447c50b0ccf6dac98bcdb93ef430d3d3d3a01a89
[openssl.git] / crypto / sha / asm / sha256-586.pl
1 #!/usr/bin/env perl
2 #
3 # ====================================================================
4 # 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 # ====================================================================
9 #
10 # SHA256 block transform for x86. September 2007.
11 #
12 # Performance improvement over compiler generated code varies from
13 # 10% to 40% [see below]. Not very impressive on some ยต-archs, but
14 # it's 5 times smaller and optimizies amount of writes.
15 #
16 # May 2012.
17 #
18 # Optimization including two of Pavel Semjanov's ideas, alternative
19 # Maj and full unroll, resulted in ~20-25% improvement on most CPUs,
20 # ~7% on Pentium, ~40% on Atom. As fully unrolled loop body is almost
21 # 15x larger, 8KB vs. 560B, it's fired only for longer inputs. But not
22 # on P4, where it kills performance, nor Sandy Bridge, where folded
23 # loop is approximately as fast...
24 #
25 # June 2012.
26 #
27 # Add AMD XOP-specific code path, >30% improvement on Bulldozer over
28 # May version, >60% over original. Add AVX+shrd code path, >25%
29 # improvement on Sandy Bridge over May version, 60% over original.
30 #
31 # Performance in clock cycles per processed byte (less is better):
32 #
33 #               PIII    P4      AMD K8  Core2   SB      Atom    Bldzr
34 # gcc           36      41      27      26      25      50      36
35 # icc           33      38      25      23      -       -       -
36 # x86 asm(*)    27/24   28      19/15.5 18/15.6 12.3    30/25   16.6
37 # x86_64 asm(**)        17.5    15.1    13.9    11.6    22      13.7
38 #
39 # (*)   numbers after slash are for unrolled loop, where available,
40 #       otherwise best applicable such as AVX/XOP;
41 # (**)  x86_64 assembly performance is presented for reference
42 #       purposes.
43
44 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
45 push(@INC,"${dir}","${dir}../../perlasm");
46 require "x86asm.pl";
47
48 &asm_init($ARGV[0],"sha512-586.pl",$ARGV[$#ARGV] eq "386");
49
50 $xmm=$ymm=0;
51 for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
52
53 $ymm=1 if ($xmm &&
54                 `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
55                         =~ /GNU assembler version ([2-9]\.[0-9]+)/ &&
56                 $1>=2.19);      # first version supporting AVX
57
58 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32n" &&
59                 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ &&
60                 $1>=2.03);      # first version supporting AVX
61
62 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32" &&
63                 `ml 2>&1` =~ /Version ([0-9]+)\./ &&
64                 $1>=10);        # first version supporting AVX
65
66 $unroll_after = 64*4;   # If pre-evicted from L1P cache first spin of
67                         # fully unrolled loop was measured to run about
68                         # 3-4x slower. If slowdown coefficient is N and
69                         # unrolled loop is m times faster, then you break
70                         # even at (N-1)/(m-1) blocks. Then it needs to be
71                         # adjusted for probability of code being evicted,
72                         # code size/cache size=1/4. Typical m is 1.15...
73
74 $A="eax";
75 $E="edx";
76 $T="ebx";
77 $Aoff=&DWP(4,"esp");
78 $Boff=&DWP(8,"esp");
79 $Coff=&DWP(12,"esp");
80 $Doff=&DWP(16,"esp");
81 $Eoff=&DWP(20,"esp");
82 $Foff=&DWP(24,"esp");
83 $Goff=&DWP(28,"esp");
84 $Hoff=&DWP(32,"esp");
85 $Xoff=&DWP(36,"esp");
86 $K256="ebp";
87
88 sub BODY_16_63() {
89         &mov    ($T,"ecx");                     # "ecx" is preloaded
90          &mov   ("esi",&DWP(4*(9+15+16-14),"esp"));
91         &ror    ("ecx",18-7);
92          &mov   ("edi","esi");
93         &ror    ("esi",19-17);
94          &xor   ("ecx",$T);
95          &shr   ($T,3);
96         &ror    ("ecx",7);
97          &xor   ("esi","edi");
98          &xor   ($T,"ecx");                     # T = sigma0(X[-15])
99         &ror    ("esi",17);
100          &add   ($T,&DWP(4*(9+15+16),"esp"));   # T += X[-16]
101         &shr    ("edi",10);
102          &add   ($T,&DWP(4*(9+15+16-9),"esp")); # T += X[-7]
103         #&xor   ("edi","esi")                   # sigma1(X[-2])
104         # &add  ($T,"edi");                     # T += sigma1(X[-2])
105         # &mov  (&DWP(4*(9+15),"esp"),$T);      # save X[0]
106
107         &BODY_00_15(1);
108 }
109 sub BODY_00_15() {
110     my $in_16_63=shift;
111
112         &mov    ("ecx",$E);
113          &xor   ("edi","esi")                   if ($in_16_63); # sigma1(X[-2])
