3 # ====================================================================
4 # Written by Andy Polyakov <appro@openssl.org> 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. 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:-)
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.
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.
45 # Optimization including one of Pavel Semjanov's ideas, alternative
46 # Maj, resulted in >=5% improvement on most CPUs, +20% SHA256 and
47 # unfortunately -2% SHA512 on P4 [which nobody should care about
52 # Add SIMD code paths, see below for improvement coefficients. SSSE3
53 # code path was not attempted for SHA512, because improvement is not
54 # estimated to be high enough, noticeably less than 9%, to justify
55 # the effort, not on pre-AVX processors. [Obviously with exclusion
56 # for VIA Nano, but it has SHA512 instruction that is faster and
57 # should be used instead.] For reference, corresponding estimated
58 # upper limit for improvement for SSSE3 SHA256 is 28%. The fact that
59 # higher coefficients are observed on VIA Nano and Bulldozer has more
60 # to do with specifics of their architecture [which is topic for
61 # separate discussion].
65 # Add AVX2 code path. Two consecutive input blocks are loaded to
66 # 256-bit %ymm registers, with data from first block to least
67 # significant 128-bit halves and data from second to most significant.
68 # The data is then processed with same SIMD instruction sequence as
69 # for AVX, but with %ymm as operands. Side effect is increased stack
70 # frame, 448 additional bytes in SHA256 and 1152 in SHA512, and 1.2KB
75 # Add support for Intel SHA Extensions.
77 ######################################################################
78 # Current performance in cycles per processed byte (less is better):
80 # SHA256 SSSE3 AVX/XOP(*) SHA512 AVX/XOP(*)
82 # AMD K8 14.9 - - 9.57 -
84 # Core 2 15.6 13.8(+13%) - 9.97 -
85 # Westmere 14.8 12.3(+19%) - 9.58 -
86 # Sandy Bridge 17.4 14.2(+23%) 11.6(+50%(**)) 11.2 8.10(+38%(**))
87 # Ivy Bridge 12.6 10.5(+20%) 10.3(+22%) 8.17 7.22(+13%)
88 # Haswell 12.2 9.28(+31%) 7.80(+56%) 7.66 5.40(+42%)
89 # Bulldozer 21.1 13.6(+54%) 13.6(+54%(***)) 13.5 8.58(+57%)
90 # VIA Nano 23.0 16.5(+39%) - 14.7 -
91 # Atom 23.0 18.9(+22%) - 14.7 -
92 # Silvermont 27.4 20.6(+33%) - 17.5 -
94 # (*) whichever best applicable;
95 # (**) switch from ror to shrd stands for fair share of improvement;
96 # (***) execution time is fully determined by remaining integer-only
97 # part, body_00_15; reducing the amount of SIMD instructions
98 # below certain limit makes no difference/sense; to conserve
99 # space SHA256 XOP code path is therefore omitted;
103 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
105 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
107 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
108 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
109 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
110 die "can't locate x86_64-xlate.pl";
112 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
113 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
114 $avx = ($1>=2.19) + ($1>=2.22);
117 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
118 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
119 $avx = ($1>=2.09) + ($1>=2.10);
122 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
123 `ml64 2>&1` =~ /Version ([0-9]+)\./) {
124 $avx = ($1>=10) + ($1>=11);
127 if (!$avx && `$ENV{CC} -v 2>&1` =~ /(^clang version|based on LLVM) ([3-9]\.[0-9]+)/) {
128 $avx = ($2>=3.0) + ($2>3.0);
131 $shaext=1; ### set to zero if compiling for 1.0.1
132 $avx=1 if (!$shaext && $avx);
134 open OUT,"| \"$^X\" $xlate $flavour $output";
137 if ($output =~ /512/) {
138 $func="sha512_block_data_order";
141 @ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%rax","%rbx","%rcx","%rdx",
142 "%r8", "%r9", "%r10","%r11");
143 ($T1,$a0,$a1,$a2,$a3)=("%r12","%r13","%r14","%r15","%rdi");
150 $func="sha256_block_data_order";
153 @ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%eax","%ebx","%ecx","%edx",
154 "%r8d","%r9d","%r10d","%r11d");
155 ($T1,$a0,$a1,$a2,$a3)=("%r12d","%r13d","%r14d","%r15d","%edi");
163 $ctx="%rdi"; # 1st arg, zapped by $a3
164 $inp="%rsi"; # 2nd arg
167 $_ctx="16*$SZ+0*8(%rsp)";
168 $_inp="16*$SZ+1*8(%rsp)";
169 $_end="16*$SZ+2*8(%rsp)";
170 $_rsp="16*$SZ+3*8(%rsp)";
171 $framesz="16*$SZ+4*8";
175 { my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
177 $STRIDE += 16 if ($i%(16/$SZ)==(16/$SZ-1));
180 ror \$`$Sigma1[2]-$Sigma1[1]`,$a0
184 ror \$`$Sigma0[2]-$Sigma0[1]`,$a1
187 mov $T1,`$SZ*($i&0xf)`(%rsp)
191 ror \$`$Sigma1[1]-$Sigma1[0]`,$a0
193 xor $g,$a2 # Ch(e,f,g)=((f^g)&e)^g
195 ror \$`$Sigma0[1]-$Sigma0[0]`,$a1
197 add $a2,$T1 # T1+=Ch(e,f,g)
200 add ($Tbl),$T1 # T1+=K[round]
203 xor $b,$a2 # a^b, b^c in next round
204 ror \$$Sigma1[0],$a0 # Sigma1(e)
208 ror \$$Sigma0[0],$a1 # Sigma0(a)
209 add $a0,$T1 # T1+=Sigma1(e)
211 xor $a3,$h # h=Maj(a,b,c)=Ch(a^b,c,b)
215 lea $STRIDE($Tbl),$Tbl # round++
217 $code.=<<___ if ($i<15);
218 add $a1,$h # h+=Sigma0(a)
220 ($a2,$a3) = ($a3,$a2);
224 { my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
227 mov `$SZ*(($i+1)&0xf)`(%rsp),$a0
228 mov `$SZ*(($i+14)&0xf)`(%rsp),$a2
231 ror \$`$sigma0[1]-$sigma0[0]`,$a0
232 add $a1,$a # modulo-scheduled h+=Sigma0(a)
234 ror \$`$sigma1[1]-$sigma1[0]`,$a2
243 xor $a0,$T1 # sigma0(X[(i+1)&0xf])
244 xor $a1,$a2 # sigma1(X[(i+14)&0xf])
245 add `$SZ*(($i+9)&0xf)`(%rsp),$T1
247 add `$SZ*($i&0xf)`(%rsp),$T1
258 .extern OPENSSL_ia32cap_P
260 .type $func,\@function,3
264 $code.=<<___ if ($SZ==4 || $avx);
265 lea OPENSSL_ia32cap_P(%rip),%r11
270 $code.=<<___ if ($SZ==4 && $shaext);
271 test \$`1<<29`,%r11d # check for SHA
274 $code.=<<___ if ($avx && $SZ==8);
275 test \$`1<<11`,%r10d # check for XOP
278 $code.=<<___ if ($avx>1);
279 and \$`1<<8|1<<5|1<<3`,%r11d # check for BMI2+AVX2+BMI1
280 cmp \$`1<<8|1<<5|1<<3`,%r11d
283 $code.=<<___ if ($avx);
284 and \$`1<<30`,%r9d # mask "Intel CPU" bit
285 and \$`1<<28|1<<9`,%r10d # mask AVX and SSSE3 bits
287 cmp \$`1<<28|1<<9|1<<30`,%r10d
290 $code.=<<___ if ($SZ==4);
301 mov %rsp,%r11 # copy %rsp
302 shl \$4,%rdx # num*16
304 lea ($inp,%rdx,$SZ),%rdx # inp+num*16*$SZ
305 and \$-64,%rsp # align stack frame
306 mov $ctx,$_ctx # save ctx, 1st arg
307 mov $inp,$_inp # save inp, 2nd arh
308 mov %rdx,$_end # save end pointer, "3rd" arg
309 mov %r11,$_rsp # save copy of %rsp
325 lea $TABLE(%rip),$Tbl
328 for($i=0;$i<16;$i++) {
329 $code.=" mov $SZ*$i($inp),$T1\n";
330 $code.=" mov @ROT[4],$a0\n";
331 $code.=" mov @ROT[0],$a1\n";
332 $code.=" bswap $T1\n";
333 &ROUND_00_15($i,@ROT);
334 unshift(@ROT,pop(@ROT));
342 &ROUND_16_XX($i,@ROT);
343 unshift(@ROT,pop(@ROT));
347 cmpb \$0,`$SZ-1`($Tbl)
351 add $a1,$A # modulo-scheduled h+=Sigma0(a)
352 lea 16*$SZ($inp),$inp
391 .type $TABLE,\@object
393 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
394 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
395 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
396 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
397 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
398 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
399 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
400 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
401 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
402 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
403 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
404 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
405 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
406 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
407 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
