1258
1259 &vpslld (@X[4],@X[4],2);
1260 &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer
1261 eval(shift(@insns));
1262 eval(shift(@insns));
1263 &vpxor (@X[0],@X[0],@X[3]);
1264 eval(shift(@insns));
1265 eval(shift(@insns));
1266 eval(shift(@insns));
1267 eval(shift(@insns));
1268
1269 &vpxor (@X[0],@X[0],@X[4]); # "X[0]"^=("X[0]"<<96)<<<2
1270 eval(shift(@insns));
1271 eval(shift(@insns));
1272 &vmovdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX
1273 eval(shift(@insns));
1274 eval(shift(@insns));
1275
1276 foreach (@insns) { eval; } # remaining instructions [if any]
1277
1278 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1279 }
1280
1281 sub Xupdate_avx_32_79()
1282 { use integer;
1283 my $body = shift;
1284 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
1285 my ($a,$b,$c,$d,$e);
1286
1287 &vpalignr(@X[2],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1288 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1289 eval(shift(@insns)); # body_20_39
1290 eval(shift(@insns));
1291 eval(shift(@insns));
1292 eval(shift(@insns)); # rol
1293
1294 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1295 &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer
1296 eval(shift(@insns));
1297 eval(shift(@insns));
1298 if ($Xi%5) {
1299 &vmovdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX...
1300 } else { # ... or load next one
1301 &vmovdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp"));
1302 }
1303 &vpaddd (@X[3],@X[3],@X[-1&7]);
1304 eval(shift(@insns)); # ror
1305 eval(shift(@insns));
1306
1307 &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-6]"
1308 eval(shift(@insns)); # body_20_39
1309 eval(shift(@insns));
1310 eval(shift(@insns));
1311 eval(shift(@insns)); # rol
1312
1313 &vpsrld (@X[2],@X[0],30);
1314 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
1315 eval(shift(@insns));
1316 eval(shift(@insns));
1317 eval(shift(@insns)); # ror
1318 eval(shift(@insns));
1319
1320 &vpslld (@X[0],@X[0],2);
1321 eval(shift(@insns)); # body_20_39
1322 eval(shift(@insns));
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));
1329
1330 &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=2
1331 eval(shift(@insns)); # body_20_39
1332 eval(shift(@insns));
1333 &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer
1334 eval(shift(@insns));
1335 eval(shift(@insns)); # rol
1336 eval(shift(@insns));
1337 eval(shift(@insns));
1338 eval(shift(@insns)); # ror
1339 eval(shift(@insns));
1340
1341 foreach (@insns) { eval; } # remaining instructions
1342
1343 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1344 }
1345
1346 sub Xuplast_avx_80()
1347 { use integer;
1348 my $body = shift;
1349 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1350 my ($a,$b,$c,$d,$e);
1351
1352 eval(shift(@insns));
1353 &vpaddd (@X[3],@X[3],@X[-1&7]);
1354 eval(shift(@insns));
1355 eval(shift(@insns));
1356 eval(shift(@insns));
1357 eval(shift(@insns));
1358
1359 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU
1360
1361 foreach (@insns) { eval; } # remaining instructions
1362
1363 &mov ($inp=@T[1],&DWP(192+4,"esp"));
1364 &cmp ($inp,&DWP(192+8,"esp"));
1365 &je (&label("done"));
1366
1367 &vmovdqa(@X[3],&QWP(112+48,"esp")); # K_00_19
1368 &vmovdqa(@X[2],&QWP(112+64,"esp")); # pbswap mask
1369 &vmovdqu(@X[-4&7],&QWP(0,$inp)); # load input
1370 &vmovdqu(@X[-3&7],&QWP(16,$inp));
1371 &vmovdqu(@X[-2&7],&QWP(32,$inp));
1372 &vmovdqu(@X[-1&7],&QWP(48,$inp));
1373 &add ($inp,64);
1374 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1375 &mov (&DWP(192+4,"esp"),$inp);
1376 &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
