1 #!/usr/bin/env perl
2 #
3 # ====================================================================
4 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
5 # project. Rights for redistribution and usage in source and binary
6 # forms are granted according to the OpenSSL license.
7 # ====================================================================
8 #
9 # sha256/512_block procedure for x86_64.
10 #
11 # 40% improvement over compiler-generated code on Opteron. On EM64T
12 # sha256 was observed to run >80% faster and sha512 - >40%. No magical
13 # tricks, just straight implementation... I really wonder why gcc
14 # [being armed with inline assembler] fails to generate as fast code.
15 # The only thing which is cool about this module is that it's very
16 # same instruction sequence used for both SHA-256 and SHA-512. In
17 # former case the instructions operate on 32-bit operands, while in
18 # latter - on 64-bit ones. All I had to do is to get one flavor right,
19 # the other one passed the test right away:-)
20 #
21 # sha256_block runs in ~1005 cycles on Opteron, which gives you
22 # asymptotic performance of 64*1000/1005=63.7MBps times CPU clock
23 # frequency in GHz. sha512_block runs in ~1275 cycles, which results
24 # in 128*1000/1275=100MBps per GHz. Is there room for improvement?
25 # Well, if you compare it to IA-64 implementation, which maintains
26 # X[16] in register bank[!], tends to 4 instructions per CPU clock
27 # cycle and runs in 1003 cycles, 1275 is very good result for 3-way
28 # issue Opteron pipeline and X[16] maintained in memory. So that *if*
29 # there is a way to improve it, *then* the only way would be to try to
30 # offload X[16] updates to SSE unit, but that would require "deeper"
31 # loop unroll, which in turn would naturally cause size blow-up, not
32 # to mention increased complexity! And once again, only *if* it's
33 # actually possible to noticeably improve overall ILP, instruction
34 # level parallelism, on a given CPU implementation in this case.
35 #
36 # Special note on Intel EM64T. While Opteron CPU exhibits perfect
37 # perfromance ratio of 1.5 between 64- and 32-bit flavors [see above],
38 # [currently available] EM64T CPUs apparently are far from it. On the
39 # contrary, 64-bit version, sha512_block, is ~30% *slower* than 32-bit
40 # sha256_block:-( This is presumably because 64-bit shifts/rotates
41 # apparently are not atomic instructions, but implemented in microcode.
42
43 #
44 # OpenSolaris OS modifications
45 #
46 # Sun elects to use this software under the BSD license.
47 #
48 # This source originates from OpenSSL file sha512-x86_64.pl at
49 # ftp://ftp.openssl.org/snapshot/openssl-0.9.8-stable-SNAP-20080131.tar.gz
50 # (presumably for future OpenSSL release 0.9.8h), with these changes:
51 #
52 # 1. Added perl "use strict" and declared variables.
53 #
54 # 2. Added OpenSolaris ENTRY_NP/SET_SIZE macros from
55 # /usr/include/sys/asm_linkage.h, .ident keywords, and lint(1B) guards.
56 #
57 # 3. Added perl function &lea_offset_eax_register_register() to handle
58 # OpenSolaris as(1) bug.
59 #
60 # 4. Removed x86_64-xlate.pl script (not needed for as(1) or gas(1)
61 # assemblers). Replaced the .picmeup macro with assembler code.
62 #
63 # 5. Added 8 to $ctx, as OpenSolaris OS has an extra 4-byte field, "algotype",
64 # at the beginning of SHA2_CTX (the next field is 8-byte aligned).
65 #
66
67 use strict;
68 my ($code, $func, $TABLE, $SZ, @Sigma0, @Sigma1, @sigma0, @sigma1, $rounds,
69 @ROT, $A, $B, $C, $D, $E, $F, $G, $H, $T1, $a0, $a1, $a2, $i,
70 $ctx, $round, $inp, $Tbl, $_ctx, $_inp, $_end, $_rsp, $framesz);
71 my $output = shift;
72 open STDOUT,">$output";
73
74 #
75 # OpenSSL library:
76 # void sha512_block_data_order(SHA512_CTX *ctx, const void *in, size_t num);
77 # void sha256_block_data_order(SHA256_CTX *ctx, const void *in, size_t num);
78 #
79 # OpenSolaris OS:
80 # void SHA512TransformBlocks(SHA2_CTX *ctx, const void *in, size_t num);
81 # void SHA256TransformBlocks(SHA2_CTX *ctx, const void *in, size_t num);
82 # Note: the OpenSolaris SHA2 structure has an extra 8 byte field at the
83 # beginning (over OpenSSL's SHA512 or SHA256 structure).