408 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
409 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
410 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
411 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
412 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
413 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
414 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
415 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070
416 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070
417 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
418 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
419 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
420 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
421 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
422 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
423 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
424 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
426 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f
427 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f
428 .long 0x03020100,0x0b0a0908,0xffffffff,0xffffffff
429 .long 0x03020100,0x0b0a0908,0xffffffff,0xffffffff
430 .long 0xffffffff,0xffffffff,0x03020100,0x0b0a0908
431 .long 0xffffffff,0xffffffff,0x03020100,0x0b0a0908
432 .asciz "SHA256 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
437 .type $TABLE,\@object
439 .quad 0x428a2f98d728ae22,0x7137449123ef65cd
440 .quad 0x428a2f98d728ae22,0x7137449123ef65cd
441 .quad 0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
442 .quad 0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
443 .quad 0x3956c25bf348b538,0x59f111f1b605d019
444 .quad 0x3956c25bf348b538,0x59f111f1b605d019
445 .quad 0x923f82a4af194f9b,0xab1c5ed5da6d8118
446 .quad 0x923f82a4af194f9b,0xab1c5ed5da6d8118
447 .quad 0xd807aa98a3030242,0x12835b0145706fbe
448 .quad 0xd807aa98a3030242,0x12835b0145706fbe
449 .quad 0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
450 .quad 0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
451 .quad 0x72be5d74f27b896f,0x80deb1fe3b1696b1
452 .quad 0x72be5d74f27b896f,0x80deb1fe3b1696b1
453 .quad 0x9bdc06a725c71235,0xc19bf174cf692694
454 .quad 0x9bdc06a725c71235,0xc19bf174cf692694
455 .quad 0xe49b69c19ef14ad2,0xefbe4786384f25e3
456 .quad 0xe49b69c19ef14ad2,0xefbe4786384f25e3
457 .quad 0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
458 .quad 0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
459 .quad 0x2de92c6f592b0275,0x4a7484aa6ea6e483
460 .quad 0x2de92c6f592b0275,0x4a7484aa6ea6e483
461 .quad 0x5cb0a9dcbd41fbd4,0x76f988da831153b5
462 .quad 0x5cb0a9dcbd41fbd4,0x76f988da831153b5
463 .quad 0x983e5152ee66dfab,0xa831c66d2db43210
464 .quad 0x983e5152ee66dfab,0xa831c66d2db43210
465 .quad 0xb00327c898fb213f,0xbf597fc7beef0ee4
466 .quad 0xb00327c898fb213f,0xbf597fc7beef0ee4
467 .quad 0xc6e00bf33da88fc2,0xd5a79147930aa725
468 .quad 0xc6e00bf33da88fc2,0xd5a79147930aa725
469 .quad 0x06ca6351e003826f,0x142929670a0e6e70
470 .quad 0x06ca6351e003826f,0x142929670a0e6e70
471 .quad 0x27b70a8546d22ffc,0x2e1b21385c26c926
472 .quad 0x27b70a8546d22ffc,0x2e1b21385c26c926
473 .quad 0x4d2c6dfc5ac42aed,0x53380d139d95b3df
474 .quad 0x4d2c6dfc5ac42aed,0x53380d139d95b3df
475 .quad 0x650a73548baf63de,0x766a0abb3c77b2a8
476 .quad 0x650a73548baf63de,0x766a0abb3c77b2a8
477 .quad 0x81c2c92e47edaee6,0x92722c851482353b
478 .quad 0x81c2c92e47edaee6,0x92722c851482353b
479 .quad 0xa2bfe8a14cf10364,0xa81a664bbc423001
480 .quad 0xa2bfe8a14cf10364,0xa81a664bbc423001
481 .quad 0xc24b8b70d0f89791,0xc76c51a30654be30
482 .quad 0xc24b8b70d0f89791,0xc76c51a30654be30
483 .quad 0xd192e819d6ef5218,0xd69906245565a910
484 .quad 0xd192e819d6ef5218,0xd69906245565a910
485 .quad 0xf40e35855771202a,0x106aa07032bbd1b8
486 .quad 0xf40e35855771202a,0x106aa07032bbd1b8
487 .quad 0x19a4c116b8d2d0c8,0x1e376c085141ab53
488 .quad 0x19a4c116b8d2d0c8,0x1e376c085141ab53
489 .quad 0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
490 .quad 0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
491 .quad 0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
492 .quad 0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
493 .quad 0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
494 .quad 0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
495 .quad 0x748f82ee5defb2fc,0x78a5636f43172f60
496 .quad 0x748f82ee5defb2fc,0x78a5636f43172f60
497 .quad 0x84c87814a1f0ab72,0x8cc702081a6439ec
498 .quad 0x84c87814a1f0ab72,0x8cc702081a6439ec
499 .quad 0x90befffa23631e28,0xa4506cebde82bde9
500 .quad 0x90befffa23631e28,0xa4506cebde82bde9
501 .quad 0xbef9a3f7b2c67915,0xc67178f2e372532b
502 .quad 0xbef9a3f7b2c67915,0xc67178f2e372532b
503 .quad 0xca273eceea26619c,0xd186b8c721c0c207
504 .quad 0xca273eceea26619c,0xd186b8c721c0c207
505 .quad 0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
506 .quad 0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
507 .quad 0x06f067aa72176fba,0x0a637dc5a2c898a6
508 .quad 0x06f067aa72176fba,0x0a637dc5a2c898a6
509 .quad 0x113f9804bef90dae,0x1b710b35131c471b
510 .quad 0x113f9804bef90dae,0x1b710b35131c471b
511 .quad 0x28db77f523047d84,0x32caab7b40c72493
512 .quad 0x28db77f523047d84,0x32caab7b40c72493
513 .quad 0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
514 .quad 0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
515 .quad 0x4cc5d4becb3e42b6,0x597f299cfc657e2a
516 .quad 0x4cc5d4becb3e42b6,0x597f299cfc657e2a
517 .quad 0x5fcb6fab3ad6faec,0x6c44198c4a475817
518 .quad 0x5fcb6fab3ad6faec,0x6c44198c4a475817
520 .quad 0x0001020304050607,0x08090a0b0c0d0e0f
521 .quad 0x0001020304050607,0x08090a0b0c0d0e0f
522 .asciz "SHA512 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
526 ######################################################################
529 if ($SZ==4 && $shaext) {{{
530 ######################################################################
531 # Intel SHA Extensions implementation of SHA256 update function.
533 my ($ctx,$inp,$num,$Tbl)=("%rdi","%rsi","%rdx","%rcx");
535 my ($Wi,$ABEF,$CDGH,$TMP,$BSWAP,$ABEF_SAVE,$CDGH_SAVE)=map("%xmm$_",(0..2,7..10));
536 my @MSG=map("%xmm$_",(3..6));
539 .type sha256_block_data_order_shaext,\@function,3
541 sha256_block_data_order_shaext:
544 $code.=<<___ if ($win64);
545 lea `-8-5*16`(%rsp),%rsp
546 movaps %xmm6,-8-5*16(%rax)
547 movaps %xmm7,-8-4*16(%rax)
548 movaps %xmm8,-8-3*16(%rax)
549 movaps %xmm9,-8-2*16(%rax)
550 movaps %xmm10,-8-1*16(%rax)
554 lea K256+0x80(%rip),$Tbl
555 movdqu ($ctx),$ABEF # DCBA
556 movdqu 16($ctx),$CDGH # HGFE
557 movdqa 0x200-0x80($Tbl),$TMP # byte swap mask
559 pshufd \$0x1b,$ABEF,$Wi # ABCD
560 pshufd \$0xb1,$ABEF,$ABEF # CDAB
561 pshufd \$0x1b,$CDGH,$CDGH # EFGH
562 movdqa $TMP,$BSWAP # offload
563 palignr \$8,$CDGH,$ABEF # ABEF
564 punpcklqdq $Wi,$CDGH # CDGH
569 movdqu ($inp),@MSG[0]
570 movdqu 0x10($inp),@MSG[1]
571 movdqu 0x20($inp),@MSG[2]
573 movdqu 0x30($inp),@MSG[3]
575 movdqa 0*32-0x80($Tbl),$Wi
578 movdqa $CDGH,$CDGH_SAVE # offload
579 sha256rnds2 $ABEF,$CDGH # 0-3
580 pshufd \$0x0e,$Wi,$Wi
582 movdqa $ABEF,$ABEF_SAVE # offload
583 sha256rnds2 $CDGH,$ABEF
585 movdqa 1*32-0x80($Tbl),$Wi
588 sha256rnds2 $ABEF,$CDGH # 4-7
589 pshufd \$0x0e,$Wi,$Wi
591 sha256msg1 @MSG[1],@MSG[0]
592 sha256rnds2 $CDGH,$ABEF
594 movdqa 2*32-0x80($Tbl),$Wi
597 sha256rnds2 $ABEF,$CDGH # 8-11
598 pshufd \$0x0e,$Wi,$Wi
600 palignr \$4,@MSG[2],$TMP
603 sha256msg1 @MSG[2],@MSG[1]
604 sha256rnds2 $CDGH,$ABEF
606 movdqa 3*32-0x80($Tbl),$Wi
608 sha256msg2 @MSG[3],@MSG[0]
609 sha256rnds2 $ABEF,$CDGH # 12-15
610 pshufd \$0x0e,$Wi,$Wi
612 palignr \$4,@MSG[3],$TMP
615 sha256msg1 @MSG[3],@MSG[2]
616 sha256rnds2 $CDGH,$ABEF
618 for($i=4;$i<16-3;$i++) {
620 movdqa $i*32-0x80($Tbl),$Wi
622 sha256msg2 @MSG[0],@MSG[1]
623 sha256rnds2 $ABEF,$CDGH # 16-19...
