2 # Copyright 2006-2016 The OpenSSL Project Authors. All Rights Reserved.
4 # Licensed under the OpenSSL license (the "License"). You may not use
5 # this file except in compliance with the License. You can obtain a copy
6 # in the file LICENSE in the source distribution or at
7 # https://www.openssl.org/source/license.html
10 # ====================================================================
11 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12 # project. The module is, however, dual licensed under OpenSSL and
13 # CRYPTOGAMS licenses depending on where you obtain it. For further
14 # details see http://www.openssl.org/~appro/cryptogams/.
15 # ====================================================================
17 # sha1_block procedure for x86_64.
19 # It was brought to my attention that on EM64T compiler-generated code
20 # was far behind 32-bit assembler implementation. This is unlike on
21 # Opteron where compiler-generated code was only 15% behind 32-bit
22 # assembler, which originally made it hard to motivate the effort.
23 # There was suggestion to mechanically translate 32-bit code, but I
24 # dismissed it, reasoning that x86_64 offers enough register bank
25 # capacity to fully utilize SHA-1 parallelism. Therefore this fresh
26 # implementation:-) However! While 64-bit code does perform better
27 # on Opteron, I failed to beat 32-bit assembler on EM64T core. Well,
28 # x86_64 does offer larger *addressable* bank, but out-of-order core
29 # reaches for even more registers through dynamic aliasing, and EM64T
30 # core must have managed to run-time optimize even 32-bit code just as
31 # good as 64-bit one. Performance improvement is summarized in the
34 # gcc 3.4 32-bit asm cycles/byte
35 # Opteron +45% +20% 6.8
36 # Xeon P4 +65% +0% 9.9
41 # The code was revised to minimize code size and to maximize
42 # "distance" between instructions producing input to 'lea'
43 # instruction and the 'lea' instruction itself, which is essential
44 # for Intel Atom core.
48 # Add SSSE3, Supplemental[!] SSE3, implementation. The idea behind it
49 # is to offload message schedule denoted by Wt in NIST specification,
50 # or Xupdate in OpenSSL source, to SIMD unit. See sha1-586.pl module
51 # for background and implementation details. The only difference from
52 # 32-bit code is that 64-bit code doesn't have to spill @X[] elements
53 # to free temporary registers.
57 # Add AVX code path. See sha1-586.pl for further information.
61 # Add AVX2+BMI code path. Initial attempt (utilizing BMI instructions
62 # and loading pair of consecutive blocks to 256-bit %ymm registers)
63 # did not provide impressive performance improvement till a crucial
64 # hint regarding the number of Xupdate iterations to pre-compute in
65 # advance was provided by Ilya Albrekht of Intel Corp.
69 # Add support for Intel SHA Extensions.
71 ######################################################################
72 # Current performance is summarized in following table. Numbers are
73 # CPU clock cycles spent to process single byte (less is better).
78 # Core2 6.55 6.05/+8% -
79 # Westmere 6.73 5.30/+27% -
80 # Sandy Bridge 7.70 6.10/+26% 4.99/+54%
81 # Ivy Bridge 6.06 4.67/+30% 4.60/+32%
82 # Haswell 5.45 4.15/+31% 3.57/+53%
83 # Skylake 5.18 4.06/+28% 3.54/+46%
84 # Bulldozer 9.11 5.95/+53%
85 # VIA Nano 9.32 7.15/+30%
87 # Silvermont 13.1(*) 9.37/+40%
89 # (*) obviously suboptimal result, nothing was done about it,
90 # because SSSE3 code is compiled unconditionally;
94 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
96 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
98 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
99 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
100 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
101 die "can't locate x86_64-xlate.pl";
103 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
104 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
105 $avx = ($1>=2.19) + ($1>=2.22);
108 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
109 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
110 $avx = ($1>=2.09) + ($1>=2.10);
113 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
114 `ml64 2>&1` =~ /Version ([0-9]+)\./) {
115 $avx = ($1>=10) + ($1>=11);
118 if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([2-9]\.[0-9]+)/) {
119 $avx = ($2>=3.0) + ($2>3.0);
122 $shaext=1; ### set to zero if compiling for 1.0.1
123 $avx=1 if (!$shaext && $avx);
125 open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\"";
128 $ctx="%rdi"; # 1st arg
129 $inp="%rsi"; # 2nd arg
130 $num="%rdx"; # 3rd arg
132 # reassign arguments in order to produce more compact code
140 @xi=("%edx","%ebp","%r14d");
150 my ($i,$a,$b,$c,$d,$e)=@_;
152 $code.=<<___ if ($i==0);
153 mov `4*$i`($inp),$xi[0]
156 $code.=<<___ if ($i<15);
157 mov `4*$j`($inp),$xi[1]
159 mov $xi[0],`4*$i`(%rsp)
165 lea 0x5a827999($xi[0],$e),$e
171 $code.=<<___ if ($i>=15);
172 xor `4*($j%16)`(%rsp),$xi[1]
174 mov $xi[0],`4*($i%16)`(%rsp)
176 xor `4*(($j+2)%16)`(%rsp),$xi[1]
179 xor `4*(($j+8)%16)`(%rsp),$xi[1]
181 lea 0x5a827999($xi[0],$e),$e
188 push(@xi,shift(@xi));
192 my ($i,$a,$b,$c,$d,$e)=@_;
194 my $K=($i<40)?0x6ed9eba1:0xca62c1d6;
195 $code.=<<___ if ($i<79);
196 xor `4*($j%16)`(%rsp),$xi[1]
198 `"mov $xi[0],".4*($i%16)."(%rsp)" if ($i<72)`
200 xor `4*(($j+2)%16)`(%rsp),$xi[1]
203 xor `4*(($j+8)%16)`(%rsp),$xi[1]
211 $code.=<<___ if ($i==79);
222 push(@xi,shift(@xi));
226 my ($i,$a,$b,$c,$d,$e)=@_;
229 xor `4*($j%16)`(%rsp),$xi[1]
231 mov $xi[0],`4*($i%16)`(%rsp)
233 xor `4*(($j+2)%16)`(%rsp),$xi[1]
236 xor `4*(($j+8)%16)`(%rsp),$xi[1]
237 lea 0x8f1bbcdc($xi[0],$e),$e
247 push(@xi,shift(@xi));
252 .extern OPENSSL_ia32cap_P
254 .globl sha1_block_data_order
255 .type sha1_block_data_order,\@function,3
257 sha1_block_data_order:
258 mov OPENSSL_ia32cap_P+0(%rip),%r9d
259 mov OPENSSL_ia32cap_P+4(%rip),%r8d
260 mov OPENSSL_ia32cap_P+8(%rip),%r10d
261 test \$`1<<9`,%r8d # check SSSE3 bit
264 $code.=<<___ if ($shaext);
265 test \$`1<<29`,%r10d # check SHA bit
268 $code.=<<___ if ($avx>1);
269 and \$`1<<3|1<<5|1<<8`,%r10d # check AVX2+BMI1+BMI2
270 cmp \$`1<<3|1<<5|1<<8`,%r10d
273 $code.=<<___ if ($avx);
274 and \$`1<<28`,%r8d # mask AVX bit
275 and \$`1<<30`,%r9d # mask "Intel CPU" bit
277 cmp \$`1<<28|1<<30`,%r8d
291 mov %rdi,$ctx # reassigned argument
293 mov %rsi,$inp # reassigned argument
295 mov %rdx,$num # reassigned argument
296 mov %rax,`16*4`(%rsp)
309 for($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); }
310 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
311 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
312 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
326 lea `16*4`($inp),$inp
329 mov `16*4`(%rsp),%rsi
338 .size sha1_block_data_order,.-sha1_block_data_order
341 ######################################################################
342 # Intel SHA Extensions implementation of SHA1 update function.