114          &mov   ("esi",$Foff);
115         &ror    ("ecx",25-11);
116          &add   ($T,"edi")                      if ($in_16_63); # T += sigma1(X[-2])
117          &mov   ("edi",$Goff);
118         &xor    ("ecx",$E);
119          &xor   ("esi","edi");
120          &mov   ($T,&DWP(4*(9+15),"esp"))       if (!$in_16_63);
121          &mov   (&DWP(4*(9+15),"esp"),$T)       if ($in_16_63); # save X[0]
122         &ror    ("ecx",11-6);
123          &and   ("esi",$E);
124          &mov   ($Eoff,$E);             # modulo-scheduled
125         &xor    ($E,"ecx");
126          &add   ($T,$Hoff);             # T += h
127          &xor   ("esi","edi");          # Ch(e,f,g)
128         &ror    ($E,6);                 # Sigma1(e)
129          &mov   ("ecx",$A);
130          &add   ($T,"esi");             # T += Ch(e,f,g)
131
132         &ror    ("ecx",22-13);
133          &add   ($T,$E);                # T += Sigma1(e)
134          &mov   ("edi",$Boff);
135         &xor    ("ecx",$A);
136          &mov   ($Aoff,$A);             # modulo-scheduled
137          &lea   ("esp",&DWP(-4,"esp"));
138         &ror    ("ecx",13-2);
139          &mov   ("esi",&DWP(0,$K256));
140         &xor    ("ecx",$A);
141          &mov   ($E,$Eoff);             # e in next iteration, d in this one
142          &xor   ($A,"edi");             # a ^= b
143         &ror    ("ecx",2);              # Sigma0(a)
144
145          &add   ($T,"esi");             # T+= K[i]
146          &mov   (&DWP(0,"esp"),$A);     # (b^c) in next round
147         &add    ($E,$T);                # d += T
148          &and   ($A,&DWP(4,"esp"));     # a &= (b^c)
149         &add    ($T,"ecx");             # T += Sigma0(a)
150          &xor   ($A,"edi");             # h = Maj(a,b,c) = Ch(a^b,c,b)
151          &mov   ("ecx",&DWP(4*(9+15+16-1),"esp"))       if ($in_16_63); # preload T
152         &add    ($K256,4);
153          &add   ($A,$T);                # h += T
154 }
155
156 &external_label("OPENSSL_ia32cap_P")            if (!$i386);
157
158 &function_begin("sha256_block_data_order");
159         &mov    ("esi",wparam(0));      # ctx
160         &mov    ("edi",wparam(1));      # inp
161         &mov    ("eax",wparam(2));      # num
162         &mov    ("ebx","esp");          # saved sp
163
164         &call   (&label("pic_point"));  # make it PIC!
165 &set_label("pic_point");
166         &blindpop($K256);
167         &lea    ($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256));
168
169         &sub    ("esp",16);
170         &and    ("esp",-64);
171
172         &shl    ("eax",6);
173         &add    ("eax","edi");
174         &mov    (&DWP(0,"esp"),"esi");  # ctx
175         &mov    (&DWP(4,"esp"),"edi");  # inp
176         &mov    (&DWP(8,"esp"),"eax");  # inp+num*128
177         &mov    (&DWP(12,"esp"),"ebx"); # saved sp
178                                                 if (!$i386) {
179         &picmeup("edx","OPENSSL_ia32cap_P",$K256,&label("K256"));
180         &mov    ("ecx",&DWP(0,"edx"));
181         &mov    ("edx",&DWP(4,"edx"));
182         &test   ("ecx",1<<20);          # check for P4
183         &jnz    (&label("loop"));
184         &test   ("edx",1<<11);          # check for XOP
185         &jnz    (&label("XOP"))         if ($ymm);
186         &and    ("ecx",1<<30);          # mask "Intel CPU" bit
187         &and    ("edx",1<<28);          # mask AVX bit
188         &or     ("ecx","edx");
189         &cmp    ("ecx",1<<28|1<<30);
190         &je     (&label("AVX"))         if ($ymm);
191         &je     (&label("loop_shrd"))   if (!$ymm);
192                                                 if ($unroll_after) {
193         &sub    ("eax","edi");
194         &cmp    ("eax",$unroll_after);
195         &jae    (&label("unrolled"));
196                                                 } }
197         &jmp    (&label("loop"));
198
199 sub COMPACT_LOOP() {
200 my $suffix=shift;
201
202 &set_label("loop$suffix",16);
203     # copy input block to stack reversing byte and dword order
204     for($i=0;$i<4;$i++) {
205         &mov    ("eax",&DWP($i*16+0,"edi"));
206         &mov    ("ebx",&DWP($i*16+4,"edi"));
207         &mov    ("ecx",&DWP($i*16+8,"edi"));
208         &bswap  ("eax");
209         &mov    ("edx",&DWP($i*16+12,"edi"));
210         &bswap  ("ebx");
211         &push   ("eax");
212         &bswap  ("ecx");
213         &push   ("ebx");
214         &bswap  ("edx");
215         &push   ("ecx");
216         &push   ("edx");
217     }
218         &add    ("edi",64);
219         &lea    ("esp",&DWP(-4*9,"esp"));# place for A,B,C,D,E,F,G,H
220         &mov    (&DWP(4*(9+16)+4,"esp"),"edi");
221
222         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