84 #
85
86 if ($output =~ /512/) {
87 $func="SHA512TransformBlocks";
88 $TABLE="K512";
89 $SZ=8;
90 @ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%rax","%rbx","%rcx","%rdx",
91 "%r8", "%r9", "%r10","%r11");
92 ($T1,$a0,$a1,$a2)=("%r12","%r13","%r14","%r15");
93 @Sigma0=(28,34,39);
94 @Sigma1=(14,18,41);
95 @sigma0=(1, 8, 7);
96 @sigma1=(19,61, 6);
97 $rounds=80;
98 } else {
99 $func="SHA256TransformBlocks";
100 $TABLE="K256";
101 $SZ=4;
102 @ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%eax","%ebx","%ecx","%edx",
103 "%r8d","%r9d","%r10d","%r11d");
104 ($T1,$a0,$a1,$a2)=("%r12d","%r13d","%r14d","%r15d");
105 @Sigma0=( 2,13,22);
106 @Sigma1=( 6,11,25);
107 @sigma0=( 7,18, 3);
108 @sigma1=(17,19,10);
109 $rounds=64;
110 }
111
112 $ctx="%rdi"; # 1st arg
113 $round="%rdi"; # zaps $ctx
114 $inp="%rsi"; # 2nd arg
115 $Tbl="%rbp";
116
117 $_ctx="16*$SZ+0*8(%rsp)";
118 $_inp="16*$SZ+1*8(%rsp)";
119 $_end="16*$SZ+2*8(%rsp)";
120 $_rsp="16*$SZ+3*8(%rsp)";
121 $framesz="16*$SZ+4*8";
122
123
124 sub ROUND_00_15()
125 { my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
126
127 $code.=<<___;
128 mov $e,$a0
129 mov $e,$a1
130 mov $f,$a2
131
132 ror \$$Sigma1[0],$a0
133 ror \$$Sigma1[1],$a1
134 xor $g,$a2 # f^g
135
136 xor $a1,$a0
137 ror \$`$Sigma1[2]-$Sigma1[1]`,$a1
138 and $e,$a2 # (f^g)&e
139 mov $T1,`$SZ*($i&0xf)`(%rsp)
140
141 xor $a1,$a0 # Sigma1(e)
142 xor $g,$a2 # Ch(e,f,g)=((f^g)&e)^g
143 add $h,$T1 # T1+=h
144
145 mov $a,$h
146 add $a0,$T1 # T1+=Sigma1(e)
147
148 add $a2,$T1 # T1+=Ch(e,f,g)
149 mov $a,$a0
150 mov $a,$a1
151
152 ror \$$Sigma0[0],$h
153 ror \$$Sigma0[1],$a0
154 mov $a,$a2
155 add ($Tbl,$round,$SZ),$T1 # T1+=K[round]
156
157 xor $a0,$h
158 ror \$`$Sigma0[2]-$Sigma0[1]`,$a0
159 or $c,$a1 # a|c
160
161 xor $a0,$h # h=Sigma0(a)
162 and $c,$a2 # a&c
163 add $T1,$d # d+=T1
164
165 and $b,$a1 # (a|c)&b
166 add $T1,$h # h+=T1
167
168 or $a2,$a1 # Maj(a,b,c)=((a|c)&b)|(a&c)
169 lea 1($round),$round # round++
170
171 add $a1,$h # h+=Maj(a,b,c)
172 ___
173 }
174
175 sub ROUND_16_XX()
176 { my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
177
178 $code.=<<___;
179 mov `$SZ*(($i+1)&0xf)`(%rsp),$a0
180 mov `$SZ*(($i+14)&0xf)`(%rsp),$T1
181
182 mov $a0,$a2
183
184 shr \$$sigma0[2],$a0
185 ror \$$sigma0[0],$a2
186
187 xor $a2,$a0
188 ror \$`$sigma0[1]-$sigma0[0]`,$a2
189
190 xor $a2,$a0 # sigma0(X[(i+1)&0xf])
191 mov $T1,$a1
192
193 shr \$$sigma1[2],$T1
194 ror \$$sigma1[0],$a1
195
196 xor $a1,$T1
197 ror \$`$sigma1[1]-$sigma1[0]`,$a1
198
199 xor $a1,$T1 # sigma1(X[(i+14)&0xf])
200
201 add $a0,$T1
202
203 add `$SZ*(($i+9)&0xf)`(%rsp),$T1
204
205 add `$SZ*($i&0xf)`(%rsp),$T1
206 ___
207 &ROUND_00_15(@_);
208 }
209
210 #
211 # Execution begins here
212 #
213
214 $code=<<___;
215 #if !defined(lint) && !defined(__lint)