624 pshufd \$0x0e,$Wi,$Wi
626 palignr \$4,@MSG[0],$TMP
629 sha256msg1 @MSG[0],@MSG[3]
630 sha256rnds2 $CDGH,$ABEF
632 push(@MSG,shift(@MSG));
635 movdqa 13*32-0x80($Tbl),$Wi
637 sha256msg2 @MSG[0],@MSG[1]
638 sha256rnds2 $ABEF,$CDGH # 52-55
639 pshufd \$0x0e,$Wi,$Wi
641 palignr \$4,@MSG[0],$TMP
642 sha256rnds2 $CDGH,$ABEF
645 movdqa 14*32-0x80($Tbl),$Wi
647 sha256rnds2 $ABEF,$CDGH # 56-59
648 pshufd \$0x0e,$Wi,$Wi
649 sha256msg2 @MSG[1],@MSG[2]
651 sha256rnds2 $CDGH,$ABEF
653 movdqa 15*32-0x80($Tbl),$Wi
656 sha256rnds2 $ABEF,$CDGH # 60-63
657 pshufd \$0x0e,$Wi,$Wi
660 sha256rnds2 $CDGH,$ABEF
662 paddd $CDGH_SAVE,$CDGH
663 paddd $ABEF_SAVE,$ABEF
666 pshufd \$0xb1,$CDGH,$CDGH # DCHG
667 pshufd \$0x1b,$ABEF,$TMP # FEBA
668 pshufd \$0xb1,$ABEF,$ABEF # BAFE
669 punpckhqdq $CDGH,$ABEF # DCBA
670 palignr \$8,$TMP,$CDGH # HGFE
673 movdqu $CDGH,16($ctx)
675 $code.=<<___ if ($win64);
676 movaps -8-5*16(%rax),%xmm6
677 movaps -8-4*16(%rax),%xmm7
678 movaps -8-3*16(%rax),%xmm8
679 movaps -8-2*16(%rax),%xmm9
680 movaps -8-1*16(%rax),%xmm10
686 .size sha256_block_data_order_shaext,.-sha256_block_data_order_shaext
692 my ($a,$b,$c,$d,$e,$f,$g,$h);
694 sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm
695 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
697 $arg = "\$$arg" if ($arg*1 eq $arg);
698 $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
703 '($a,$b,$c,$d,$e,$f,$g,$h)=@ROT;'.
705 '&ror ($a0,$Sigma1[2]-$Sigma1[1])',
709 '&ror ($a1,$Sigma0[2]-$Sigma0[1])',
711 '&xor ($a4,$g)', # f^g
713 '&ror ($a0,$Sigma1[1]-$Sigma1[0])',
715 '&and ($a4,$e)', # (f^g)&e
718 '&add ($h,$SZ*($i&15)."(%rsp)")', # h+=X[i]+K[i]
721 '&xor ($a4,$g)', # Ch(e,f,g)=((f^g)&e)^g
722 '&ror ($a1,$Sigma0[1]-$Sigma0[0])',
723 '&xor ($a2,$b)', # a^b, b^c in next round
725 '&add ($h,$a4)', # h+=Ch(e,f,g)
726 '&ror ($a0,$Sigma1[0])', # Sigma1(e)
727 '&and ($a3,$a2)', # (b^c)&(a^b)
730 '&add ($h,$a0)', # h+=Sigma1(e)
731 '&xor ($a3,$b)', # Maj(a,b,c)=Ch(a^b,c,b)
733 '&ror ($a1,$Sigma0[0])', # Sigma0(a)
734 '&add ($d,$h)', # d+=h
735 '&add ($h,$a3)', # h+=Maj(a,b,c)
738 '&add ($a1,$h);'. # h+=Sigma0(a)
739 '($a2,$a3) = ($a3,$a2); unshift(@ROT,pop(@ROT)); $i++;'
743 ######################################################################
746 if ($SZ==4) { # SHA256 only
747 my @X = map("%xmm$_",(0..3));
748 my ($t0,$t1,$t2,$t3, $t4,$t5) = map("%xmm$_",(4..9));
751 .type ${func}_ssse3,\@function,3
761 mov %rsp,%r11 # copy %rsp
762 shl \$4,%rdx # num*16
763 sub \$`$framesz+$win64*16*4`,%rsp
764 lea ($inp,%rdx,$SZ),%rdx # inp+num*16*$SZ
765 and \$-64,%rsp # align stack frame
766 mov $ctx,$_ctx # save ctx, 1st arg
767 mov $inp,$_inp # save inp, 2nd arh
768 mov %rdx,$_end # save end pointer, "3rd" arg
769 mov %r11,$_rsp # save copy of %rsp
771 $code.=<<___ if ($win64);
772 movaps %xmm6,16*$SZ+32(%rsp)
773 movaps %xmm7,16*$SZ+48(%rsp)
774 movaps %xmm8,16*$SZ+64(%rsp)
775 movaps %xmm9,16*$SZ+80(%rsp)
791 #movdqa $TABLE+`$SZ*2*$rounds`+32(%rip),$t4
792 #movdqa $TABLE+`$SZ*2*$rounds`+64(%rip),$t5
796 movdqa $TABLE+`$SZ*2*$rounds`(%rip),$t3
797 movdqu 0x00($inp),@X[0]
798 movdqu 0x10($inp),@X[1]
799 movdqu 0x20($inp),@X[2]
801 movdqu 0x30($inp),@X[3]
802 lea $TABLE(%rip),$Tbl
804 movdqa 0x00($Tbl),$t0
805 movdqa 0x20($Tbl),$t1
808 movdqa 0x40($Tbl),$t2
810 movdqa 0x60($Tbl),$t3
814 movdqa $t0,0x00(%rsp)
816 movdqa $t1,0x10(%rsp)
818 movdqa $t2,0x20(%rsp)
820 movdqa $t3,0x30(%rsp)
826 sub \$`-16*2*$SZ`,$Tbl # size optimization
828 sub Xupdate_256_SSSE3 () {
830 '&movdqa ($t0,@X[1]);',
831 '&movdqa ($t3,@X[3])',
832 '&palignr ($t0,@X[0],$SZ)', # X[1..4]
833 '&palignr ($t3,@X[2],$SZ);', # X[9..12]
835 '&movdqa ($t2,$t0);',
836 '&psrld ($t0,$sigma0[2])',
837 '&paddd (@X[0],$t3);', # X[0..3] += X[9..12]
838 '&psrld ($t2,$sigma0[0])',
839 '&pshufd ($t3,@X[3],0b11111010)',# X[14..15]
840 '&pslld ($t1,8*$SZ-$sigma0[1]);'.
842 '&psrld ($t2,$sigma0[1]-$sigma0[0]);'.
844 '&pslld ($t1,$sigma0[1]-$sigma0[0]);'.
847 '&pxor ($t0,$t1);', # sigma0(X[1..4])
848 '&psrld ($t3,$sigma1[2])',
849 '&paddd (@X[0],$t0);', # X[0..3] += sigma0(X[1..4])
850 '&psrlq ($t2,$sigma1[0])',
852 '&psrlq ($t2,$sigma1[1]-$sigma1[0])',
854 '&pshufb ($t3,$t4)', # sigma1(X[14..15])
855 '&paddd (@X[0],$t3)', # X[0..1] += sigma1(X[14..15])
856 '&pshufd ($t3,@X[0],0b01010000)',# X[16..17]
857 '&movdqa ($t2,$t3);',
858 '&psrld ($t3,$sigma1[2])',
859 '&psrlq ($t2,$sigma1[0])',
861 '&psrlq ($t2,$sigma1[1]-$sigma1[0])',
863 '&movdqa ($t2,16*2*$j."($Tbl)")',
865 '&paddd (@X[0],$t3)' # X[2..3] += sigma1(X[16..17])
869 sub SSSE3_256_00_47 () {
873 my @insns = (&$body,&$body,&$body,&$body); # 104 instructions
876 foreach (Xupdate_256_SSSE3()) { # 36 instructions
882 } else { # squeeze extra 4% on Westmere and 19% on Atom
883 eval(shift(@insns)); #@
888 eval(shift(@insns)); #@
891 eval(shift(@insns)); #@
893 &palignr ($t0,@X[0],$SZ); # X[1..4]
896 &palignr ($t3,@X[2],$SZ); # X[9..12]
900 eval(shift(@insns)); #@
905 eval(shift(@insns)); #@
907 &psrld ($t0,$sigma0[2]);
911 &paddd (@X[0],$t3); # X[0..3] += X[9..12]
912 eval(shift(@insns)); #@
914 &psrld ($t2,$sigma0[0]);
917 &pshufd ($t3,@X[3],0b11111010); # X[4..15]
919 eval(shift(@insns)); #@
920 &pslld ($t1,8*$SZ-$sigma0[1]);
924 eval(shift(@insns)); #@
927 eval(shift(@insns)); #@
928 &psrld ($t2,$sigma0[1]-$sigma0[0]);
933 &pslld ($t1,$sigma0[1]-$sigma0[0]);
938 eval(shift(@insns)); #@
942 &pxor ($t0,$t1); # sigma0(X[1..4])
943 eval(shift(@insns)); #@
946 &psrld ($t3,$sigma1[2]);
949 &paddd (@X[0],$t0); # X[0..3] += sigma0(X[1..4])
950 eval(shift(@insns)); #@
952 &psrlq ($t2,$sigma1[0]);
957 eval(shift(@insns)); #@
960 eval(shift(@insns)); #@
961 &psrlq ($t2,$sigma1[1]-$sigma1[0]);