344 my ($ctx,$inp,$num)=("%rdi","%rsi","%rdx");
345 my ($ABCD,$E,$E_,$BSWAP,$ABCD_SAVE,$E_SAVE)=map("%xmm$_",(0..3,8,9));
346 my @MSG=map("%xmm$_",(4..7));
349 .type sha1_block_data_order_shaext,\@function,3
351 sha1_block_data_order_shaext:
354 $code.=<<___ if ($win64);
355 lea `-8-4*16`(%rsp),%rsp
356 movaps %xmm6,-8-4*16(%rax)
357 movaps %xmm7,-8-3*16(%rax)
358 movaps %xmm8,-8-2*16(%rax)
359 movaps %xmm9,-8-1*16(%rax)
365 movdqa K_XX_XX+0xa0(%rip),$BSWAP # byte-n-word swap
367 movdqu ($inp),@MSG[0]
368 pshufd \$0b00011011,$ABCD,$ABCD # flip word order
369 movdqu 0x10($inp),@MSG[1]
370 pshufd \$0b00011011,$E,$E # flip word order
371 movdqu 0x20($inp),@MSG[2]
372 pshufb $BSWAP,@MSG[0]
373 movdqu 0x30($inp),@MSG[3]
374 pshufb $BSWAP,@MSG[1]
375 pshufb $BSWAP,@MSG[2]
376 movdqa $E,$E_SAVE # offload $E
377 pshufb $BSWAP,@MSG[3]
383 lea 0x40($inp),%rax # next input block
386 movdqa $ABCD,$ABCD_SAVE # offload $ABCD
388 for($i=0;$i<20-4;$i+=2) {
390 sha1msg1 @MSG[1],@MSG[0]
392 sha1rnds4 \$`int($i/5)`,$E,$ABCD # 0-3...
393 sha1nexte @MSG[1],$E_
395 sha1msg1 @MSG[2],@MSG[1]
396 sha1msg2 @MSG[3],@MSG[0]
399 sha1rnds4 \$`int(($i+1)/5)`,$E_,$ABCD
402 sha1msg2 @MSG[0],@MSG[1]
404 push(@MSG,shift(@MSG)); push(@MSG,shift(@MSG));
407 movdqu ($inp),@MSG[0]
409 sha1rnds4 \$3,$E,$ABCD # 64-67
410 sha1nexte @MSG[1],$E_
411 movdqu 0x10($inp),@MSG[1]
412 pshufb $BSWAP,@MSG[0]
415 sha1rnds4 \$3,$E_,$ABCD # 68-71
417 movdqu 0x20($inp),@MSG[2]
418 pshufb $BSWAP,@MSG[1]
421 sha1rnds4 \$3,$E,$ABCD # 72-75
422 sha1nexte @MSG[3],$E_
423 movdqu 0x30($inp),@MSG[3]
424 pshufb $BSWAP,@MSG[2]
427 sha1rnds4 \$3,$E_,$ABCD # 76-79
429 pshufb $BSWAP,@MSG[3]
431 paddd $ABCD_SAVE,$ABCD
432 movdqa $E,$E_SAVE # offload $E
436 pshufd \$0b00011011,$ABCD,$ABCD
437 pshufd \$0b00011011,$E,$E
441 $code.=<<___ if ($win64);
442 movaps -8-4*16(%rax),%xmm6
443 movaps -8-3*16(%rax),%xmm7
444 movaps -8-2*16(%rax),%xmm8
445 movaps -8-1*16(%rax),%xmm9
451 .size sha1_block_data_order_shaext,.-sha1_block_data_order_shaext
456 my @X=map("%xmm$_",(4..7,0..3));
457 my @Tx=map("%xmm$_",(8..10));
459 my @V=($A,$B,$C,$D,$E)=("%eax","%ebx","%ecx","%edx","%ebp"); # size optimization
460 my @T=("%esi","%edi");
465 my $_rol=sub { &rol(@_) };
466 my $_ror=sub { &ror(@_) };
472 jmp .Lalign32_$sn # see "Decoded ICache" in manual
480 .type sha1_block_data_order_ssse3,\@function,3
482 sha1_block_data_order_ssse3:
488 push %r13 # redundant, done to share Win64 SE handler
490 lea `-64-($win64?6*16:0)`(%rsp),%rsp
492 $code.=<<___ if ($win64);
493 movaps %xmm6,-40-6*16(%rax)
494 movaps %xmm7,-40-5*16(%rax)
495 movaps %xmm8,-40-4*16(%rax)
496 movaps %xmm9,-40-3*16(%rax)
497 movaps %xmm10,-40-2*16(%rax)
498 movaps %xmm11,-40-1*16(%rax)
502 mov %rax,%r14 # original %rsp
504 mov %rdi,$ctx # reassigned argument
505 mov %rsi,$inp # reassigned argument
506 mov %rdx,$num # reassigned argument
510 lea K_XX_XX+64(%rip),$K_XX_XX
512 mov 0($ctx),$A # load context
516 mov $B,@T[0] # magic seed
522 movdqa 64($K_XX_XX),@X[2] # pbswap mask
523 movdqa -64($K_XX_XX),@Tx[1] # K_00_19
524 movdqu 0($inp),@X[-4&7] # load input to %xmm[0-3]
525 movdqu 16($inp),@X[-3&7]
526 movdqu 32($inp),@X[-2&7]
527 movdqu 48($inp),@X[-1&7]
528 pshufb @X[2],@X[-4&7] # byte swap
529 pshufb @X[2],@X[-3&7]
530 pshufb @X[2],@X[-2&7]
532 paddd @Tx[1],@X[-4&7] # add K_00_19
533 pshufb @X[2],@X[-1&7]
534 paddd @Tx[1],@X[-3&7]
535 paddd @Tx[1],@X[-2&7]
536 movdqa @X[-4&7],0(%rsp) # X[]+K xfer to IALU
537 psubd @Tx[1],@X[-4&7] # restore X[]
538 movdqa @X[-3&7],16(%rsp)
539 psubd @Tx[1],@X[-3&7]
540 movdqa @X[-2&7],32(%rsp)
541 psubd @Tx[1],@X[-2&7]
545 sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm
546 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
548 $arg = "\$$arg" if ($arg*1 eq $arg);
549 $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
552 sub Xupdate_ssse3_16_31() # recall that $Xi starts wtih 4
555 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
558 eval(shift(@insns)); # ror