216 .ident "@(#)sha512-x86_64.pl 1.2 08/03/20 SMI"
217 #include <sys/asm_linkage.h>
218
219 ENTRY_NP($func)
220 push %rbx
221 push %rbp
222 push %r12
223 push %r13
224 push %r14
225 push %r15
226 mov %rsp,%rbp # copy %rsp
227 shl \$4,%rdx # num*16
228 sub \$$framesz,%rsp
229 lea ($inp,%rdx,$SZ),%rdx # inp+num*16*$SZ
230 and \$-64,%rsp # align stack frame
231 add \$8,$ctx # Skip OpenSolaris field, "algotype"
232 mov $ctx,$_ctx # save ctx, 1st arg
233 mov $inp,$_inp # save inp, 2nd arg
234 mov %rdx,$_end # save end pointer, "3rd" arg
235 mov %rbp,$_rsp # save copy of %rsp
236
237 /.picmeup $Tbl
238 / The .picmeup pseudo-directive, from perlasm/x86_64_xlate.pl, puts
239 / the address of the "next" instruction into the target register
240 / ($Tbl). This generates these 2 instructions:
241 lea .Llea(%rip),$Tbl
242 /nop / .picmeup generates a nop for mod 8 alignment--not needed here
243
244 .Llea:
245 lea $TABLE-.($Tbl),$Tbl
246
247 mov $SZ*0($ctx),$A
248 mov $SZ*1($ctx),$B
249 mov $SZ*2($ctx),$C
250 mov $SZ*3($ctx),$D
251 mov $SZ*4($ctx),$E
252 mov $SZ*5($ctx),$F
253 mov $SZ*6($ctx),$G
254 mov $SZ*7($ctx),$H
255 jmp .Lloop
256
257 .align 16
258 .Lloop:
259 xor $round,$round
260 ___
261 for($i=0;$i<16;$i++) {
262 $code.=" mov $SZ*$i($inp),$T1\n";
263 $code.=" bswap $T1\n";
264 &ROUND_00_15($i,@ROT);
265 unshift(@ROT,pop(@ROT));
266 }
267 $code.=<<___;
268 jmp .Lrounds_16_xx
269 .align 16
270 .Lrounds_16_xx:
271 ___
272 for(;$i<32;$i++) {
273 &ROUND_16_XX($i,@ROT);
274 unshift(@ROT,pop(@ROT));
275 }
276
277 $code.=<<___;
278 cmp \$$rounds,$round
279 jb .Lrounds_16_xx
280
281 mov $_ctx,$ctx
282 lea 16*$SZ($inp),$inp
283
284 add $SZ*0($ctx),$A
285 add $SZ*1($ctx),$B
286 add $SZ*2($ctx),$C
287 add $SZ*3($ctx),$D
288 add $SZ*4($ctx),$E
289 add $SZ*5($ctx),$F
290 add $SZ*6($ctx),$G
291 add $SZ*7($ctx),$H
292
293 cmp $_end,$inp
294
295 mov $A,$SZ*0($ctx)
296 mov $B,$SZ*1($ctx)
297 mov $C,$SZ*2($ctx)
298 mov $D,$SZ*3($ctx)
299 mov $E,$SZ*4($ctx)
300 mov $F,$SZ*5($ctx)
301 mov $G,$SZ*6($ctx)
302 mov $H,$SZ*7($ctx)
303 jb .Lloop
304
305 mov $_rsp,%rsp
306 pop %r15
307 pop %r14
308 pop %r13
309 pop %r12
310 pop %rbp
311 pop %rbx
312
313 ret
314 SET_SIZE($func)
315
316 ___
317
318 if ($SZ==4) {
319 # SHA256
320 $code.=<<___;
321 .align 64
322 .type $TABLE,\@object
323 $TABLE:
324 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
325 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
326 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
327 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
328 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
329 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
330 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
331 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
332 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