965 eval(shift(@insns)); #@
968 #&pshufb ($t3,$t4); # sigma1(X[14..15])
969 &pshufd ($t3,$t3,0b10000000);
975 eval(shift(@insns)); #@
978 eval(shift(@insns)); #@
979 &paddd (@X[0],$t3); # X[0..1] += sigma1(X[14..15])
983 &pshufd ($t3,@X[0],0b01010000); # X[16..17]
985 eval(shift(@insns)); #@
990 &psrld ($t3,$sigma1[2]);
992 eval(shift(@insns)); #@
993 &psrlq ($t2,$sigma1[0]);
997 eval(shift(@insns)); #@
1000 eval(shift(@insns)); #@
1001 eval(shift(@insns));
1002 &psrlq ($t2,$sigma1[1]-$sigma1[0]);
1003 eval(shift(@insns));
1004 eval(shift(@insns));
1005 eval(shift(@insns));
1007 eval(shift(@insns));
1008 eval(shift(@insns));
1009 eval(shift(@insns)); #@
1011 &pshufd ($t3,$t3,0b00001000);
1012 eval(shift(@insns));
1013 eval(shift(@insns));
1014 &movdqa ($t2,16*2*$j."($Tbl)");
1015 eval(shift(@insns)); #@
1016 eval(shift(@insns));
1018 eval(shift(@insns));
1019 eval(shift(@insns));
1020 eval(shift(@insns));
1021 &paddd (@X[0],$t3); # X[2..3] += sigma1(X[16..17])
1022 eval(shift(@insns)); #@
1023 eval(shift(@insns));
1024 eval(shift(@insns));
1027 foreach (@insns) { eval; } # remaining instructions
1028 &movdqa (16*$j."(%rsp)",$t2);
1031 for ($i=0,$j=0; $j<4; $j++) {
1032 &SSSE3_256_00_47($j,\&body_00_15,@X);
1033 push(@X,shift(@X)); # rotate(@X)
1035 &cmpb ($SZ-1+16*2*$SZ."($Tbl)",0);
1036 &jne (".Lssse3_00_47");
1038 for ($i=0; $i<16; ) {
1039 foreach(body_00_15()) { eval; }
1046 lea 16*$SZ($inp),$inp
1069 $code.=<<___ if ($win64);
1070 movaps 16*$SZ+32(%rsp),%xmm6
1071 movaps 16*$SZ+48(%rsp),%xmm7
1072 movaps 16*$SZ+64(%rsp),%xmm8
1073 movaps 16*$SZ+80(%rsp),%xmm9
1085 .size ${func}_ssse3,.-${func}_ssse3
1090 ######################################################################
1093 if ($SZ==8) { # SHA512 only
1095 .type ${func}_xop,\@function,3
1105 mov %rsp,%r11 # copy %rsp
1106 shl \$4,%rdx # num*16
1107 sub \$`$framesz+$win64*16*($SZ==4?4:6)`,%rsp
1108 lea ($inp,%rdx,$SZ),%rdx # inp+num*16*$SZ
1109 and \$-64,%rsp # align stack frame
1110 mov $ctx,$_ctx # save ctx, 1st arg
1111 mov $inp,$_inp # save inp, 2nd arh
1112 mov %rdx,$_end # save end pointer, "3rd" arg
1113 mov %r11,$_rsp # save copy of %rsp
1115 $code.=<<___ if ($win64);
1116 movaps %xmm6,16*$SZ+32(%rsp)
1117 movaps %xmm7,16*$SZ+48(%rsp)
1118 movaps %xmm8,16*$SZ+64(%rsp)
1119 movaps %xmm9,16*$SZ+80(%rsp)
1121 $code.=<<___ if ($win64 && $SZ>4);
1122 movaps %xmm10,16*$SZ+96(%rsp)
1123 movaps %xmm11,16*$SZ+112(%rsp)
1139 if ($SZ==4) { # SHA256
1140 my @X = map("%xmm$_",(0..3));
1141 my ($t0,$t1,$t2,$t3) = map("%xmm$_",(4..7));
1146 vmovdqa $TABLE+`$SZ*2*$rounds`(%rip),$t3
1147 vmovdqu 0x00($inp),@X[0]
1148 vmovdqu 0x10($inp),@X[1]
1149 vmovdqu 0x20($inp),@X[2]
1150 vmovdqu 0x30($inp),@X[3]
1151 vpshufb $t3,@X[0],@X[0]
1152 lea $TABLE(%rip),$Tbl
1153 vpshufb $t3,@X[1],@X[1]
1154 vpshufb $t3,@X[2],@X[2]
1155 vpaddd 0x00($Tbl),@X[0],$t0
1156 vpshufb $t3,@X[3],@X[3]
1157 vpaddd 0x20($Tbl),@X[1],$t1
1158 vpaddd 0x40($Tbl),@X[2],$t2
1159 vpaddd 0x60($Tbl),@X[3],$t3
1160 vmovdqa $t0,0x00(%rsp)
1162 vmovdqa $t1,0x10(%rsp)
1164 vmovdqa $t2,0x20(%rsp)
1166 vmovdqa $t3,0x30(%rsp)
1172 sub \$`-16*2*$SZ`,$Tbl # size optimization
1174 sub XOP_256_00_47 () {
1178 my @insns = (&$body,&$body,&$body,&$body); # 104 instructions
1180 &vpalignr ($t0,@X[1],@X[0],$SZ); # X[1..4]
1181 eval(shift(@insns));
1182 eval(shift(@insns));
1183 &vpalignr ($t3,@X[3],@X[2],$SZ); # X[9..12]
1184 eval(shift(@insns));
1185 eval(shift(@insns));
1186 &vprotd ($t1,$t0,8*$SZ-$sigma0[1]);
1187 eval(shift(@insns));
1188 eval(shift(@insns));
1189 &vpsrld ($t0,$t0,$sigma0[2]);
1190 eval(shift(@insns));
1191 eval(shift(@insns));
1192 &vpaddd (@X[0],@X[0],$t3); # X[0..3] += X[9..12]
1193 eval(shift(@insns));
1194 eval(shift(@insns));
1195 eval(shift(@insns));
1196 eval(shift(@insns));
1197 &vprotd ($t2,$t1,$sigma0[1]-$sigma0[0]);
1198 eval(shift(@insns));
1199 eval(shift(@insns));
1200 &vpxor ($t0,$t0,$t1);
1201 eval(shift(@insns));
1202 eval(shift(@insns));
1203 eval(shift(@insns));
1204 eval(shift(@insns));
1205 &vprotd ($t3,@X[3],8*$SZ-$sigma1[1]);
1206 eval(shift(@insns));
1207 eval(shift(@insns));
1208 &vpxor ($t0,$t0,$t2); # sigma0(X[1..4])
1209 eval(shift(@insns));
1210 eval(shift(@insns));
1211 &vpsrld ($t2,@X[3],$sigma1[2]);
1212 eval(shift(@insns));
1213 eval(shift(@insns));
1214 &vpaddd (@X[0],@X[0],$t0); # X[0..3] += sigma0(X[1..4])
1215 eval(shift(@insns));
1216 eval(shift(@insns));
1217 &vprotd ($t1,$t3,$sigma1[1]-$sigma1[0]);
1218 eval(shift(@insns));
1219 eval(shift(@insns));
1220 &vpxor ($t3,$t3,$t2);
1221 eval(shift(@insns));
1222 eval(shift(@insns));
1223 eval(shift(@insns));
1224 eval(shift(@insns));
1225 &vpxor ($t3,$t3,$t1); # sigma1(X[14..15])
1226 eval(shift(@insns));
1227 eval(shift(@insns));
1228 eval(shift(@insns));
1229 eval(shift(@insns));
1230 &vpsrldq ($t3,$t3,8);
1231 eval(shift(@insns));
1232 eval(shift(@insns));
1233 eval(shift(@insns));
1234 eval(shift(@insns));
1235 &vpaddd (@X[0],@X[0],$t3); # X[0..1] += sigma1(X[14..15])
1236 eval(shift(@insns));
1237 eval(shift(@insns));
1238 eval(shift(@insns));
1239 eval(shift(@insns));
1240 &vprotd ($t3,@X[0],8*$SZ-$sigma1[1]);
1241 eval(shift(@insns));
1242 eval(shift(@insns));
1243 &vpsrld ($t2,@X[0],$sigma1[2]);
1244 eval(shift(@insns));
1245 eval(shift(@insns));
1246 &vprotd ($t1,$t3,$sigma1[1]-$sigma1[0]);
1247 eval(shift(@insns));
1248 eval(shift(@insns));
1249 &vpxor ($t3,$t3,$t2);
1250 eval(shift(@insns));
1251 eval(shift(@insns));
1252 eval(shift(@insns));
1253 eval(shift(@insns));
1254 &vpxor ($t3,$t3,$t1); # sigma1(X[16..17])
1255 eval(shift(@insns));
1256 eval(shift(@insns));
1257 eval(shift(@insns));
1258 eval(shift(@insns));
1259 &vpslldq ($t3,$t3,8); # 22 instructions