223         &mov    ($A,&DWP(0,"esi"));
224         &mov    ("ebx",&DWP(4,"esi"));
225         &mov    ("ecx",&DWP(8,"esi"));
226         &mov    ("edi",&DWP(12,"esi"));
227         # &mov  ($Aoff,$A);
228         &mov    ($Boff,"ebx");
229         &xor    ("ebx","ecx");
230         &mov    ($Coff,"ecx");
231         &mov    ($Doff,"edi");
232         &mov    (&DWP(0,"esp"),"ebx");  # magic
233         &mov    ($E,&DWP(16,"esi"));    
234         &mov    ("ebx",&DWP(20,"esi"));
235         &mov    ("ecx",&DWP(24,"esi"));
236         &mov    ("edi",&DWP(28,"esi"));
237         # &mov  ($Eoff,$E);
238         &mov    ($Foff,"ebx");
239         &mov    ($Goff,"ecx");
240         &mov    ($Hoff,"edi");
241
242 &set_label("00_15$suffix",16);
243
244         &BODY_00_15();
245
246         &cmp    ("esi",0xc19bf174);
247         &jne    (&label("00_15$suffix"));
248
249         &mov    ("ecx",&DWP(4*(9+15+16-1),"esp"));      # preloaded in BODY_00_15(1)
250         &jmp    (&label("16_63$suffix"));
251
252 &set_label("16_63$suffix",16);
253
254         &BODY_16_63();
255
256         &cmp    ("esi",0xc67178f2);
257         &jne    (&label("16_63$suffix"));
258
259         &mov    ("esi",&DWP(4*(9+16+64)+0,"esp"));#ctx
260         # &mov  ($A,$Aoff);
261         &mov    ("ebx",$Boff);
262         # &mov  ("edi",$Coff);
263         &mov    ("ecx",$Doff);
264         &add    ($A,&DWP(0,"esi"));
265         &add    ("ebx",&DWP(4,"esi"));
266         &add    ("edi",&DWP(8,"esi"));
267         &add    ("ecx",&DWP(12,"esi"));
268         &mov    (&DWP(0,"esi"),$A);
269         &mov    (&DWP(4,"esi"),"ebx");
270         &mov    (&DWP(8,"esi"),"edi");
271         &mov    (&DWP(12,"esi"),"ecx");
272         # &mov  ($E,$Eoff);
273         &mov    ("eax",$Foff);
274         &mov    ("ebx",$Goff);
275         &mov    ("ecx",$Hoff);
276         &mov    ("edi",&DWP(4*(9+16+64)+4,"esp"));#inp
277         &add    ($E,&DWP(16,"esi"));
278         &add    ("eax",&DWP(20,"esi"));
279         &add    ("ebx",&DWP(24,"esi"));
280         &add    ("ecx",&DWP(28,"esi"));
281         &mov    (&DWP(16,"esi"),$E);
282         &mov    (&DWP(20,"esi"),"eax");
283         &mov    (&DWP(24,"esi"),"ebx");
284         &mov    (&DWP(28,"esi"),"ecx");
285
286         &lea    ("esp",&DWP(4*(9+16+64),"esp"));# destroy frame
287         &sub    ($K256,4*64);                   # rewind K
288
289         &cmp    ("edi",&DWP(8,"esp"));          # are we done yet?
290         &jb     (&label("loop$suffix"));
291 }
292         &COMPACT_LOOP();
293         &mov    ("esp",&DWP(12,"esp"));         # restore sp
294 &function_end_A();
295                                                 if (!$i386 && !$ymm) {
296         # ~20% improvement on Sandy Bridge
297         local *ror = sub { &shrd(@_[0],@_) };
298         &COMPACT_LOOP("_shrd");
299         &mov    ("esp",&DWP(12,"esp"));         # restore sp
300 &function_end_A();
301                                                 }
302
303 &set_label("K256",64);  # Yes! I keep it in the code segment!
304 @K256=( 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5,
305         0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5,
306         0xd807aa98,0x12835b01,0x243185be,0x550c7dc3,
307         0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174,
308         0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc,
309         0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da,
310         0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7,
311         0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967,
312         0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13,
313         0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85,
314         0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3,
315         0xd192e819,0xd6990624,0xf40e3585,0x106aa070,
316         0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5,
317         0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3,
318         0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208,
319         0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2     );
320 &data_word(@K256);
321 &data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f);
322
323 if (!$i386 && $unroll_after) {
324 my @AH=($A,$K256);
325
326 &set_label("unrolled",16);
327         &lea    ("esp",&DWP(-96,"esp"));
328         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
329         &mov    ($AH[0],&DWP(0,"esi"));
330         &mov    ($AH[1],&DWP(4,"esi"));
331         &mov    ("ecx",&DWP(8,"esi"));
332         &mov    ("ebx",&DWP(12,"esi"));
333         #&mov   (&DWP(0,"esp"),$AH[0]);