333 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
334 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
335 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070
336 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
337 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
338 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
339 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
340 ___
341 } else {
342 # SHA512
343 $code.=<<___;
344 .align 64
345 .type $TABLE,\@object
346 $TABLE:
347 .quad 0x428a2f98d728ae22,0x7137449123ef65cd
348 .quad 0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
349 .quad 0x3956c25bf348b538,0x59f111f1b605d019
350 .quad 0x923f82a4af194f9b,0xab1c5ed5da6d8118
351 .quad 0xd807aa98a3030242,0x12835b0145706fbe
352 .quad 0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
353 .quad 0x72be5d74f27b896f,0x80deb1fe3b1696b1
354 .quad 0x9bdc06a725c71235,0xc19bf174cf692694
355 .quad 0xe49b69c19ef14ad2,0xefbe4786384f25e3
356 .quad 0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
357 .quad 0x2de92c6f592b0275,0x4a7484aa6ea6e483
358 .quad 0x5cb0a9dcbd41fbd4,0x76f988da831153b5
359 .quad 0x983e5152ee66dfab,0xa831c66d2db43210
360 .quad 0xb00327c898fb213f,0xbf597fc7beef0ee4
361 .quad 0xc6e00bf33da88fc2,0xd5a79147930aa725
362 .quad 0x06ca6351e003826f,0x142929670a0e6e70
363 .quad 0x27b70a8546d22ffc,0x2e1b21385c26c926
364 .quad 0x4d2c6dfc5ac42aed,0x53380d139d95b3df
365 .quad 0x650a73548baf63de,0x766a0abb3c77b2a8
366 .quad 0x81c2c92e47edaee6,0x92722c851482353b
367 .quad 0xa2bfe8a14cf10364,0xa81a664bbc423001
368 .quad 0xc24b8b70d0f89791,0xc76c51a30654be30
369 .quad 0xd192e819d6ef5218,0xd69906245565a910
370 .quad 0xf40e35855771202a,0x106aa07032bbd1b8
371 .quad 0x19a4c116b8d2d0c8,0x1e376c085141ab53
372 .quad 0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
373 .quad 0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
374 .quad 0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
375 .quad 0x748f82ee5defb2fc,0x78a5636f43172f60
376 .quad 0x84c87814a1f0ab72,0x8cc702081a6439ec
377 .quad 0x90befffa23631e28,0xa4506cebde82bde9
378 .quad 0xbef9a3f7b2c67915,0xc67178f2e372532b
379 .quad 0xca273eceea26619c,0xd186b8c721c0c207
380 .quad 0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
381 .quad 0x06f067aa72176fba,0x0a637dc5a2c898a6
382 .quad 0x113f9804bef90dae,0x1b710b35131c471b
383 .quad 0x28db77f523047d84,0x32caab7b40c72493
384 .quad 0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
385 .quad 0x4cc5d4becb3e42b6,0x597f299cfc657e2a
386 .quad 0x5fcb6fab3ad6faec,0x6c44198c4a475817
387 ___
388 }
389 $code.=<<___;
390
391 #else
392 /* LINTED */
393 /* Nothing to be linted in this file--it's pure assembly source. */
394 #endif /* !lint && !__lint */
395 ___
396
397 $code =~ s/\`([^\`]*)\`/eval $1/gem;
398 print $code;
399 close STDOUT;