1260 eval(shift(@insns));
1261 eval(shift(@insns));
1262 eval(shift(@insns));
1263 eval(shift(@insns));
1264 &vpaddd (@X[0],@X[0],$t3); # X[2..3] += sigma1(X[16..17])
1265 eval(shift(@insns));
1266 eval(shift(@insns));
1267 eval(shift(@insns));
1268 eval(shift(@insns));
1269 &vpaddd ($t2,@X[0],16*2*$j."($Tbl)");
1270 foreach (@insns) { eval; } # remaining instructions
1271 &vmovdqa (16*$j."(%rsp)",$t2);
1274 for ($i=0,$j=0; $j<4; $j++) {
1275 &XOP_256_00_47($j,\&body_00_15,@X);
1276 push(@X,shift(@X)); # rotate(@X)
1278 &cmpb ($SZ-1+16*2*$SZ."($Tbl)",0);
1279 &jne (".Lxop_00_47");
1281 for ($i=0; $i<16; ) {
1282 foreach(body_00_15()) { eval; }
1286 my @X = map("%xmm$_",(0..7));
1287 my ($t0,$t1,$t2,$t3) = map("%xmm$_",(8..11));
1292 vmovdqa $TABLE+`$SZ*2*$rounds`(%rip),$t3
1293 vmovdqu 0x00($inp),@X[0]
1294 lea $TABLE+0x80(%rip),$Tbl # size optimization
1295 vmovdqu 0x10($inp),@X[1]
1296 vmovdqu 0x20($inp),@X[2]
1297 vpshufb $t3,@X[0],@X[0]
1298 vmovdqu 0x30($inp),@X[3]
1299 vpshufb $t3,@X[1],@X[1]
1300 vmovdqu 0x40($inp),@X[4]
1301 vpshufb $t3,@X[2],@X[2]
1302 vmovdqu 0x50($inp),@X[5]
1303 vpshufb $t3,@X[3],@X[3]
1304 vmovdqu 0x60($inp),@X[6]
1305 vpshufb $t3,@X[4],@X[4]
1306 vmovdqu 0x70($inp),@X[7]
1307 vpshufb $t3,@X[5],@X[5]
1308 vpaddq -0x80($Tbl),@X[0],$t0
1309 vpshufb $t3,@X[6],@X[6]
1310 vpaddq -0x60($Tbl),@X[1],$t1
1311 vpshufb $t3,@X[7],@X[7]
1312 vpaddq -0x40($Tbl),@X[2],$t2
1313 vpaddq -0x20($Tbl),@X[3],$t3
1314 vmovdqa $t0,0x00(%rsp)
1315 vpaddq 0x00($Tbl),@X[4],$t0
1316 vmovdqa $t1,0x10(%rsp)
1317 vpaddq 0x20($Tbl),@X[5],$t1
1318 vmovdqa $t2,0x20(%rsp)
1319 vpaddq 0x40($Tbl),@X[6],$t2
1320 vmovdqa $t3,0x30(%rsp)
1321 vpaddq 0x60($Tbl),@X[7],$t3
1322 vmovdqa $t0,0x40(%rsp)
1324 vmovdqa $t1,0x50(%rsp)
1326 vmovdqa $t2,0x60(%rsp)
1328 vmovdqa $t3,0x70(%rsp)
1334 add \$`16*2*$SZ`,$Tbl
1336 sub XOP_512_00_47 () {
1340 my @insns = (&$body,&$body); # 52 instructions
1342 &vpalignr ($t0,@X[1],@X[0],$SZ); # X[1..2]
1343 eval(shift(@insns));
1344 eval(shift(@insns));
1345 &vpalignr ($t3,@X[5],@X[4],$SZ); # X[9..10]
1346 eval(shift(@insns));
1347 eval(shift(@insns));
1348 &vprotq ($t1,$t0,8*$SZ-$sigma0[1]);
1349 eval(shift(@insns));
1350 eval(shift(@insns));
1351 &vpsrlq ($t0,$t0,$sigma0[2]);
1352 eval(shift(@insns));
1353 eval(shift(@insns));
1354 &vpaddq (@X[0],@X[0],$t3); # X[0..1] += X[9..10]
1355 eval(shift(@insns));
1356 eval(shift(@insns));
1357 eval(shift(@insns));
1358 eval(shift(@insns));
1359 &vprotq ($t2,$t1,$sigma0[1]-$sigma0[0]);
1360 eval(shift(@insns));
1361 eval(shift(@insns));
1362 &vpxor ($t0,$t0,$t1);
1363 eval(shift(@insns));
1364 eval(shift(@insns));
1365 eval(shift(@insns));
1366 eval(shift(@insns));
1367 &vprotq ($t3,@X[7],8*$SZ-$sigma1[1]);
1368 eval(shift(@insns));
1369 eval(shift(@insns));
1370 &vpxor ($t0,$t0,$t2); # sigma0(X[1..2])
1371 eval(shift(@insns));
1372 eval(shift(@insns));
1373 &vpsrlq ($t2,@X[7],$sigma1[2]);
1374 eval(shift(@insns));
1375 eval(shift(@insns));
1376 &vpaddq (@X[0],@X[0],$t0); # X[0..1] += sigma0(X[1..2])
1377 eval(shift(@insns));
1378 eval(shift(@insns));
1379 &vprotq ($t1,$t3,$sigma1[1]-$sigma1[0]);
1380 eval(shift(@insns));
1381 eval(shift(@insns));
1382 &vpxor ($t3,$t3,$t2);
1383 eval(shift(@insns));
1384 eval(shift(@insns));
1385 eval(shift(@insns));
1386 eval(shift(@insns));
1387 &vpxor ($t3,$t3,$t1); # sigma1(X[14..15])
1388 eval(shift(@insns));
1389 eval(shift(@insns));
1390 eval(shift(@insns));
1391 eval(shift(@insns));
1392 &vpaddq (@X[0],@X[0],$t3); # X[0..1] += sigma1(X[14..15])
1393 eval(shift(@insns));
1394 eval(shift(@insns));
1395 eval(shift(@insns));
1396 eval(shift(@insns));
1397 &vpaddq ($t2,@X[0],16*2*$j-0x80."($Tbl)");
1398 foreach (@insns) { eval; } # remaining instructions
1399 &vmovdqa (16*$j."(%rsp)",$t2);
1402 for ($i=0,$j=0; $j<8; $j++) {
1403 &XOP_512_00_47($j,\&body_00_15,@X);
1404 push(@X,shift(@X)); # rotate(@X)
1406 &cmpb ($SZ-1+16*2*$SZ-0x80."($Tbl)",0);
1407 &jne (".Lxop_00_47");
1409 for ($i=0; $i<16; ) {
1410 foreach(body_00_15()) { eval; }
1418 lea 16*$SZ($inp),$inp
1442 $code.=<<___ if ($win64);
1443 movaps 16*$SZ+32(%rsp),%xmm6
1444 movaps 16*$SZ+48(%rsp),%xmm7
1445 movaps 16*$SZ+64(%rsp),%xmm8
1446 movaps 16*$SZ+80(%rsp),%xmm9
1448 $code.=<<___ if ($win64 && $SZ>4);
1449 movaps 16*$SZ+96(%rsp),%xmm10
1450 movaps 16*$SZ+112(%rsp),%xmm11
1462 .size ${func}_xop,.-${func}_xop
1465 ######################################################################
1466 # AVX+shrd code path
1468 local *ror = sub { &shrd(@_[0],@_) };
1471 .type ${func}_avx,\@function,3
1481 mov %rsp,%r11 # copy %rsp
1482 shl \$4,%rdx # num*16
1483 sub \$`$framesz+$win64*16*($SZ==4?4:6)`,%rsp
1484 lea ($inp,%rdx,$SZ),%rdx # inp+num*16*$SZ
1485 and \$-64,%rsp # align stack frame
1486 mov $ctx,$_ctx # save ctx, 1st arg
1487 mov $inp,$_inp # save inp, 2nd arh
1488 mov %rdx,$_end # save end pointer, "3rd" arg
1489 mov %r11,$_rsp # save copy of %rsp
1491 $code.=<<___ if ($win64);
1492 movaps %xmm6,16*$SZ+32(%rsp)
1493 movaps %xmm7,16*$SZ+48(%rsp)
1494 movaps %xmm8,16*$SZ+64(%rsp)
1495 movaps %xmm9,16*$SZ+80(%rsp)
1497 $code.=<<___ if ($win64 && $SZ>4);
1498 movaps %xmm10,16*$SZ+96(%rsp)
1499 movaps %xmm11,16*$SZ+112(%rsp)
1514 if ($SZ==4) { # SHA256
1515 my @X = map("%xmm$_",(0..3));
1516 my ($t0,$t1,$t2,$t3, $t4,$t5) = map("%xmm$_",(4..9));
1519 vmovdqa $TABLE+`$SZ*2*$rounds`+32(%rip),$t4
1520 vmovdqa $TABLE+`$SZ*2*$rounds`+64(%rip),$t5
1524 vmovdqa $TABLE+`$SZ*2*$rounds`(%rip),$t3
1525 vmovdqu 0x00($inp),@X[0]
1526 vmovdqu 0x10($inp),@X[1]
1527 vmovdqu 0x20($inp),@X[2]
1528 vmovdqu 0x30($inp),@X[3]
1529 vpshufb $t3,@X[0],@X[0]
1530 lea $TABLE(%rip),$Tbl