334         &mov    (&DWP(4,"esp"),$AH[1]);
335         &xor    ($AH[1],"ecx");         # magic
336         &mov    (&DWP(8,"esp"),"ecx");
337         &mov    (&DWP(12,"esp"),"ebx");
338         &mov    ($E,&DWP(16,"esi"));    
339         &mov    ("ebx",&DWP(20,"esi"));
340         &mov    ("ecx",&DWP(24,"esi"));
341         &mov    ("esi",&DWP(28,"esi"));
342         #&mov   (&DWP(16,"esp"),$E);
343         &mov    (&DWP(20,"esp"),"ebx");
344         &mov    (&DWP(24,"esp"),"ecx");
345         &mov    (&DWP(28,"esp"),"esi");
346         &jmp    (&label("grand_loop"));
347
348 &set_label("grand_loop",16);
349     # copy input block to stack reversing byte order
350     for($i=0;$i<5;$i++) {
351         &mov    ("ebx",&DWP(12*$i+0,"edi"));
352         &mov    ("ecx",&DWP(12*$i+4,"edi"));
353         &bswap  ("ebx");
354         &mov    ("esi",&DWP(12*$i+8,"edi"));
355         &bswap  ("ecx");
356         &mov    (&DWP(32+12*$i+0,"esp"),"ebx");
357         &bswap  ("esi");
358         &mov    (&DWP(32+12*$i+4,"esp"),"ecx");
359         &mov    (&DWP(32+12*$i+8,"esp"),"esi");
360     }
361         &mov    ("ebx",&DWP($i*12,"edi"));
362         &add    ("edi",64);
363         &bswap  ("ebx");
364         &mov    (&DWP(96+4,"esp"),"edi");
365         &mov    (&DWP(32+12*$i,"esp"),"ebx");
366
367     my ($t1,$t2) = ("ecx","esi");
368     my ($a,$b,$c,$d,$e,$f,$g,$h)=(0..7);        # offsets
369     sub off { &DWP(4*(((shift)-$i)&7),"esp"); }
370
371     for ($i=0;$i<64;$i++) {
372
373       if ($i>=16) {
374         &mov    ($T,$t1);                       # $t1 is preloaded
375         # &mov  ($t2,&DWP(32+4*(($i+14)&15),"esp"));
376         &ror    ($t1,18-7);
377          &mov   ("edi",$t2);
378         &ror    ($t2,19-17);
379          &xor   ($t1,$T);
380          &shr   ($T,3);
381         &ror    ($t1,7);
382          &xor   ($t2,"edi");
383          &xor   ($T,$t1);                       # T = sigma0(X[-15])
384         &ror    ($t2,17);
385          &add   ($T,&DWP(32+4*($i&15),"esp"));  # T += X[-16]
386         &shr    ("edi",10);
387          &add   ($T,&DWP(32+4*(($i+9)&15),"esp"));      # T += X[-7]
388         #&xor   ("edi",$t2)                     # sigma1(X[-2])
389         # &add  ($T,"edi");                     # T += sigma1(X[-2])
390         # &mov  (&DWP(4*(9+15),"esp"),$T);      # save X[0]
391       }
392         &mov    ($t1,$E);
393          &xor   ("edi",$t2)                     if ($i>=16);    # sigma1(X[-2])
394          &mov   ($t2,&off($f));
395         &ror    ($E,25-11);
396          &add   ($T,"edi")                      if ($i>=16);    # T += sigma1(X[-2])
397          &mov   ("edi",&off($g));
398         &xor    ($E,$t1);
399          &mov   ($T,&DWP(32+4*($i&15),"esp"))   if ($i<16);     # X[i]
400          &mov   (&DWP(32+4*($i&15),"esp"),$T)   if ($i>=16 && $i<62);   # save X[0]
401          &xor   ($t2,"edi");
402         &ror    ($E,11-6);
403          &and   ($t2,$t1);
404          &mov   (&off($e),$t1);         # save $E, modulo-scheduled
405         &xor    ($E,$t1);
406          &add   ($T,&off($h));          # T += h
407          &xor   ("edi",$t2);            # Ch(e,f,g)
408         &ror    ($E,6);                 # Sigma1(e)
409          &mov   ($t1,$AH[0]);
410          &add   ($T,"edi");             # T += Ch(e,f,g)
411
412         &ror    ($t1,22-13);
413          &mov   ($t2,$AH[0]);
414          &mov   ("edi",&off($b));
415         &xor    ($t1,$AH[0]);
416          &mov   (&off($a),$AH[0]);      # save $A, modulo-scheduled
417          &xor   ($AH[0],"edi");         # a ^= b, (b^c) in next round
418         &ror    ($t1,13-2);
419          &and   ($AH[1],$AH[0]);        # (b^c) &= (a^b)
420          &lea   ($E,&DWP(@K256[$i],$T,$E));     # T += Sigma1(1)+K[i]
421         &xor    ($t1,$t2);
422          &xor   ($AH[1],"edi");         # h = Maj(a,b,c) = Ch(a^b,c,b)
423          &mov   ($t2,&DWP(32+4*(($i+2)&15),"esp"))      if ($i>=15 && $i<63);
424         &ror    ($t1,2);                # Sigma0(a)
425
426          &add   ($AH[1],$E);            # h += T
427          &add   ($E,&off($d));          # d += T
428         &add    ($AH[1],$t1);           # h += Sigma0(a)
429          &mov   ($t1,&DWP(32+4*(($i+15)&15),"esp"))     if ($i>=15 && $i<63);
430
431         @AH = reverse(@AH);             # rotate(a,h)
432         ($t1,$t2) = ($t2,$t1);          # rotate(t1,t2)
433     }
434         &mov    ("esi",&DWP(96,"esp")); #ctx
435                                         #&mov   ($AH[0],&DWP(0,"esp"));
436         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
437                                         #&mov   ("edi", &DWP(8,"esp"));
438         &mov    ("ecx",&DWP(12,"esp"));