559 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
561 &movdqa (@Tx[0],@X[-1&7]);
562 &paddd (@Tx[1],@X[-1&7]);
566 &punpcklqdq(@X[0],@X[-3&7]); # compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8);
568 eval(shift(@insns)); # rol
570 &psrldq (@Tx[0],4); # "X[-3]", 3 dwords
574 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
576 eval(shift(@insns)); # ror
577 &pxor (@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
582 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
584 eval(shift(@insns)); # rol
585 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
589 &movdqa (@Tx[2],@X[0]);
592 eval(shift(@insns)); # ror
593 &movdqa (@Tx[0],@X[0]);
596 &pslldq (@Tx[2],12); # "X[0]"<<96, extract one dword
597 &paddd (@X[0],@X[0]);
603 eval(shift(@insns)); # rol
605 &movdqa (@Tx[1],@Tx[2]);
611 eval(shift(@insns)); # ror
612 &por (@X[0],@Tx[0]); # "X[0]"<<<=1
618 &pxor (@X[0],@Tx[2]);
620 &movdqa (@Tx[2],eval(2*16*(($Xi)/5)-64)."($K_XX_XX)"); # K_XX_XX
621 eval(shift(@insns)); # rol
625 &pxor (@X[0],@Tx[1]); # "X[0]"^=("X[0]">>96)<<<2
626 &pshufd (@Tx[1],@X[-1&7],0xee) if ($Xi==7); # was &movdqa (@Tx[0],@X[-1&7]) in Xupdate_ssse3_32_79
628 foreach (@insns) { eval; } # remaining instructions [if any]
630 $Xi++; push(@X,shift(@X)); # "rotate" X[]
631 push(@Tx,shift(@Tx));
634 sub Xupdate_ssse3_32_79()
637 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
640 eval(shift(@insns)) if ($Xi==8);
641 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
642 eval(shift(@insns)) if ($Xi==8);
643 eval(shift(@insns)); # body_20_39
645 eval(shift(@insns)) if (@insns[1] =~ /_ror/);
646 eval(shift(@insns)) if (@insns[0] =~ /_ror/);
647 &punpcklqdq(@Tx[0],@X[-1&7]); # compose "X[-6]", was &palignr(@Tx[0],@X[-2&7],8);
649 eval(shift(@insns)); # rol
651 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
655 &movdqa (@Tx[2],@Tx[1]);# "perpetuate" K_XX_XX...
656 } else { # ... or load next one
657 &movdqa (@Tx[2],eval(2*16*($Xi/5)-64)."($K_XX_XX)");
659 eval(shift(@insns)); # ror
660 &paddd (@Tx[1],@X[-1&7]);
663 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-6]"
664 eval(shift(@insns)); # body_20_39
667 eval(shift(@insns)); # rol
668 eval(shift(@insns)) if (@insns[0] =~ /_ror/);
670 &movdqa (@Tx[0],@X[0]);
673 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
674 eval(shift(@insns)); # ror
676 eval(shift(@insns)); # body_20_39
682 eval(shift(@insns)) if (@insns[0] =~ /_rol/);# rol
685 eval(shift(@insns)); # ror
687 &por (@X[0],@Tx[0]); # "X[0]"<<<=2
689 eval(shift(@insns)); # body_20_39
690 eval(shift(@insns)) if (@insns[1] =~ /_rol/);
691 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
692 &pshufd(@Tx[1],@X[-1&7],0xee) if ($Xi<19); # was &movdqa (@Tx[1],@X[0])
694 eval(shift(@insns)); # rol
697 eval(shift(@insns)); # rol
700 foreach (@insns) { eval; } # remaining instructions
702 $Xi++; push(@X,shift(@X)); # "rotate" X[]
703 push(@Tx,shift(@Tx));
706 sub Xuplast_ssse3_80()
709 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
716 &paddd (@Tx[1],@X[-1&7]);
720 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU
722 foreach (@insns) { eval; } # remaining instructions
725 &je (".Ldone_ssse3");
727 unshift(@Tx,pop(@Tx));
729 &movdqa (@X[2],"64($K_XX_XX)"); # pbswap mask
730 &movdqa (@Tx[1],"-64($K_XX_XX)"); # K_00_19
731 &movdqu (@X[-4&7],"0($inp)"); # load input
732 &movdqu (@X[-3&7],"16($inp)");
733 &movdqu (@X[-2&7],"32($inp)");
734 &movdqu (@X[-1&7],"48($inp)");
735 &pshufb (@X[-4&7],@X[2]); # byte swap
744 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
750 &pshufb (@X[($Xi-3)&7],@X[2]);
755 &paddd (@X[($Xi-4)&7],@Tx[1]);
760 &movdqa (eval(16*$Xi)."(%rsp)",@X[($Xi-4)&7]); # X[]+K xfer to IALU
765 &psubd (@X[($Xi-4)&7],@Tx[1]);
767 foreach (@insns) { eval; }
774 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
777 foreach (@insns) { eval; }
780 sub body_00_19 () { # ((c^d)&b)^d
781 # on start @T[0]=(c^d)&b
782 return &body_20_39() if ($rx==19); $rx++;
784 '($a,$b,$c,$d,$e)=@V;'.
