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 # ====================================================================
9 # sha256/512_block procedure for x86_64.
11 # 40% improvement over compiler-generated code on Opteron. No magical
12 # tricks, just straight implementation... I really wonder why gcc
13 # [being armed with inline assembler] fails to generate as fast code.
14 # The only thing which is cool about this module is that it's very
15 # same instruction sequence used for both SHA-256 and SHA-512. In
16 # former case the instructions operate on 32-bit operands, while in
17 # latter - on 64-bit ones. All I had to do is to get one flavor right,
18 # the other one passed the test right away:-)
20 # sha256_block runs in ~1005 cycles on Opteron, which gives you
21 # asymptotic performance of 64*1000/1005=63.7MBps times CPU clock
22 # frequency in GHz. sha512_block runs in ~1275 cycles, which results
23 # in 128*1000/1275=100MBps per GHz. Is there room for improvement?
24 # Well, if you compare it to IA-64 implementation, which maintains
25 # X[16] in register bank[!], tends to 4 instructions per CPU clock
26 # cycle and runs in 1003 cycles, 1275 is very good result for 3-way
27 # issue Opteron pipeline and X[16] maintained in memory. So that *if*
28 # there is a way to improve it, *then* the only way would be to try to
29 # offload X[16] updates to SSE unit, but that would require "deeper"
30 # loop unroll, which in turn would naturally cause size blow-up, not
31 # to mention increased complexity! And once again, only *if* it's
32 # actually possible to noticeably improve overall ILP, instruction
33 # level parallelism, on a given CPU implementation in this case.
35 # Special note on Intel EM64T. While Opteron CPU exhibits perfect
36 # perfromance ratio of 1.5 between 64- and 32-bit flavors [see above],
37 # [currently available] EM64T CPUs apparently are far from it. 64-bit
38 # version, sha512_block, is hardly faster than 32-bit one. This is
39 # presumably because 64-bit shifts/rotates apparently are not atomic
40 # instructions, but implemented in microcode.
43 open STDOUT,"| $^X ../perlasm/x86_64-xlate.pl $output";
45 if ($output =~ /512/) {
49 @ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%rax","%rbx","%rcx","%rdx",
50 "%r8", "%r9", "%r10","%r11");
51 ($T1,$a0,$a1,$a2)=("%r12","%r13","%r14","%r15");
61 @ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%eax","%ebx","%ecx","%edx",
62 "%r8d","%r9d","%r10d","%r11d");
63 ($T1,$a0,$a1,$a2)=("%r12d","%r13d","%r14d","%r15d");
71 $ctx="%rdi"; # 1st arg
72 $round="%rdi"; # zaps $ctx
73 $inp="%rsi"; # 2nd arg
76 $_ctx="16*$SZ+0*8(%rsp)";
77 $_inp="16*$SZ+1*8(%rsp)";
78 $_end="16*$SZ+2*8(%rsp)";
79 $_ord="16*$SZ+3*8(%rsp)";
80 $_rsp="16*$SZ+4*8(%rsp)";
81 $framesz="16*$SZ+5*8";
85 { my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
97 ror \$`$Sigma1[2]-$Sigma1[1]`,$a1
99 mov $T1,`$SZ*($i&0xf)`(%rsp)
101 xor $a1,$a0 # Sigma1(e)
102 xor $g,$a2 # Ch(e,f,g)=((f^g)&e)^g
106 add $a0,$T1 # T1+=Sigma1(e)
108 add $a2,$T1 # T1+=Ch(e,f,g)
115 add ($Tbl,$round,$SZ),$T1 # T1+=K[round]
118 ror \$`$Sigma0[2]-$Sigma0[1]`,$a0
121 xor $a0,$h # h=Sigma0(a)
128 or $a2,$a1 # Maj(a,b,c)=((a|c)&b)|(a&c)
129 lea 1($round),$round # round++
131 add $a1,$h # h+=Maj(a,b,c)
136 { my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
139 mov `$SZ*(($i+1)&0xf)`(%rsp),$a0
140 mov `$SZ*(($i+14)&0xf)`(%rsp),$T1
148 ror \$`$sigma0[1]-$sigma0[0]`,$a2
150 xor $a2,$a0 # sigma0(X[(i+1)&0xf])
157 ror \$`$sigma1[1]-$sigma1[0]`,$a1
159 xor $a1,$T1 # sigma1(X[(i+14)&0xf])
163 add `$SZ*(($i+9)&0xf)`(%rsp),$T1
165 add `$SZ*($i&0xf)`(%rsp),$T1
174 .type $func,\@function,4
183 mov %rsp,%rbp # copy %rsp
184 shl \$4,%rdx # num*16