1531 vpshufb $t3,@X[1],@X[1]
1532 vpshufb $t3,@X[2],@X[2]
1533 vpaddd 0x00($Tbl),@X[0],$t0
1534 vpshufb $t3,@X[3],@X[3]
1535 vpaddd 0x20($Tbl),@X[1],$t1
1536 vpaddd 0x40($Tbl),@X[2],$t2
1537 vpaddd 0x60($Tbl),@X[3],$t3
1538 vmovdqa $t0,0x00(%rsp)
1540 vmovdqa $t1,0x10(%rsp)
1542 vmovdqa $t2,0x20(%rsp)
1544 vmovdqa $t3,0x30(%rsp)
1550 sub \$`-16*2*$SZ`,$Tbl # size optimization
1552 sub Xupdate_256_AVX () {
1554 '&vpalignr ($t0,@X[1],@X[0],$SZ)', # X[1..4]
1555 '&vpalignr ($t3,@X[3],@X[2],$SZ)', # X[9..12]
1556 '&vpsrld ($t2,$t0,$sigma0[0]);',
1557 '&vpaddd (@X[0],@X[0],$t3)', # X[0..3] += X[9..12]
1558 '&vpsrld ($t3,$t0,$sigma0[2])',
1559 '&vpslld ($t1,$t0,8*$SZ-$sigma0[1]);',
1560 '&vpxor ($t0,$t3,$t2)',
1561 '&vpshufd ($t3,@X[3],0b11111010)',# X[14..15]
1562 '&vpsrld ($t2,$t2,$sigma0[1]-$sigma0[0]);',
1563 '&vpxor ($t0,$t0,$t1)',
1564 '&vpslld ($t1,$t1,$sigma0[1]-$sigma0[0]);',
1565 '&vpxor ($t0,$t0,$t2)',
1566 '&vpsrld ($t2,$t3,$sigma1[2]);',
1567 '&vpxor ($t0,$t0,$t1)', # sigma0(X[1..4])
1568 '&vpsrlq ($t3,$t3,$sigma1[0]);',
1569 '&vpaddd (@X[0],@X[0],$t0)', # X[0..3] += sigma0(X[1..4])
1570 '&vpxor ($t2,$t2,$t3);',
1571 '&vpsrlq ($t3,$t3,$sigma1[1]-$sigma1[0])',
1572 '&vpxor ($t2,$t2,$t3)',
1573 '&vpshufb ($t2,$t2,$t4)', # sigma1(X[14..15])
1574 '&vpaddd (@X[0],@X[0],$t2)', # X[0..1] += sigma1(X[14..15])
1575 '&vpshufd ($t3,@X[0],0b01010000)',# X[16..17]
1576 '&vpsrld ($t2,$t3,$sigma1[2])',
1577 '&vpsrlq ($t3,$t3,$sigma1[0])',
1578 '&vpxor ($t2,$t2,$t3);',
1579 '&vpsrlq ($t3,$t3,$sigma1[1]-$sigma1[0])',
1580 '&vpxor ($t2,$t2,$t3)',
1581 '&vpshufb ($t2,$t2,$t5)',
1582 '&vpaddd (@X[0],@X[0],$t2)' # X[2..3] += sigma1(X[16..17])
1586 sub AVX_256_00_47 () {
1590 my @insns = (&$body,&$body,&$body,&$body); # 104 instructions
1592 foreach (Xupdate_256_AVX()) { # 29 instructions
1594 eval(shift(@insns));
1595 eval(shift(@insns));
1596 eval(shift(@insns));
1598 &vpaddd ($t2,@X[0],16*2*$j."($Tbl)");
1599 foreach (@insns) { eval; } # remaining instructions
1600 &vmovdqa (16*$j."(%rsp)",$t2);
1603 for ($i=0,$j=0; $j<4; $j++) {
1604 &AVX_256_00_47($j,\&body_00_15,@X);
1605 push(@X,shift(@X)); # rotate(@X)
1607 &cmpb ($SZ-1+16*2*$SZ."($Tbl)",0);
1608 &jne (".Lavx_00_47");
1610 for ($i=0; $i<16; ) {
1611 foreach(body_00_15()) { eval; }
1615 my @X = map("%xmm$_",(0..7));
1616 my ($t0,$t1,$t2,$t3) = map("%xmm$_",(8..11));
1622 vmovdqa $TABLE+`$SZ*2*$rounds`(%rip),$t3
1623 vmovdqu 0x00($inp),@X[0]
1624 lea $TABLE+0x80(%rip),$Tbl # size optimization
1625 vmovdqu 0x10($inp),@X[1]
1626 vmovdqu 0x20($inp),@X[2]
1627 vpshufb $t3,@X[0],@X[0]
1628 vmovdqu 0x30($inp),@X[3]
1629 vpshufb $t3,@X[1],@X[1]
1630 vmovdqu 0x40($inp),@X[4]
1631 vpshufb $t3,@X[2],@X[2]
1632 vmovdqu 0x50($inp),@X[5]
1633 vpshufb $t3,@X[3],@X[3]
1634 vmovdqu 0x60($inp),@X[6]
1635 vpshufb $t3,@X[4],@X[4]
1636 vmovdqu 0x70($inp),@X[7]
1637 vpshufb $t3,@X[5],@X[5]
1638 vpaddq -0x80($Tbl),@X[0],$t0
1639 vpshufb $t3,@X[6],@X[6]
1640 vpaddq -0x60($Tbl),@X[1],$t1
1641 vpshufb $t3,@X[7],@X[7]
1642 vpaddq -0x40($Tbl),@X[2],$t2
1643 vpaddq -0x20($Tbl),@X[3],$t3
1644 vmovdqa $t0,0x00(%rsp)
1645 vpaddq 0x00($Tbl),@X[4],$t0
1646 vmovdqa $t1,0x10(%rsp)
1647 vpaddq 0x20($Tbl),@X[5],$t1
1648 vmovdqa $t2,0x20(%rsp)
1649 vpaddq 0x40($Tbl),@X[6],$t2
1650 vmovdqa $t3,0x30(%rsp)
1651 vpaddq 0x60($Tbl),@X[7],$t3
1652 vmovdqa $t0,0x40(%rsp)
1654 vmovdqa $t1,0x50(%rsp)
1656 vmovdqa $t2,0x60(%rsp)
1658 vmovdqa $t3,0x70(%rsp)
1664 add \$`16*2*$SZ`,$Tbl
1666 sub Xupdate_512_AVX () {
1668 '&vpalignr ($t0,@X[1],@X[0],$SZ)', # X[1..2]
1669 '&vpalignr ($t3,@X[5],@X[4],$SZ)', # X[9..10]
1670 '&vpsrlq ($t2,$t0,$sigma0[0])',
1671 '&vpaddq (@X[0],@X[0],$t3);', # X[0..1] += X[9..10]
1672 '&vpsrlq ($t3,$t0,$sigma0[2])',
1673 '&vpsllq ($t1,$t0,8*$SZ-$sigma0[1]);',
1674 '&vpxor ($t0,$t3,$t2)',
1675 '&vpsrlq ($t2,$t2,$sigma0[1]-$sigma0[0]);',
1676 '&vpxor ($t0,$t0,$t1)',
1677 '&vpsllq ($t1,$t1,$sigma0[1]-$sigma0[0]);',
1678 '&vpxor ($t0,$t0,$t2)',
1679 '&vpsrlq ($t3,@X[7],$sigma1[2]);',
1680 '&vpxor ($t0,$t0,$t1)', # sigma0(X[1..2])
1681 '&vpsllq ($t2,@X[7],8*$SZ-$sigma1[1]);',
1682 '&vpaddq (@X[0],@X[0],$t0)', # X[0..1] += sigma0(X[1..2])
1683 '&vpsrlq ($t1,@X[7],$sigma1[0]);',
1684 '&vpxor ($t3,$t3,$t2)',
1685 '&vpsllq ($t2,$t2,$sigma1[1]-$sigma1[0]);',
1686 '&vpxor ($t3,$t3,$t1)',
1687 '&vpsrlq ($t1,$t1,$sigma1[1]-$sigma1[0]);',
1688 '&vpxor ($t3,$t3,$t2)',
1689 '&vpxor ($t3,$t3,$t1)', # sigma1(X[14..15])
1690 '&vpaddq (@X[0],@X[0],$t3)', # X[0..1] += sigma1(X[14..15])
1694 sub AVX_512_00_47 () {
1698 my @insns = (&$body,&$body); # 52 instructions
1700 foreach (Xupdate_512_AVX()) { # 23 instructions
1702 eval(shift(@insns));
1703 eval(shift(@insns));
1705 &vpaddq ($t2,@X[0],16*2*$j-0x80."($Tbl)");
1706 foreach (@insns) { eval; } # remaining instructions
1707 &vmovdqa (16*$j."(%rsp)",$t2);
1710 for ($i=0,$j=0; $j<8; $j++) {
1711 &AVX_512_00_47($j,\&body_00_15,@X);
1712 push(@X,shift(@X)); # rotate(@X)
1714 &cmpb ($SZ-1+16*2*$SZ-0x80."($Tbl)",0);
1715 &jne (".Lavx_00_47");
1717 for ($i=0; $i<16; ) {
1718 foreach(body_00_15()) { eval; }
1726 lea 16*$SZ($inp),$inp
1750 $code.=<<___ if ($win64);
1751 movaps 16*$SZ+32(%rsp),%xmm6
1752 movaps 16*$SZ+48(%rsp),%xmm7
1753 movaps 16*$SZ+64(%rsp),%xmm8
1754 movaps 16*$SZ+80(%rsp),%xmm9
1756 $code.=<<___ if ($win64 && $SZ>4);
1757 movaps 16*$SZ+96(%rsp),%xmm10
1758 movaps 16*$SZ+112(%rsp),%xmm11
1770 .size ${func}_avx,.-${func}_avx
1774 ######################################################################
1775 # AVX2+BMI code path
1777 my $a5=$SZ==4?"%esi":"%rsi"; # zap $inp
1781 sub bodyx_00_15 () {
1782 # at start $a1 should be zero, $a3 - $b^$c and $a4 copy of $f
1784 '($a,$b,$c,$d,$e,$f,$g,$h)=@ROT;'.