785 '&$_ror ($b,$j?7:2)', # $b>>>2
787 '&mov (@T[1],$a)', # $b for next round
789 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
790 '&xor ($b,$c)', # $c^$d for next round
794 '&and (@T[1],$b)', # ($b&($c^$d)) for next round
796 '&xor ($b,$c)', # restore $b
797 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
801 sub body_20_39 () { # b^d^c
803 return &body_40_59() if ($rx==39); $rx++;
805 '($a,$b,$c,$d,$e)=@V;'.
806 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
807 '&xor (@T[0],$d) if($j==19);'.
808 '&xor (@T[0],$c) if($j> 19)', # ($b^$d^$c)
809 '&mov (@T[1],$a)', # $b for next round
813 '&xor (@T[1],$c) if ($j< 79)', # $b^$d for next round
815 '&$_ror ($b,7)', # $b>>>2
816 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
820 sub body_40_59 () { # ((b^c)&(c^d))^c
821 # on entry @T[0]=(b^c), (c^=d)
824 '($a,$b,$c,$d,$e)=@V;'.
825 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
826 '&and (@T[0],$c) if ($j>=40)', # (b^c)&(c^d)
827 '&xor ($c,$d) if ($j>=40)', # restore $c
829 '&$_ror ($b,7)', # $b>>>2
830 '&mov (@T[1],$a)', # $b for next round
835 '&xor (@T[1],$c) if ($j==59);'.
836 '&xor (@T[1],$b) if ($j< 59)', # b^c for next round
838 '&xor ($b,$c) if ($j< 59)', # c^d for next round
839 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
846 &Xupdate_ssse3_16_31(\&body_00_19);
847 &Xupdate_ssse3_16_31(\&body_00_19);
848 &Xupdate_ssse3_16_31(\&body_00_19);
849 &Xupdate_ssse3_16_31(\&body_00_19);
850 &Xupdate_ssse3_32_79(\&body_00_19);
851 &Xupdate_ssse3_32_79(\&body_20_39);
852 &Xupdate_ssse3_32_79(\&body_20_39);
853 &Xupdate_ssse3_32_79(\&body_20_39);
854 &Xupdate_ssse3_32_79(\&body_20_39);
855 &Xupdate_ssse3_32_79(\&body_20_39);
856 &Xupdate_ssse3_32_79(\&body_40_59);
857 &Xupdate_ssse3_32_79(\&body_40_59);
858 &Xupdate_ssse3_32_79(\&body_40_59);
859 &Xupdate_ssse3_32_79(\&body_40_59);
860 &Xupdate_ssse3_32_79(\&body_40_59);
861 &Xupdate_ssse3_32_79(\&body_20_39);
862 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done"
864 $saved_j=$j; @saved_V=@V;
866 &Xloop_ssse3(\&body_20_39);
867 &Xloop_ssse3(\&body_20_39);
868 &Xloop_ssse3(\&body_20_39);
871 add 0($ctx),$A # update context
878 mov @T[0],$B # magic seed
890 $j=$saved_j; @V=@saved_V;
892 &Xtail_ssse3(\&body_20_39);
893 &Xtail_ssse3(\&body_20_39);
894 &Xtail_ssse3(\&body_20_39);
897 add 0($ctx),$A # update context
908 $code.=<<___ if ($win64);
909 movaps -40-6*16(%r14),%xmm6
910 movaps -40-5*16(%r14),%xmm7
911 movaps -40-4*16(%r14),%xmm8
912 movaps -40-3*16(%r14),%xmm9
913 movaps -40-2*16(%r14),%xmm10
914 movaps -40-1*16(%r14),%xmm11
926 .size sha1_block_data_order_ssse3,.-sha1_block_data_order_ssse3
930 $Xi=4; # reset variables
931 @X=map("%xmm$_",(4..7,0..3));
932 @Tx=map("%xmm$_",(8..10));
936 my $done_avx_label=".Ldone_avx";
938 my $_rol=sub { &shld(@_[0],@_) };
939 my $_ror=sub { &shrd(@_[0],@_) };
942 .type sha1_block_data_order_avx,\@function,3
944 sha1_block_data_order_avx:
950 push %r13 # redundant, done to share Win64 SE handler
952 lea `-64-($win64?6*16:0)`(%rsp),%rsp
955 $code.=<<___ if ($win64);
956 vmovaps %xmm6,-40-6*16(%rax)
957 vmovaps %xmm7,-40-5*16(%rax)
958 vmovaps %xmm8,-40-4*16(%rax)
959 vmovaps %xmm9,-40-3*16(%rax)
960 vmovaps %xmm10,-40-2*16(%rax)
961 vmovaps %xmm11,-40-1*16(%rax)
965 mov %rax,%r14 # original %rsp
967 mov %rdi,$ctx # reassigned argument
968 mov %rsi,$inp # reassigned argument
969 mov %rdx,$num # reassigned argument
973 lea K_XX_XX+64(%rip),$K_XX_XX
975 mov 0($ctx),$A # load context
979 mov $B,@T[0] # magic seed
985 vmovdqa 64($K_XX_XX),@X[2] # pbswap mask
986 vmovdqa -64($K_XX_XX),$Kx # K_00_19
987 vmovdqu 0($inp),@X[-4&7] # load input to %xmm[0-3]
988 vmovdqu 16($inp),@X[-3&7]
989 vmovdqu 32($inp),@X[-2&7]
990 vmovdqu 48($inp),@X[-1&7]
991 vpshufb @X[2],@X[-4&7],@X[-4&7] # byte swap
993 vpshufb @X[2],@X[-3&7],@X[-3&7]
994 vpshufb @X[2],@X[-2&7],@X[-2&7]
995 vpshufb @X[2],@X[-1&7],@X[-1&7]
996 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19
997 vpaddd $Kx,@X[-3&7],@X[1]
998 vpaddd $Kx,@X[-2&7],@X[2]
999 vmovdqa @X[0],0(%rsp) # X[]+K xfer to IALU
1000 vmovdqa @X[1],16(%rsp)
1001 vmovdqa @X[2],32(%rsp)
1005 sub Xupdate_avx_16_31() # recall that $Xi starts wtih 4
1008 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
1009 my ($a,$b,$c,$d,$e);
1011 eval(shift(@insns));
1012 eval(shift(@insns));
1013 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1014 eval(shift(@insns));
1015 eval(shift(@insns));