439         &add    ($AH[0],&DWP(0,"esi"));
440         &add    ($AH[1],&DWP(4,"esi"));
441         &add    ("edi",&DWP(8,"esi"));
442         &add    ("ecx",&DWP(12,"esi"));
443         &mov    (&DWP(0,"esi"),$AH[0]);
444         &mov    (&DWP(4,"esi"),$AH[1]);
445         &mov    (&DWP(8,"esi"),"edi");
446         &mov    (&DWP(12,"esi"),"ecx");
447          #&mov  (&DWP(0,"esp"),$AH[0]);
448          &mov   (&DWP(4,"esp"),$AH[1]);
449          &xor   ($AH[1],"edi");         # magic
450          &mov   (&DWP(8,"esp"),"edi");
451          &mov   (&DWP(12,"esp"),"ecx");
452         #&mov   ($E,&DWP(16,"esp"));
453         &mov    ("edi",&DWP(20,"esp"));
454         &mov    ("ebx",&DWP(24,"esp"));
455         &mov    ("ecx",&DWP(28,"esp"));
456         &add    ($E,&DWP(16,"esi"));
457         &add    ("edi",&DWP(20,"esi"));
458         &add    ("ebx",&DWP(24,"esi"));
459         &add    ("ecx",&DWP(28,"esi"));
460         &mov    (&DWP(16,"esi"),$E);
461         &mov    (&DWP(20,"esi"),"edi");
462         &mov    (&DWP(24,"esi"),"ebx");
463         &mov    (&DWP(28,"esi"),"ecx");
464          #&mov  (&DWP(16,"esp"),$E);
465          &mov   (&DWP(20,"esp"),"edi");
466         &mov    ("edi",&DWP(96+4,"esp"));       # inp
467          &mov   (&DWP(24,"esp"),"ebx");
468          &mov   (&DWP(28,"esp"),"ecx");
469
470         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
471         &jb     (&label("grand_loop"));
472
473         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
474 &function_end_A();
475
476                                                 if ($ymm) {{{
477 my @X = map("xmm$_",(0..3));
478 my ($t0,$t1,$t2,$t3) = map("xmm$_",(4..7));
479 my @AH = ($A,$T);
480
481 &set_label("XOP",16);
482         &lea    ("esp",&DWP(-96,"esp"));
483         &vzeroall       ();
484         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
485         &mov    ($AH[0],&DWP(0,"esi"));
486         &mov    ($AH[1],&DWP(4,"esi"));
487         &mov    ("ecx",&DWP(8,"esi"));
488         &mov    ("edi",&DWP(12,"esi"));
489         #&mov   (&DWP(0,"esp"),$AH[0]);
490         &mov    (&DWP(4,"esp"),$AH[1]);
491         &xor    ($AH[1],"ecx");                 # magic
492         &mov    (&DWP(8,"esp"),"ecx");
493         &mov    (&DWP(12,"esp"),"edi");
494         &mov    ($E,&DWP(16,"esi"));
495         &mov    ("edi",&DWP(20,"esi"));
496         &mov    ("ecx",&DWP(24,"esi"));
497         &mov    ("esi",&DWP(28,"esi"));
498         #&mov   (&DWP(16,"esp"),$E);
499         &mov    (&DWP(20,"esp"),"edi");
500         &mov    ("edi",&DWP(96+4,"esp"));       # inp
501         &mov    (&DWP(24,"esp"),"ecx");
502         &mov    (&DWP(28,"esp"),"esi");
503         &vmovdqa        ($t3,&QWP(256,$K256));
504         &jmp    (&label("grand_xop"));
505
506 &set_label("grand_xop",16);
507         # load input, reverse byte order, add K256[0..15], save to stack
508         &vmovdqu        (@X[0],&QWP(0,"edi"));
509         &vmovdqu        (@X[1],&QWP(16,"edi"));
510         &vmovdqu        (@X[2],&QWP(32,"edi"));
511         &vmovdqu        (@X[3],&QWP(48,"edi"));
512         &add            ("edi",64);
513         &vpshufb        (@X[0],@X[0],$t3);
514         &mov            (&DWP(96+4,"esp"),"edi");
515         &vpshufb        (@X[1],@X[1],$t3);
516         &vpshufb        (@X[2],@X[2],$t3);
517         &vpaddd         ($t0,@X[0],&QWP(0,$K256));
518         &vpshufb        (@X[3],@X[3],$t3);
519         &vpaddd         ($t1,@X[1],&QWP(16,$K256));
520         &vpaddd         ($t2,@X[2],&QWP(32,$K256));
521         &vpaddd         ($t3,@X[3],&QWP(48,$K256));
522         &vmovdqa        (&QWP(32+0,"esp"),$t0);
523         &vmovdqa        (&QWP(32+16,"esp"),$t1);
524         &vmovdqa        (&QWP(32+32,"esp"),$t2);
525         &vmovdqa        (&QWP(32+48,"esp"),$t3);
526         &jmp            (&label("xop_00_47"));
527
528 &set_label("xop_00_47",16);
529         &add            ($K256,64);
530
531 sub XOP_00_47 () {
532 my $j = shift;
533 my $body = shift;
534 my @X = @_;
535 my @insns = (&$body,&$body,&$body,&$body);      # 120 instructions
536
537         &vpalignr       ($t0,@X[1],@X[0],4);    # X[1..4]
538           eval(shift(@insns));
539           eval(shift(@insns));
540          &vpalignr      ($t3,@X[3],@X[2],4);    # X[9..12]
541           eval(shift(@insns));
542           eval(shift(@insns));
543         &vprotd         ($t1,$t0,14);
544           eval(shift(@insns));
545           eval(shift(@insns));
546         &vpsrld         ($t0,$t0,3);
547          &vpaddd        (@X[0],@X[0],$t3);      # X[0..3] += X[9..12]
548           eval(shift(@insns));
549           eval(shift(@insns));