1786 '&add ($h,(32*($i/(16/$SZ))+$SZ*($i%(16/$SZ)))%$PUSH8.$base)', # h+=X[i]+K[i]
1787 '&and ($a4,$e)', # f&e
1788 '&rorx ($a0,$e,$Sigma1[2])',
1789 '&rorx ($a2,$e,$Sigma1[1])',
1791 '&lea ($a,"($a,$a1)")', # h+=Sigma0(a) from the past
1792 '&lea ($h,"($h,$a4)")',
1793 '&andn ($a4,$e,$g)', # ~e&g
1796 '&rorx ($a1,$e,$Sigma1[0])',
1797 '&lea ($h,"($h,$a4)")', # h+=Ch(e,f,g)=(e&f)+(~e&g)
1798 '&xor ($a0,$a1)', # Sigma1(e)
1801 '&rorx ($a4,$a,$Sigma0[2])',
1802 '&lea ($h,"($h,$a0)")', # h+=Sigma1(e)
1803 '&xor ($a2,$b)', # a^b, b^c in next round
1804 '&rorx ($a1,$a,$Sigma0[1])',
1806 '&rorx ($a0,$a,$Sigma0[0])',
1807 '&lea ($d,"($d,$h)")', # d+=h
1808 '&and ($a3,$a2)', # (b^c)&(a^b)
1811 '&xor ($a3,$b)', # Maj(a,b,c)=Ch(a^b,c,b)
1812 '&xor ($a1,$a0)', # Sigma0(a)
1813 '&lea ($h,"($h,$a3)");'. # h+=Maj(a,b,c)
1814 '&mov ($a4,$e)', # copy of f in future
1816 '($a2,$a3) = ($a3,$a2); unshift(@ROT,pop(@ROT)); $i++;'
1818 # and at the finish one has to $a+=$a1
1822 .type ${func}_avx2,\@function,3
1832 mov %rsp,%r11 # copy %rsp
1833 sub \$`2*$SZ*$rounds+4*8+$win64*16*($SZ==4?4:6)`,%rsp
1834 shl \$4,%rdx # num*16
1835 and \$-256*$SZ,%rsp # align stack frame
1836 lea ($inp,%rdx,$SZ),%rdx # inp+num*16*$SZ
1837 add \$`2*$SZ*($rounds-8)`,%rsp
1838 mov $ctx,$_ctx # save ctx, 1st arg
1839 mov $inp,$_inp # save inp, 2nd arh
1840 mov %rdx,$_end # save end pointer, "3rd" arg
1841 mov %r11,$_rsp # save copy of %rsp
1843 $code.=<<___ if ($win64);
1844 movaps %xmm6,16*$SZ+32(%rsp)
1845 movaps %xmm7,16*$SZ+48(%rsp)
1846 movaps %xmm8,16*$SZ+64(%rsp)
1847 movaps %xmm9,16*$SZ+80(%rsp)
1849 $code.=<<___ if ($win64 && $SZ>4);
1850 movaps %xmm10,16*$SZ+96(%rsp)
1851 movaps %xmm11,16*$SZ+112(%rsp)
1857 sub \$-16*$SZ,$inp # inp++, size optimization
1859 mov $inp,%r12 # borrow $T1
1861 cmp %rdx,$inp # $_end
1863 cmove %rsp,%r12 # next block or random data
1870 if ($SZ==4) { # SHA256
1871 my @X = map("%ymm$_",(0..3));
1872 my ($t0,$t1,$t2,$t3, $t4,$t5) = map("%ymm$_",(4..9));
1875 vmovdqa $TABLE+`$SZ*2*$rounds`+32(%rip),$t4
1876 vmovdqa $TABLE+`$SZ*2*$rounds`+64(%rip),$t5
1880 vmovdqa $TABLE+`$SZ*2*$rounds`(%rip),$t3
1881 vmovdqu -16*$SZ+0($inp),%xmm0
1882 vmovdqu -16*$SZ+16($inp),%xmm1
1883 vmovdqu -16*$SZ+32($inp),%xmm2
1884 vmovdqu -16*$SZ+48($inp),%xmm3
1885 #mov $inp,$_inp # offload $inp
1886 vinserti128 \$1,(%r12),@X[0],@X[0]
1887 vinserti128 \$1,16(%r12),@X[1],@X[1]
1888 vpshufb $t3,@X[0],@X[0]
1889 vinserti128 \$1,32(%r12),@X[2],@X[2]
1890 vpshufb $t3,@X[1],@X[1]
1891 vinserti128 \$1,48(%r12),@X[3],@X[3]
1893 lea $TABLE(%rip),$Tbl
1894 vpshufb $t3,@X[2],@X[2]
1895 vpaddd 0x00($Tbl),@X[0],$t0
1896 vpshufb $t3,@X[3],@X[3]
1897 vpaddd 0x20($Tbl),@X[1],$t1
1898 vpaddd 0x40($Tbl),@X[2],$t2
1899 vpaddd 0x60($Tbl),@X[3],$t3
1900 vmovdqa $t0,0x00(%rsp)
1902 vmovdqa $t1,0x20(%rsp)
1903 lea -$PUSH8(%rsp),%rsp
1905 vmovdqa $t2,0x00(%rsp)
1907 vmovdqa $t3,0x20(%rsp)
1909 sub \$-16*2*$SZ,$Tbl # size optimization
1916 sub AVX2_256_00_47 () {
1920 my @insns = (&$body,&$body,&$body,&$body); # 96 instructions
1921 my $base = "+2*$PUSH8(%rsp)";
1923 &lea ("%rsp","-$PUSH8(%rsp)") if (($j%2)==0);
1924 foreach (Xupdate_256_AVX()) { # 29 instructions
1926 eval(shift(@insns));
1927 eval(shift(@insns));
1928 eval(shift(@insns));
1930 &vpaddd ($t2,@X[0],16*2*$j."($Tbl)");
1931 foreach (@insns) { eval; } # remaining instructions
1932 &vmovdqa ((32*$j)%$PUSH8."(%rsp)",$t2);
1935 for ($i=0,$j=0; $j<4; $j++) {
1936 &AVX2_256_00_47($j,\&bodyx_00_15,@X);
1937 push(@X,shift(@X)); # rotate(@X)
1939 &lea ($Tbl,16*2*$SZ."($Tbl)");
1940 &cmpb (($SZ-1)."($Tbl)",0);
1941 &jne (".Lavx2_00_47");
1943 for ($i=0; $i<16; ) {
1944 my $base=$i<8?"+$PUSH8(%rsp)":"(%rsp)";
1945 foreach(bodyx_00_15()) { eval; }
1948 my @X = map("%ymm$_",(0..7));
1949 my ($t0,$t1,$t2,$t3) = map("%ymm$_",(8..11));
1955 vmovdqu -16*$SZ($inp),%xmm0
1956 vmovdqu -16*$SZ+16($inp),%xmm1
1957 vmovdqu -16*$SZ+32($inp),%xmm2
1958 lea $TABLE+0x80(%rip),$Tbl # size optimization
1959 vmovdqu -16*$SZ+48($inp),%xmm3
1960 vmovdqu -16*$SZ+64($inp),%xmm4
1961 vmovdqu -16*$SZ+80($inp),%xmm5
1962 vmovdqu -16*$SZ+96($inp),%xmm6
1963 vmovdqu -16*$SZ+112($inp),%xmm7
1964 #mov $inp,$_inp # offload $inp
1965 vmovdqa `$SZ*2*$rounds-0x80`($Tbl),$t2
1966 vinserti128 \$1,(%r12),@X[0],@X[0]
1967 vinserti128 \$1,16(%r12),@X[1],@X[1]
1968 vpshufb $t2,@X[0],@X[0]
1969 vinserti128 \$1,32(%r12),@X[2],@X[2]
1970 vpshufb $t2,@X[1],@X[1]
1971 vinserti128 \$1,48(%r12),@X[3],@X[3]
1972 vpshufb $t2,@X[2],@X[2]
1973 vinserti128 \$1,64(%r12),@X[4],@X[4]
1974 vpshufb $t2,@X[3],@X[3]
1975 vinserti128 \$1,80(%r12),@X[5],@X[5]
1976 vpshufb $t2,@X[4],@X[4]
1977 vinserti128 \$1,96(%r12),@X[6],@X[6]
1978 vpshufb $t2,@X[5],@X[5]
1979 vinserti128 \$1,112(%r12),@X[7],@X[7]
1981 vpaddq -0x80($Tbl),@X[0],$t0
1982 vpshufb $t2,@X[6],@X[6]
1983 vpaddq -0x60($Tbl),@X[1],$t1
1984 vpshufb $t2,@X[7],@X[7]
1985 vpaddq -0x40($Tbl),@X[2],$t2
1986 vpaddq -0x20($Tbl),@X[3],$t3
1987 vmovdqa $t0,0x00(%rsp)
1988 vpaddq 0x00($Tbl),@X[4],$t0
1989 vmovdqa $t1,0x20(%rsp)
1990 vpaddq 0x20($Tbl),@X[5],$t1
1991 vmovdqa $t2,0x40(%rsp)
1992 vpaddq 0x40($Tbl),@X[6],$t2
1993 vmovdqa $t3,0x60(%rsp)
1994 lea -$PUSH8(%rsp),%rsp
1995 vpaddq 0x60($Tbl),@X[7],$t3
1996 vmovdqa $t0,0x00(%rsp)
1998 vmovdqa $t1,0x20(%rsp)
2000 vmovdqa $t2,0x40(%rsp)
2002 vmovdqa $t3,0x60(%rsp)
2011 sub AVX2_512_00_47 () {
2015 my @insns = (&$body,&$body); # 48 instructions
2016 my $base = "+2*$PUSH8(%rsp)";
2018 &lea ("%rsp","-$PUSH8(%rsp)") if (($j%4)==0);
2019 foreach (Xupdate_512_AVX()) { # 23 instructions
2022 eval(shift(@insns));
2023 eval(shift(@insns));
2024 eval(shift(@insns));
2027 &vpaddq ($t2,@X[0],16*2*$j-0x80."($Tbl)");
2028 foreach (@insns) { eval; } # remaining instructions
2029 &vmovdqa ((32*$j)%$PUSH8."(%rsp)",$t2);
2032 for ($i=0,$j=0; $j<8; $j++) {