1377
1378 $Xi=0;
1379 }
1380
1381 sub Xloop_avx()
1382 { use integer;
1383 my $body = shift;
1384 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1385 my ($a,$b,$c,$d,$e);
1386
1387 eval(shift(@insns));
1388 eval(shift(@insns));
1389 &vpshufb (@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1390 eval(shift(@insns));
1391 eval(shift(@insns));
1392 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],@X[3]);
1393 eval(shift(@insns));
1394 eval(shift(@insns));
1395 eval(shift(@insns));
1396 eval(shift(@insns));
1397 &vmovdqa (&QWP(0+16*$Xi,"esp"),@X[$Xi&7]); # X[]+K xfer to IALU
1398 eval(shift(@insns));
1399 eval(shift(@insns));
1400
1401 foreach (@insns) { eval; }
1402 $Xi++;
1403 }
1404
1405 sub Xtail_avx()
1406 { use integer;
1407 my $body = shift;
1408 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1409 my ($a,$b,$c,$d,$e);
1410
1411 foreach (@insns) { eval; }
1412 }
1413
1414 &set_label("loop",16);
1415 &Xupdate_avx_16_31(\&body_00_19);
1416 &Xupdate_avx_16_31(\&body_00_19);
1417 &Xupdate_avx_16_31(\&body_00_19);
1418 &Xupdate_avx_16_31(\&body_00_19);
1419 &Xupdate_avx_32_79(\&body_00_19);
1420 &Xupdate_avx_32_79(\&body_20_39);
1421 &Xupdate_avx_32_79(\&body_20_39);
1422 &Xupdate_avx_32_79(\&body_20_39);
1423 &Xupdate_avx_32_79(\&body_20_39);
1424 &Xupdate_avx_32_79(\&body_20_39);
1425 &Xupdate_avx_32_79(\&body_40_59);
1426 &Xupdate_avx_32_79(\&body_40_59);
1427 &Xupdate_avx_32_79(\&body_40_59);
1428 &Xupdate_avx_32_79(\&body_40_59);
1429 &Xupdate_avx_32_79(\&body_40_59);
1430 &Xupdate_avx_32_79(\&body_20_39);
1431 &Xuplast_avx_80(\&body_20_39); # can jump to "done"
1432
1433 $saved_j=$j; @saved_V=@V;
1434
1435 &Xloop_avx(\&body_20_39);
1436 &Xloop_avx(\&body_20_39);
1437 &Xloop_avx(\&body_20_39);
1438
1439 &mov (@T[1],&DWP(192,"esp")); # update context
1440 &add ($A,&DWP(0,@T[1]));
1441 &add (@T[0],&DWP(4,@T[1])); # $b
1442 &add ($C,&DWP(8,@T[1]));
1443 &mov (&DWP(0,@T[1]),$A);
1444 &add ($D,&DWP(12,@T[1]));
1445 &mov (&DWP(4,@T[1]),@T[0]);
1446 &add ($E,&DWP(16,@T[1]));
1447 &mov ($B,$C);
1448 &mov (&DWP(8,@T[1]),$C);
1449 &xor ($B,$D);
1450 &mov (&DWP(12,@T[1]),$D);
1451 &mov (&DWP(16,@T[1]),$E);
1452 &mov (@T[1],@T[0]);
1453 &and (@T[0],$B);
1454 &mov ($B,@T[1]);
1455
1456 &jmp (&label("loop"));
1457
1458 &set_label("done",16); $j=$saved_j; @V=@saved_V;
1459
1460 &Xtail_avx(\&body_20_39);
1461 &Xtail_avx(\&body_20_39);
1462 &Xtail_avx(\&body_20_39);
1463
1464 &vzeroall();
1465
1466 &mov (@T[1],&DWP(192,"esp")); # update context
1467 &add ($A,&DWP(0,@T[1]));
1468 &mov ("esp",&DWP(192+12,"esp")); # restore %esp
1469 &add (@T[0],&DWP(4,@T[1])); # $b
1470 &add ($C,&DWP(8,@T[1]));
1471 &mov (&DWP(0,@T[1]),$A);
1472 &add ($D,&DWP(12,@T[1]));
1473 &mov (&DWP(4,@T[1]),@T[0]);
1474 &add ($E,&DWP(16,@T[1]));
1475 &mov (&DWP(8,@T[1]),$C);
1476 &mov (&DWP(12,@T[1]),$D);
1477 &mov (&DWP(16,@T[1]),$E);
1478 &function_end("_sha1_block_data_order_avx");
1479 }
1480 &set_label("K_XX_XX",64);
1481 &data_word(0x5a827999,0x5a827999,0x5a827999,0x5a827999); # K_00_19
1482 &data_word(0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1); # K_20_39
1483 &data_word(0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc); # K_40_59
1484 &data_word(0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6); # K_60_79
1485 &data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f); # pbswap mask
1486 &data_byte(0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0);
1487 }
1488 &asciz("SHA1 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
1489
1490 &asm_finish();
1491
1492 close STDOUT;
A mirror of the official openssl repository