550           eval(shift(@insns));
551           eval(shift(@insns));
552         &vprotd         ($t2,$t1,25-14);
553         &vpxor          ($t0,$t0,$t1);
554           eval(shift(@insns));
555           eval(shift(@insns));
556           eval(shift(@insns));
557           eval(shift(@insns));
558          &vprotd        ($t3,@X[3],13);
559         &vpxor          ($t0,$t0,$t2);          # sigma0(X[1..4])
560           eval(shift(@insns));
561           eval(shift(@insns));
562          &vpsrld        ($t2,@X[3],10);
563           eval(shift(@insns));
564           eval(shift(@insns));
565         &vpaddd         (@X[0],@X[0],$t0);      # X[0..3] += sigma0(X[1..4])
566           eval(shift(@insns));
567           eval(shift(@insns));
568          &vprotd        ($t1,$t3,15-13);
569          &vpxor         ($t3,$t3,$t2);
570           eval(shift(@insns));
571           eval(shift(@insns));
572           eval(shift(@insns));
573           eval(shift(@insns));
574          &vpxor         ($t3,$t3,$t1);          # sigma1(X[14..15])
575           eval(shift(@insns));
576           eval(shift(@insns));
577           eval(shift(@insns));
578           eval(shift(@insns));
579         &vpsrldq        ($t3,$t3,8);
580           eval(shift(@insns));
581           eval(shift(@insns));
582           eval(shift(@insns));
583           eval(shift(@insns));
584         &vpaddd         (@X[0],@X[0],$t3);      # X[0..1] += sigma1(X[14..15])
585           eval(shift(@insns));
586           eval(shift(@insns));
587           eval(shift(@insns));
588           eval(shift(@insns));
589          &vprotd        ($t3,@X[0],13);
590           eval(shift(@insns));
591           eval(shift(@insns));
592          &vpsrld        ($t2,@X[0],10);
593           eval(shift(@insns));
594           eval(shift(@insns));
595          &vprotd        ($t1,$t3,15-13);
596           eval(shift(@insns));
597           eval(shift(@insns));
598          &vpxor         ($t3,$t3,$t2);
599           eval(shift(@insns));
600           eval(shift(@insns));
601           eval(shift(@insns));
602           eval(shift(@insns));
603          &vpxor         ($t3,$t3,$t1);          # sigma1(X[16..17])
604           eval(shift(@insns));
605           eval(shift(@insns));
606           eval(shift(@insns));
607           eval(shift(@insns));
608         &vpslldq        ($t3,$t3,8);            # 22 instructions
609           eval(shift(@insns));
610           eval(shift(@insns));
611           eval(shift(@insns));
612           eval(shift(@insns));
613         &vpaddd         (@X[0],@X[0],$t3);      # X[2..3] += sigma1(X[16..17])
614           eval(shift(@insns));
615           eval(shift(@insns));
616           eval(shift(@insns));
617           eval(shift(@insns));
618         &vpaddd         ($t2,@X[0],&QWP(16*$j,$K256));
619
620         foreach (@insns) { eval; }              # remaining instructions
621
622         &vmovdqa        (&QWP(32+16*$j,"esp"),$t2);
623 }
624
625 sub body_00_15 () {
626         (
627         '&mov   ("ecx",$E);',
628          '&mov  ("esi",&off($f));',
629         '&ror   ($E,25-11);',
630          '&mov  ("edi",&off($g));',
631         '&xor   ($E,"ecx");',
632          '&xor  ("esi","edi");',
633         '&ror   ($E,11-6);',
634          '&and  ("esi","ecx");',
635          '&mov  (&off($e),"ecx");',     # save $E, modulo-scheduled
636         '&xor   ($E,"ecx");',
637          '&xor  ("edi","esi");',        # Ch(e,f,g)
638         '&ror   ($E,6);',               # T = Sigma1(e)
639          '&mov  ("ecx",$AH[0]);',
640         '&mov   ("esi",$AH[0]);',
641          '&add  ($E,&off($h));',        # T += h
642
643         '&ror   ("ecx",22-13);',
644          '&add  ($E,"edi");',           # T += Ch(e,f,g)
645          '&mov  ("edi",&off($b));',
646         '&xor   ("ecx",$AH[0]);',
647          '&mov  (&off($a),$AH[0]);',    # save $A, modulo-scheduled
648          '&xor  ($AH[0],"edi");',       # a ^= b, (b^c) in next round
649         '&ror   ("ecx",13-2);',
650          '&and  ($AH[1],$AH[0]);',      # (b^c) &= (a^b)
651          '&add  ($E,&DWP(32+4*($i&15),"esp"));',        # T += K[i]+X[i]
652         '&xor   ("ecx","esi");',
653          '&xor  ($AH[1],"edi");',       # h = Maj(a,b,c) = Ch(a^b,c,b)
654         '&ror   ("ecx",2);',            # Sigma0(a)
655
656          '&add  ($AH[1],$E);',          # h += T
657          '&add  ($E,&off($d));',        # d += T
658         '&add   ($AH[1],"ecx");'.       # h += Sigma0(a)
659
660         '@AH = reverse(@AH); $i++;'     # rotate(a,h)
661         );
662 }
663
664     for ($i=0,$j=0; $j<4; $j++) {
665         &XOP_00_47($j,\&body_00_15,@X);
666         push(@X,shift(@X));             # rotate(@X)