1017 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1018 eval(shift(@insns));
1019 eval(shift(@insns));
1020 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1021 eval(shift(@insns));
1022 eval(shift(@insns));
1023 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1024 eval(shift(@insns));
1025 eval(shift(@insns));
1027 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1028 eval(shift(@insns));
1029 eval(shift(@insns));
1030 eval(shift(@insns));
1031 eval(shift(@insns));
1033 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1034 eval(shift(@insns));
1035 eval(shift(@insns));
1036 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1037 eval(shift(@insns));
1038 eval(shift(@insns));
1040 &vpsrld (@Tx[0],@X[0],31);
1041 eval(shift(@insns));
1042 eval(shift(@insns));
1043 eval(shift(@insns));
1044 eval(shift(@insns));
1046 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1047 &vpaddd (@X[0],@X[0],@X[0]);
1048 eval(shift(@insns));
1049 eval(shift(@insns));
1050 eval(shift(@insns));
1051 eval(shift(@insns));
1053 &vpsrld (@Tx[1],@Tx[2],30);
1054 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1055 eval(shift(@insns));
1056 eval(shift(@insns));
1057 eval(shift(@insns));
1058 eval(shift(@insns));
1060 &vpslld (@Tx[2],@Tx[2],2);
1061 &vpxor (@X[0],@X[0],@Tx[1]);
1062 eval(shift(@insns));
1063 eval(shift(@insns));
1064 eval(shift(@insns));
1065 eval(shift(@insns));
1067 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1068 eval(shift(@insns));
1069 eval(shift(@insns));
1070 &vmovdqa ($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1071 eval(shift(@insns));
1072 eval(shift(@insns));
1075 foreach (@insns) { eval; } # remaining instructions [if any]
1077 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1080 sub Xupdate_avx_32_79()
1083 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
1084 my ($a,$b,$c,$d,$e);
1086 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1087 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1088 eval(shift(@insns)); # body_20_39
1089 eval(shift(@insns));
1090 eval(shift(@insns));
1091 eval(shift(@insns)); # rol
1093 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1094 eval(shift(@insns));
1095 eval(shift(@insns)) if (@insns[0] !~ /&ro[rl]/);
1096 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1097 &vmovdqa ($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0);
1098 eval(shift(@insns)); # ror
1099 eval(shift(@insns));
1101 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]"
1102 eval(shift(@insns)); # body_20_39
1103 eval(shift(@insns));
1104 eval(shift(@insns));
1105 eval(shift(@insns)); # rol
1107 &vpsrld (@Tx[0],@X[0],30);
1108 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1109 eval(shift(@insns));
1110 eval(shift(@insns));
1111 eval(shift(@insns)); # ror
1112 eval(shift(@insns));
1114 &vpslld (@X[0],@X[0],2);
1115 eval(shift(@insns)); # body_20_39
1116 eval(shift(@insns));
1117 eval(shift(@insns));
1118 eval(shift(@insns)); # rol
1119 eval(shift(@insns));
1120 eval(shift(@insns));
1121 eval(shift(@insns)); # ror
1122 eval(shift(@insns));
1124 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2
1125 eval(shift(@insns)); # body_20_39
1126 eval(shift(@insns));
1127 eval(shift(@insns));
1128 eval(shift(@insns)); # rol
1129 eval(shift(@insns));
1130 eval(shift(@insns));
1131 eval(shift(@insns)); # rol
1132 eval(shift(@insns));
1134 foreach (@insns) { eval; } # remaining instructions
1136 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1139 sub Xuplast_avx_80()
1142 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1143 my ($a,$b,$c,$d,$e);
1145 eval(shift(@insns));
1146 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1147 eval(shift(@insns));
1148 eval(shift(@insns));
1149 eval(shift(@insns));
1150 eval(shift(@insns));
1152 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU
1154 foreach (@insns) { eval; } # remaining instructions
1157 &je ($done_avx_label);
1159 &vmovdqa(@X[2],"64($K_XX_XX)"); # pbswap mask
1160 &vmovdqa($Kx,"-64($K_XX_XX)"); # K_00_19
1161 &vmovdqu(@X[-4&7],"0($inp)"); # load input
1162 &vmovdqu(@X[-3&7],"16($inp)");
1163 &vmovdqu(@X[-2&7],"32($inp)");
1164 &vmovdqu(@X[-1&7],"48($inp)");
1165 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1174 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1175 my ($a,$b,$c,$d,$e);
1177 eval(shift(@insns));
1178 eval(shift(@insns));