2033 &AVX2_512_00_47($j,\&bodyx_00_15,@X);
2034 push(@X,shift(@X)); # rotate(@X)
2036 &lea ($Tbl,16*2*$SZ."($Tbl)");
2037 &cmpb (($SZ-1-0x80)."($Tbl)",0);
2038 &jne (".Lavx2_00_47");
2040 for ($i=0; $i<16; ) {
2041 my $base=$i<8?"+$PUSH8(%rsp)":"(%rsp)";
2042 foreach(bodyx_00_15()) { eval; }
2046 mov `2*$SZ*$rounds`(%rsp),$ctx # $_ctx
2048 #mov `2*$SZ*$rounds+8`(%rsp),$inp # $_inp
2049 lea `2*$SZ*($rounds-8)`(%rsp),$Tbl
2069 cmp `$PUSH8+2*8`($Tbl),$inp # $_end
2080 for ($i=0; $i<8; ) {
2081 my $base="+16($Tbl)";
2082 foreach(bodyx_00_15()) { eval; }
2085 lea -$PUSH8($Tbl),$Tbl
2089 mov `2*$SZ*$rounds`(%rsp),$ctx # $_ctx
2091 #mov `2*$SZ*$rounds+8`(%rsp),$inp # $_inp
2092 lea `2*$SZ*($rounds-8)`(%rsp),%rsp
2100 lea `2*16*$SZ`($inp),$inp # inp+=2
2107 cmove %rsp,%r12 # next block or stale data
2124 $code.=<<___ if ($win64);
2125 movaps 16*$SZ+32(%rsp),%xmm6
2126 movaps 16*$SZ+48(%rsp),%xmm7
2127 movaps 16*$SZ+64(%rsp),%xmm8
2128 movaps 16*$SZ+80(%rsp),%xmm9
2130 $code.=<<___ if ($win64 && $SZ>4);
2131 movaps 16*$SZ+96(%rsp),%xmm10
2132 movaps 16*$SZ+112(%rsp),%xmm11
2144 .size ${func}_avx2,.-${func}_avx2
2149 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
2150 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
2158 .extern __imp_RtlVirtualUnwind
2159 .type se_handler,\@abi-omnipotent
2173 mov 120($context),%rax # pull context->Rax
2174 mov 248($context),%rbx # pull context->Rip
2176 mov 8($disp),%rsi # disp->ImageBase
2177 mov 56($disp),%r11 # disp->HanderlData
2179 mov 0(%r11),%r10d # HandlerData[0]
2180 lea (%rsi,%r10),%r10 # prologue label
2181 cmp %r10,%rbx # context->Rip<prologue label
2184 mov 152($context),%rax # pull context->Rsp
2186 mov 4(%r11),%r10d # HandlerData[1]
2187 lea (%rsi,%r10),%r10 # epilogue label
2188 cmp %r10,%rbx # context->Rip>=epilogue label
2191 $code.=<<___ if ($avx>1);
2192 lea .Lavx2_shortcut(%rip),%r10
2193 cmp %r10,%rbx # context->Rip<avx2_shortcut
2197 add \$`2*$SZ*($rounds-8)`,%rax
2201 mov %rax,%rsi # put aside Rsp
2202 mov 16*$SZ+3*8(%rax),%rax # pull $_rsp
2211 mov %rbx,144($context) # restore context->Rbx
2212 mov %rbp,160($context) # restore context->Rbp
2213 mov %r12,216($context) # restore context->R12
2214 mov %r13,224($context) # restore context->R13
2215 mov %r14,232($context) # restore context->R14
2216 mov %r15,240($context) # restore context->R15
2218 lea .Lepilogue(%rip),%r10
2220 jb .Lin_prologue # non-AVX code
2222 lea 16*$SZ+4*8(%rsi),%rsi # Xmm6- save area
2223 lea 512($context),%rdi # &context.Xmm6
2224 mov \$`$SZ==4?8:12`,%ecx
2225 .long 0xa548f3fc # cld; rep movsq
2230 mov %rax,152($context) # restore context->Rsp
2231 mov %rsi,168($context) # restore context->Rsi
2232 mov %rdi,176($context) # restore context->Rdi
2234 mov 40($disp),%rdi # disp->ContextRecord
2235 mov $context,%rsi # context
2236 mov \$154,%ecx # sizeof(CONTEXT)
2237 .long 0xa548f3fc # cld; rep movsq
2240 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
2241 mov 8(%rsi),%rdx # arg2, disp->ImageBase
2242 mov 0(%rsi),%r8 # arg3, disp->ControlPc
2243 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
2244 mov 40(%rsi),%r10 # disp->ContextRecord
2245 lea 56(%rsi),%r11 # &disp->HandlerData
2246 lea 24(%rsi),%r12 # &disp->EstablisherFrame
2247 mov %r10,32(%rsp) # arg5
2248 mov %r11,40(%rsp) # arg6
2249 mov %r12,48(%rsp) # arg7
2250 mov %rcx,56(%rsp) # arg8, (NULL)
2251 call *__imp_RtlVirtualUnwind(%rip)
2253 mov \$1,%eax # ExceptionContinueSearch
2265 .size se_handler,.-se_handler
2268 $code.=<<___ if ($SZ==4 && $shaext);
2269 .type shaext_handler,\@abi-omnipotent
2283 mov 120($context),%rax # pull context->Rax
2284 mov 248($context),%rbx # pull context->Rip
2286 lea .Lprologue_shaext(%rip),%r10
2287 cmp %r10,%rbx # context->Rip<.Lprologue
2290 lea .Lepilogue_shaext(%rip),%r10
2291 cmp %r10,%rbx # context->Rip>=.Lepilogue
2294 lea -8-5*16(%rax),%rsi
2295 lea 512($context),%rdi # &context.Xmm6
2297 .long 0xa548f3fc # cld; rep movsq
2300 .size shaext_handler,.-shaext_handler
2306 .rva .LSEH_begin_$func
2307 .rva .LSEH_end_$func
2308 .rva .LSEH_info_$func
2310 $code.=<<___ if ($SZ==4 && $shaext);
2311 .rva .LSEH_begin_${func}_shaext
2312 .rva .LSEH_end_${func}_shaext
2313 .rva .LSEH_info_${func}_shaext
2315 $code.=<<___ if ($SZ==4);
2316 .rva .LSEH_begin_${func}_ssse3
2317 .rva .LSEH_end_${func}_ssse3
2318 .rva .LSEH_info_${func}_ssse3
2320 $code.=<<___ if ($avx && $SZ==8);
2321 .rva .LSEH_begin_${func}_xop
2322 .rva .LSEH_end_${func}_xop
2323 .rva .LSEH_info_${func}_xop
2325 $code.=<<___ if ($avx);
2326 .rva .LSEH_begin_${func}_avx
2327 .rva .LSEH_end_${func}_avx
2328 .rva .LSEH_info_${func}_avx
2330 $code.=<<___ if ($avx>1);
2331 .rva .LSEH_begin_${func}_avx2
2332 .rva .LSEH_end_${func}_avx2
2333 .rva .LSEH_info_${func}_avx2
2341 .rva .Lprologue,.Lepilogue # HandlerData[]
2343 $code.=<<___ if ($SZ==4 && $shaext);
2344 .LSEH_info_${func}_shaext:
2348 $code.=<<___ if ($SZ==4);
2349 .LSEH_info_${func}_ssse3:
2352 .rva .Lprologue_ssse3,.Lepilogue_ssse3 # HandlerData[]
2354 $code.=<<___ if ($avx && $SZ==8);
2355 .LSEH_info_${func}_xop:
2358 .rva .Lprologue_xop,.Lepilogue_xop # HandlerData[]
2360 $code.=<<___ if ($avx);
2361 .LSEH_info_${func}_avx:
2364 .rva .Lprologue_avx,.Lepilogue_avx # HandlerData[]
2366 $code.=<<___ if ($avx>1);
2367 .LSEH_info_${func}_avx2:
2370 .rva .Lprologue_avx2,.Lepilogue_avx2 # HandlerData[]
2377 "sha256rnds2" => 0xcb,
2378 "sha256msg1" => 0xcc,
2379 "sha256msg2" => 0xcd );
2381 if (defined($opcodelet{$instr}) && @_[0] =~ /%xmm([0-7]),\s*%xmm([0-7])/) {
2382 my @opcode=(0x0f,0x38);
2383 push @opcode,$opcodelet{$instr};
2384 push @opcode,0xc0|($1&7)|(($2&7)<<3); # ModR/M
2385 return ".byte\t".join(',',@opcode);
2387 return $instr."\t".@_[0];
2391 foreach (split("\n",$code)) {
2392 s/\`([^\`]*)\`/eval $1/geo;
2394 s/\b(sha256[^\s]*)\s+(.*)/sha256op38($1,$2)/geo;