667     }
668         &cmp    (&DWP(16*$j,$K256),0x00010203);
669         &jne    (&label("xop_00_47"));
670
671     for ($i=0; $i<16; ) {
672         foreach(body_00_15()) { eval; }
673     }
674
675         &mov    ("esi",&DWP(96,"esp")); #ctx
676                                         #&mov   ($AH[0],&DWP(0,"esp"));
677         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
678                                         #&mov   ("edi", &DWP(8,"esp"));
679         &mov    ("ecx",&DWP(12,"esp"));
680         &add    ($AH[0],&DWP(0,"esi"));
681         &add    ($AH[1],&DWP(4,"esi"));
682         &add    ("edi",&DWP(8,"esi"));
683         &add    ("ecx",&DWP(12,"esi"));
684         &mov    (&DWP(0,"esi"),$AH[0]);
685         &mov    (&DWP(4,"esi"),$AH[1]);
686         &mov    (&DWP(8,"esi"),"edi");
687         &mov    (&DWP(12,"esi"),"ecx");
688          #&mov  (&DWP(0,"esp"),$AH[0]);
689          &mov   (&DWP(4,"esp"),$AH[1]);
690          &xor   ($AH[1],"edi");                 # magic
691          &mov   (&DWP(8,"esp"),"edi");
692          &mov   (&DWP(12,"esp"),"ecx");
693         #&mov   ($E,&DWP(16,"esp"));
694         &mov    ("edi",&DWP(20,"esp"));
695         &mov    ("ecx",&DWP(24,"esp"));
696         &add    ($E,&DWP(16,"esi"));
697         &add    ("edi",&DWP(20,"esi"));
698         &add    ("ecx",&DWP(24,"esi"));
699         &mov    (&DWP(16,"esi"),$E);
700         &mov    (&DWP(20,"esi"),"edi");
701          &mov   (&DWP(20,"esp"),"edi");
702         &mov    ("edi",&DWP(28,"esp"));
703         &mov    (&DWP(24,"esi"),"ecx");
704          #&mov  (&DWP(16,"esp"),$E);
705         &add    ("edi",&DWP(28,"esi"));
706          &mov   (&DWP(24,"esp"),"ecx");
707         &mov    (&DWP(28,"esi"),"edi");
708          &mov   (&DWP(28,"esp"),"edi");
709         &mov    ("edi",&DWP(96+4,"esp"));       # inp
710
711         &vmovdqa        ($t3,&QWP(64,$K256));
712         &sub    ($K256,3*64);                   # rewind K
713         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
714         &jb     (&label("grand_xop"));
715
716         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
717         &vzeroall       ();
718 &function_end_A();
719
720 &set_label("AVX",16);
721         &lea    ("esp",&DWP(-96,"esp"));
722         &vzeroall       ();
723         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
724         &mov    ($AH[0],&DWP(0,"esi"));
725         &mov    ($AH[1],&DWP(4,"esi"));
726         &mov    ("ecx",&DWP(8,"esi"));
727         &mov    ("edi",&DWP(12,"esi"));
728         #&mov   (&DWP(0,"esp"),$AH[0]);
729         &mov    (&DWP(4,"esp"),$AH[1]);
730         &xor    ($AH[1],"ecx");                 # magic
731         &mov    (&DWP(8,"esp"),"ecx");
732         &mov    (&DWP(12,"esp"),"edi");
733         &mov    ($E,&DWP(16,"esi"));
734         &mov    ("edi",&DWP(20,"esi"));
735         &mov    ("ecx",&DWP(24,"esi"));
736         &mov    ("esi",&DWP(28,"esi"));
737         #&mov   (&DWP(16,"esp"),$E);
738         &mov    (&DWP(20,"esp"),"edi");
739         &mov    ("edi",&DWP(96+4,"esp"));       # inp
740         &mov    (&DWP(24,"esp"),"ecx");
741         &mov    (&DWP(28,"esp"),"esi");
742         &vmovdqa        ($t3,&QWP(256,$K256));
743         &jmp    (&label("grand_avx"));
744
745 &set_label("grand_avx",16);
746         # load input, reverse byte order, add K256[0..15], save to stack
747         &vmovdqu        (@X[0],&QWP(0,"edi"));
748         &vmovdqu        (@X[1],&QWP(16,"edi"));
749         &vmovdqu        (@X[2],&QWP(32,"edi"));
750         &vmovdqu        (@X[3],&QWP(48,"edi"));
751         &add            ("edi",64);
752         &vpshufb        (@X[0],@X[0],$t3);
753         &mov            (&DWP(96+4,"esp"),"edi");
754         &vpshufb        (@X[1],@X[1],$t3);
755         &vpshufb        (@X[2],@X[2],$t3);
756         &vpaddd         ($t0,@X[0],&QWP(0,$K256));
757         &vpshufb        (@X[3],@X[3],$t3);
758         &vpaddd         ($t1,@X[1],&QWP(16,$K256));
759         &vpaddd         ($t2,@X[2],&QWP(32,$K256));
760         &vpaddd         ($t3,@X[3],&QWP(48,$K256));
761         &vmovdqa        (&QWP(32+0,"esp"),$t0);
762         &vmovdqa        (&QWP(32+16,"esp"),$t1);
763         &vmovdqa        (&QWP(32+32,"esp"),$t2);
764         &vmovdqa        (&QWP(32+48,"esp"),$t3);
765         &jmp            (&label("avx_00_47"));
766
767 &set_label("avx_00_47",16);
768         &add            ($K256,64);
769
770 sub Xupdate_AVX () {
771         (
772         '&vpalignr      ($t0,@X[1],@X[0],4);',  # X[1..4]
773          '&vpalignr     ($t3,@X[3],@X[2],4);',  # X[9..12]
774         '&vpsrld        ($t2,$t0,7);',