1179 &vpshufb(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1180 eval(shift(@insns));
1181 eval(shift(@insns));
1182 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],$Kx);
1183 eval(shift(@insns));
1184 eval(shift(@insns));
1185 eval(shift(@insns));
1186 eval(shift(@insns));
1187 &vmovdqa(eval(16*$Xi)."(%rsp)",@X[$Xi&7]); # X[]+K xfer to IALU
1188 eval(shift(@insns));
1189 eval(shift(@insns));
1191 foreach (@insns) { eval; }
1198 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1199 my ($a,$b,$c,$d,$e);
1201 foreach (@insns) { eval; }
1208 &Xupdate_avx_16_31(\&body_00_19);
1209 &Xupdate_avx_16_31(\&body_00_19);
1210 &Xupdate_avx_16_31(\&body_00_19);
1211 &Xupdate_avx_16_31(\&body_00_19);
1212 &Xupdate_avx_32_79(\&body_00_19);
1213 &Xupdate_avx_32_79(\&body_20_39);
1214 &Xupdate_avx_32_79(\&body_20_39);
1215 &Xupdate_avx_32_79(\&body_20_39);
1216 &Xupdate_avx_32_79(\&body_20_39);
1217 &Xupdate_avx_32_79(\&body_20_39);
1218 &Xupdate_avx_32_79(\&body_40_59);
1219 &Xupdate_avx_32_79(\&body_40_59);
1220 &Xupdate_avx_32_79(\&body_40_59);
1221 &Xupdate_avx_32_79(\&body_40_59);
1222 &Xupdate_avx_32_79(\&body_40_59);
1223 &Xupdate_avx_32_79(\&body_20_39);
1224 &Xuplast_avx_80(\&body_20_39); # can jump to "done"
1226 $saved_j=$j; @saved_V=@V;
1228 &Xloop_avx(\&body_20_39);
1229 &Xloop_avx(\&body_20_39);
1230 &Xloop_avx(\&body_20_39);
1233 add 0($ctx),$A # update context
1240 mov @T[0],$B # magic seed
1252 $j=$saved_j; @V=@saved_V;
1254 &Xtail_avx(\&body_20_39);
1255 &Xtail_avx(\&body_20_39);
1256 &Xtail_avx(\&body_20_39);
1261 add 0($ctx),$A # update context
1272 $code.=<<___ if ($win64);
1273 movaps -40-6*16(%r14),%xmm6
1274 movaps -40-5*16(%r14),%xmm7
1275 movaps -40-4*16(%r14),%xmm8
1276 movaps -40-3*16(%r14),%xmm9
1277 movaps -40-2*16(%r14),%xmm10
1278 movaps -40-1*16(%r14),%xmm11
1290 .size sha1_block_data_order_avx,.-sha1_block_data_order_avx
1295 $Xi=4; # reset variables
1296 @X=map("%ymm$_",(4..7,0..3));
1297 @Tx=map("%ymm$_",(8..10));
1301 my @ROTX=("%eax","%ebp","%ebx","%ecx","%edx","%esi");
1302 my ($a5,$t0)=("%r12d","%edi");
1304 my ($A,$F,$B,$C,$D,$E)=@ROTX;
1309 .type sha1_block_data_order_avx2,\@function,3
1311 sha1_block_data_order_avx2:
1321 $code.=<<___ if ($win64);
1322 lea -6*16(%rsp),%rsp
1323 vmovaps %xmm6,-40-6*16(%rax)
1324 vmovaps %xmm7,-40-5*16(%rax)
1325 vmovaps %xmm8,-40-4*16(%rax)
1326 vmovaps %xmm9,-40-3*16(%rax)
1327 vmovaps %xmm10,-40-2*16(%rax)
1328 vmovaps %xmm11,-40-1*16(%rax)
1332 mov %rax,%r14 # original %rsp
1333 mov %rdi,$ctx # reassigned argument
1334 mov %rsi,$inp # reassigned argument
1335 mov %rdx,$num # reassigned argument
1342 lea K_XX_XX+64(%rip),$K_XX_XX
1344 mov 0($ctx),$A # load context
1346 cmovae $inp,$frame # next or same block
1351 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask
1353 vmovdqu ($inp),%xmm0
1354 vmovdqu 16($inp),%xmm1
1355 vmovdqu 32($inp),%xmm2
1356 vmovdqu 48($inp),%xmm3
1358 vinserti128 \$1,($frame),@X[-4&7],@X[-4&7]
1359 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7]
1360 vpshufb @X[2],@X[-4&7],@X[-4&7]
1361 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7]
1362 vpshufb @X[2],@X[-3&7],@X[-3&7]
1363 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7]
1364 vpshufb @X[2],@X[-2&7],@X[-2&7]
1365 vmovdqu -64($K_XX_XX),$Kx # K_00_19
1366 vpshufb @X[2],@X[-1&7],@X[-1&7]
1368 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19
1369 vpaddd $Kx,@X[-3&7],@X[1]
1370 vmovdqu @X[0],0(%rsp) # X[]+K xfer to IALU
1371 vpaddd $Kx,@X[-2&7],@X[2]
1372 vmovdqu @X[1],32(%rsp)
1373 vpaddd $Kx,@X[-1&7],@X[3]
1374 vmovdqu @X[2],64(%rsp)
1375 vmovdqu @X[3],96(%rsp)
1377 for (;$Xi<8;$Xi++) { # Xupdate_avx2_16_31
1380 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1381 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1382 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1383 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1384 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1385 &vpsrld (@Tx[0],@X[0],31);
1386 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1387 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1388 &vpaddd (@X[0],@X[0],@X[0]);
1389 &vpsrld (@Tx[1],@Tx[2],30);
1390 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1391 &vpslld (@Tx[2],@Tx[2],2);
1392 &vpxor (@X[0],@X[0],@Tx[1]);
1393 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1394 &vpaddd (@Tx[1],@X[0],$Kx);