186 lea ($inp,%rdx,$SZ),%rdx # inp+num*16*$SZ
187 and \$-64,%rsp # align stack frame
188 mov $ctx,$_ctx # save ctx, 1st arg
189 mov $inp,$_inp # save inp, 2nd arh
190 mov %rdx,$_end # save end pointer, "3rd" arg
191 mov %ecx,$_ord # save host, 4th arg
192 mov %rbp,$_rsp # save copy of %rsp
195 lea $TABLE-.($Tbl),$Tbl
219 for($i=0;$i<16;$i++) {
220 $code.=" mov $SZ*$i($inp),$T1\n";
221 &ROUND_00_15($i,@ROT);
222 unshift(@ROT,pop(@ROT));
230 for($i=0;$i<16;$i++) {
231 $code.=" mov $SZ*$i($inp),$T1\n";
232 $code.=" bswap $T1\n";
233 &ROUND_00_15($i,@ROT);
234 unshift(@ROT,pop(@ROT));
242 &ROUND_16_XX($i,@ROT);
243 unshift(@ROT,pop(@ROT));
251 lea 16*$SZ($inp),$inp
289 .type $TABLE,\@object
291 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
292 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
293 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
294 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
295 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
296 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
297 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
298 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
299 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
300 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
301 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
302 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070
303 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
304 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
305 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
306 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
311 .type $TABLE,\@object
313 .quad 0x428a2f98d728ae22,0x7137449123ef65cd
314 .quad 0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
315 .quad 0x3956c25bf348b538,0x59f111f1b605d019
316 .quad 0x923f82a4af194f9b,0xab1c5ed5da6d8118
317 .quad 0xd807aa98a3030242,0x12835b0145706fbe
318 .quad 0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
319 .quad 0x72be5d74f27b896f,0x80deb1fe3b1696b1
320 .quad 0x9bdc06a725c71235,0xc19bf174cf692694
321 .quad 0xe49b69c19ef14ad2,0xefbe4786384f25e3
322 .quad 0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
323 .quad 0x2de92c6f592b0275,0x4a7484aa6ea6e483
324 .quad 0x5cb0a9dcbd41fbd4,0x76f988da831153b5
325 .quad 0x983e5152ee66dfab,0xa831c66d2db43210
326 .quad 0xb00327c898fb213f,0xbf597fc7beef0ee4
327 .quad 0xc6e00bf33da88fc2,0xd5a79147930aa725
328 .quad 0x06ca6351e003826f,0x142929670a0e6e70
329 .quad 0x27b70a8546d22ffc,0x2e1b21385c26c926
330 .quad 0x4d2c6dfc5ac42aed,0x53380d139d95b3df
331 .quad 0x650a73548baf63de,0x766a0abb3c77b2a8
332 .quad 0x81c2c92e47edaee6,0x92722c851482353b
333 .quad 0xa2bfe8a14cf10364,0xa81a664bbc423001
334 .quad 0xc24b8b70d0f89791,0xc76c51a30654be30
335 .quad 0xd192e819d6ef5218,0xd69906245565a910
336 .quad 0xf40e35855771202a,0x106aa07032bbd1b8
337 .quad 0x19a4c116b8d2d0c8,0x1e376c085141ab53
338 .quad 0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
339 .quad 0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
340 .quad 0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
341 .quad 0x748f82ee5defb2fc,0x78a5636f43172f60
342 .quad 0x84c87814a1f0ab72,0x8cc702081a6439ec
343 .quad 0x90befffa23631e28,0xa4506cebde82bde9
344 .quad 0xbef9a3f7b2c67915,0xc67178f2e372532b
345 .quad 0xca273eceea26619c,0xd186b8c721c0c207
346 .quad 0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
347 .quad 0x06f067aa72176fba,0x0a637dc5a2c898a6
348 .quad 0x113f9804bef90dae,0x1b710b35131c471b
349 .quad 0x28db77f523047d84,0x32caab7b40c72493
350 .quad 0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
351 .quad 0x4cc5d4becb3e42b6,0x597f299cfc657e2a
352 .quad 0x5fcb6fab3ad6faec,0x6c44198c4a475817
356 $code =~ s/\`([^\`]*)\`/eval $1/gem;