775          '&vpaddd       (@X[0],@X[0],$t3);',    # X[0..3] += X[9..16]
776         '&vpsrld        ($t3,$t0,3);',
777         '&vpslld        ($t1,$t0,14);',
778         '&vpxor         ($t0,$t3,$t2);',
779          '&vpshufd      ($t3,@X[3],0b11111010)',# X[14..15]
780         '&vpsrld        ($t2,$t2,18-7);',
781         '&vpxor         ($t0,$t0,$t1);',
782         '&vpslld        ($t1,$t1,25-14);',
783         '&vpxor         ($t0,$t0,$t2);',
784          '&vpsrld       ($t2,$t3,10);',
785         '&vpxor         ($t0,$t0,$t1);',        # sigma0(X[1..4])
786          '&vpsrlq       ($t1,$t3,17);',
787         '&vpaddd        (@X[0],@X[0],$t0);',    # X[0..3] += sigma0(X[1..4])
788          '&vpxor        ($t2,$t2,$t1);',
789          '&vpsrlq       ($t3,$t3,19);',
790          '&vpxor        ($t2,$t2,$t3);',        # sigma1(X[14..15]
791          '&vpshufd      ($t3,$t2,0b10000100);',
792         '&vpsrldq       ($t3,$t3,8);',
793         '&vpaddd        (@X[0],@X[0],$t3);',    # X[0..1] += sigma1(X[14..15])
794          '&vpshufd      ($t3,@X[0],0b01010000)',# X[16..17]
795          '&vpsrld       ($t2,$t3,10);',
796          '&vpsrlq       ($t1,$t3,17);',
797          '&vpxor        ($t2,$t2,$t1);',
798          '&vpsrlq       ($t3,$t3,19);',
799          '&vpxor        ($t2,$t2,$t3);',        # sigma1(X[16..17]
800          '&vpshufd      ($t3,$t2,0b11101000);',
801         '&vpslldq       ($t3,$t3,8);',
802         '&vpaddd        (@X[0],@X[0],$t3);'     # X[2..3] += sigma1(X[16..17])
803         );
804 }
805
806 local *ror = sub { &shrd(@_[0],@_) };
807 sub AVX_00_47 () {
808 my $j = shift;
809 my $body = shift;
810 my @X = @_;
811 my @insns = (&$body,&$body,&$body,&$body);      # 120 instructions
812
813         foreach (Xupdate_AVX()) {               # 31 instructions
814             eval;
815             eval(shift(@insns));
816             eval(shift(@insns));
817             eval(shift(@insns));
818         }
819         &vpaddd         ($t2,@X[0],&QWP(16*$j,$K256));
820         foreach (@insns) { eval; }              # remaining instructions
821         &vmovdqa        (&QWP(32+16*$j,"esp"),$t2);
822 }
823
824     for ($i=0,$j=0; $j<4; $j++) {
825         &AVX_00_47($j,\&body_00_15,@X);
826         push(@X,shift(@X));             # rotate(@X)
827     }
828         &cmp    (&DWP(16*$j,$K256),0x00010203);
829         &jne    (&label("avx_00_47"));
830
831     for ($i=0; $i<16; ) {
832         foreach(body_00_15()) { eval; }
833     }
834
835         &mov    ("esi",&DWP(96,"esp")); #ctx
836                                         #&mov   ($AH[0],&DWP(0,"esp"));
837         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
838                                         #&mov   ("edi", &DWP(8,"esp"));
839         &mov    ("ecx",&DWP(12,"esp"));
840         &add    ($AH[0],&DWP(0,"esi"));
841         &add    ($AH[1],&DWP(4,"esi"));
842         &add    ("edi",&DWP(8,"esi"));
843         &add    ("ecx",&DWP(12,"esi"));
844         &mov    (&DWP(0,"esi"),$AH[0]);
845         &mov    (&DWP(4,"esi"),$AH[1]);
846         &mov    (&DWP(8,"esi"),"edi");
847         &mov    (&DWP(12,"esi"),"ecx");
848          #&mov  (&DWP(0,"esp"),$AH[0]);
849          &mov   (&DWP(4,"esp"),$AH[1]);
850          &xor   ($AH[1],"edi");                 # magic
851          &mov   (&DWP(8,"esp"),"edi");
852          &mov   (&DWP(12,"esp"),"ecx");
853         #&mov   ($E,&DWP(16,"esp"));
854         &mov    ("edi",&DWP(20,"esp"));
855         &mov    ("ecx",&DWP(24,"esp"));
856         &add    ($E,&DWP(16,"esi"));
857         &add    ("edi",&DWP(20,"esi"));
858         &add    ("ecx",&DWP(24,"esi"));
859         &mov    (&DWP(16,"esi"),$E);
860         &mov    (&DWP(20,"esi"),"edi");
861          &mov   (&DWP(20,"esp"),"edi");
862         &mov    ("edi",&DWP(28,"esp"));
863         &mov    (&DWP(24,"esi"),"ecx");
864          #&mov  (&DWP(16,"esp"),$E);
865         &add    ("edi",&DWP(28,"esi"));
866          &mov   (&DWP(24,"esp"),"ecx");
867         &mov    (&DWP(28,"esi"),"edi");
868          &mov   (&DWP(28,"esp"),"edi");
869         &mov    ("edi",&DWP(96+4,"esp"));       # inp
870
871         &vmovdqa        ($t3,&QWP(64,$K256));
872         &sub    ($K256,3*64);                   # rewind K
873         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
874         &jb     (&label("grand_avx"));
875
876         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
877         &vzeroall       ();
878 &function_end_A();
879                                                 }}}
880 }
881 &function_end_B("sha256_block_data_order");
882 &asciz("SHA256 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
883
884 &asm_finish();