1395 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1397 push(@X,shift(@X)); # "rotate" X[]
1400 lea 128(%rsp),$frame
1409 sub bodyx_00_19 () { # 8 instructions, 3 cycles critical path
1410 # at start $f=(b&c)^(~b&d), $b>>>=2
1411 return &bodyx_20_39() if ($rx==19); $rx++;
1413 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1415 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1416 '&lea ($frame,"256($frame)") if ($j%32==31);',
1417 '&andn ($t0,$a,$c)', # ~b&d for next round
1419 '&add ($e,$f)', # e+=(b&c)^(~b&d)
1420 '&rorx ($a5,$a,27)', # a<<<5
1421 '&rorx ($f,$a,2)', # b>>>2 for next round
1422 '&and ($a,$b)', # b&c for next round
1424 '&add ($e,$a5)', # e+=a<<<5
1425 '&xor ($a,$t0);'. # f=(b&c)^(~b&d) for next round
1427 'unshift(@ROTX,pop(@ROTX)); $j++;'
1431 sub bodyx_20_39 () { # 7 instructions, 2 cycles critical path
1432 # on entry $f=b^c^d, $b>>>=2
1433 return &bodyx_40_59() if ($rx==39); $rx++;
1435 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1437 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1438 '&lea ($frame,"256($frame)") if ($j%32==31);',
1440 '&lea ($e,"($e,$f)")', # e+=b^c^d
1441 '&rorx ($a5,$a,27)', # a<<<5
1442 '&rorx ($f,$a,2) if ($j<79)', # b>>>2 in next round
1443 '&xor ($a,$b) if ($j<79)', # b^c for next round
1445 '&add ($e,$a5)', # e+=a<<<5
1446 '&xor ($a,$c) if ($j<79);'. # f=b^c^d for next round
1448 'unshift(@ROTX,pop(@ROTX)); $j++;'
1452 sub bodyx_40_59 () { # 10 instructions, 3 cycles critical path
1453 # on entry $f=((b^c)&(c^d)), $b>>>=2
1456 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1458 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1459 '&lea ($frame,"256($frame)") if ($j%32==31);',
1460 '&xor ($f,$c) if ($j>39)', # (b^c)&(c^d)^c
1461 '&mov ($t0,$b) if ($j<59)', # count on zero latency
1462 '&xor ($t0,$c) if ($j<59)', # c^d for next round
1464 '&lea ($e,"($e,$f)")', # e+=(b^c)&(c^d)^c
1465 '&rorx ($a5,$a,27)', # a<<<5
1466 '&rorx ($f,$a,2)', # b>>>2 in next round
1467 '&xor ($a,$b)', # b^c for next round
1469 '&add ($e,$a5)', # e+=a<<<5
1470 '&and ($a,$t0) if ($j< 59);'. # f=(b^c)&(c^d) for next round
1471 '&xor ($a,$c) if ($j==59);'. # f=b^c^d for next round
1473 'unshift(@ROTX,pop(@ROTX)); $j++;'
1477 sub Xupdate_avx2_16_31() # recall that $Xi starts wtih 4
1480 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 instructions
1481 my ($a,$b,$c,$d,$e);
1483 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1484 eval(shift(@insns));
1485 eval(shift(@insns));
1486 eval(shift(@insns));
1487 eval(shift(@insns));
1489 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1490 eval(shift(@insns));
1491 eval(shift(@insns));
1492 eval(shift(@insns));
1494 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1495 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1496 eval(shift(@insns));
1497 eval(shift(@insns));
1499 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1500 eval(shift(@insns));
1501 eval(shift(@insns));
1502 eval(shift(@insns));
1503 eval(shift(@insns));
1505 &vpsrld (@Tx[0],@X[0],31);
1506 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1507 eval(shift(@insns));
1508 eval(shift(@insns));
1509 eval(shift(@insns));
1511 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1512 &vpaddd (@X[0],@X[0],@X[0]);
1513 eval(shift(@insns));
1514 eval(shift(@insns));
1516 &vpsrld (@Tx[1],@Tx[2],30);
1517 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1518 eval(shift(@insns));
1519 eval(shift(@insns));
1521 &vpslld (@Tx[2],@Tx[2],2);
1522 &vpxor (@X[0],@X[0],@Tx[1]);
1523 eval(shift(@insns));
1524 eval(shift(@insns));
1526 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1527 eval(shift(@insns));
1528 eval(shift(@insns));
1529 eval(shift(@insns));
1531 &vpaddd (@Tx[1],@X[0],$Kx);
1532 eval(shift(@insns));
1533 eval(shift(@insns));
1534 eval(shift(@insns));
1535 &vmovdqu(eval(32*($Xi))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1537 foreach (@insns) { eval; } # remaining instructions [if any]
1540 push(@X,shift(@X)); # "rotate" X[]
1543 sub Xupdate_avx2_32_79()
1546 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 to 50 instructions
1547 my ($a,$b,$c,$d,$e);
1549 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1550 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1551 eval(shift(@insns));
1552 eval(shift(@insns));
1554 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1555 &vmovdqu($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0);
1556 eval(shift(@insns));
1557 eval(shift(@insns));
1558 eval(shift(@insns));
1560 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]"
1561 eval(shift(@insns));
1562 eval(shift(@insns));
1563 eval(shift(@insns));
1565 &vpsrld (@Tx[0],@X[0],30);
1566 &vpslld (@X[0],@X[0],2);
1567 eval(shift(@insns));
1568 eval(shift(@insns));
1569 eval(shift(@insns));
1571 #&vpslld (@X[0],@X[0],2);
1572 eval(shift(@insns));
1573 eval(shift(@insns));
1574 eval(shift(@insns));
1576 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2
1577 eval(shift(@insns));
1578 eval(shift(@insns));
1579 eval(shift(@insns));
1580 eval(shift(@insns));
1582 &vpaddd (@Tx[1],@X[0],$Kx);
1583 eval(shift(@insns));
1584 eval(shift(@insns));
1585 eval(shift(@insns));
1586 eval(shift(@insns));
1588 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1590 foreach (@insns) { eval; } # remaining instructions
1593 push(@X,shift(@X)); # "rotate" X[]
1599 my @insns = (&$body,&$body,&$body,&$body,&$body); # 32 instructions
1600 my ($a,$b,$c,$d,$e);
1602 foreach (@insns) { eval; }
1606 &Xupdate_avx2_32_79(\&bodyx_00_19);
1607 &Xupdate_avx2_32_79(\&bodyx_00_19);
1608 &Xupdate_avx2_32_79(\&bodyx_00_19);
1609 &Xupdate_avx2_32_79(\&bodyx_00_19);
1611 &Xupdate_avx2_32_79(\&bodyx_20_39);
1612 &Xupdate_avx2_32_79(\&bodyx_20_39);
1613 &Xupdate_avx2_32_79(\&bodyx_20_39);
1614 &Xupdate_avx2_32_79(\&bodyx_20_39);
1617 &Xupdate_avx2_32_79(\&bodyx_40_59);
1618 &Xupdate_avx2_32_79(\&bodyx_40_59);
1619 &Xupdate_avx2_32_79(\&bodyx_40_59);
1620 &Xupdate_avx2_32_79(\&bodyx_40_59);
1622 &Xloop_avx2(\&bodyx_20_39);
1623 &Xloop_avx2(\&bodyx_20_39);
1624 &Xloop_avx2(\&bodyx_20_39);
1625 &Xloop_avx2(\&bodyx_20_39);
1628 lea 128($inp),$frame
1629 lea 128($inp),%rdi # borrow $t0
1631 cmovae $inp,$frame # next or previous block
1633 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c
1634 add 0($ctx),@ROTX[0] # update context
1635 add 4($ctx),@ROTX[1]
1636 add 8($ctx),@ROTX[3]
1637 mov @ROTX[0],0($ctx)
1638 add 12($ctx),@ROTX[4]
1639 mov @ROTX[1],4($ctx)
1640 mov @ROTX[0],$A # A=d
1641 add 16($ctx),@ROTX[5]
1643 mov @ROTX[3],8($ctx)
1644 mov @ROTX[4],$D # D=b
1645 #xchg @ROTX[5],$F # F=c, C=f
1646 mov @ROTX[4],12($ctx)
1647 mov @ROTX[1],$F # F=e
1648 mov @ROTX[5],16($ctx)
1650 mov @ROTX[5],$E # E=c
1652 #xchg $F,$E # E=c, F=e
1658 $Xi=4; # reset variables
1659 @X=map("%ymm$_",(4..7,0..3));
1662 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask
1663 cmp $num,%rdi # borrowed $t0
1666 vmovdqu -64(%rdi),%xmm0 # low part of @X[-4&7]
1667 vmovdqu -48(%rdi),%xmm1
1668 vmovdqu -32(%rdi),%xmm2
1669 vmovdqu -16(%rdi),%xmm3
1670 vinserti128 \$1,0($frame),@X[-4&7],@X[-4&7]
1671 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7]
1672 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7]
1673 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7]
1678 lea 128+16(%rsp),$frame
1685 $rx=$j=0; @ROTX=($A,$F,$B,$C,$D,$E);
1687 &Xloop_avx2 (\&bodyx_00_19);
1688 &Xloop_avx2 (\&bodyx_00_19);
1689 &Xloop_avx2 (\&bodyx_00_19);
1690 &Xloop_avx2 (\&bodyx_00_19);
1692 &Xloop_avx2 (\&bodyx_20_39);
1693 &vmovdqu ($Kx,"-64($K_XX_XX)"); # K_00_19
1694 &vpshufb (@X[-4&7],@X[-4&7],@X[2]); # byte swap
1695 &Xloop_avx2 (\&bodyx_20_39);
1696 &vpshufb (@X[-3&7],@X[-3&7],@X[2]);
1697 &vpaddd (@Tx[0],@X[-4&7],$Kx); # add K_00_19
1698 &Xloop_avx2 (\&bodyx_20_39);
1699 &vmovdqu ("0(%rsp)",@Tx[0]);
1700 &vpshufb (@X[-2&7],@X[-2&7],@X[2]);
1701 &vpaddd (@Tx[1],@X[-3&7],$Kx);
1702 &Xloop_avx2 (\&bodyx_20_39);
1703 &vmovdqu ("32(%rsp)",@Tx[1]);
1704 &vpshufb (@X[-1&7],@X[-1&7],@X[2]);
1705 &vpaddd (@X[2],@X[-2&7],$Kx);
1707 &Xloop_avx2 (\&bodyx_40_59);
1709 &vmovdqu ("64(%rsp)",@X[2]);
1710 &vpaddd (@X[3],@X[-1&7],$Kx);
1711 &Xloop_avx2 (\&bodyx_40_59);
1712 &vmovdqu ("96(%rsp)",@X[3]);
1713 &Xloop_avx2 (\&bodyx_40_59);
1714 &Xupdate_avx2_16_31(\&bodyx_40_59);
1716 &Xupdate_avx2_16_31(\&bodyx_20_39);
1717 &Xupdate_avx2_16_31(\&bodyx_20_39);
1718 &Xupdate_avx2_16_31(\&bodyx_20_39);
1719 &Xloop_avx2 (\&bodyx_20_39);
1722 lea 128(%rsp),$frame
1724 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c
1725 add 0($ctx),@ROTX[0] # update context
1726 add 4($ctx),@ROTX[1]
1727 add 8($ctx),@ROTX[3]
1728 mov @ROTX[0],0($ctx)
1729 add 12($ctx),@ROTX[4]
1730 mov @ROTX[1],4($ctx)
1731 mov @ROTX[0],$A # A=d
1732 add 16($ctx),@ROTX[5]
1734 mov @ROTX[3],8($ctx)
1735 mov @ROTX[4],$D # D=b
1736 #xchg @ROTX[5],$F # F=c, C=f
1737 mov @ROTX[4],12($ctx)
1738 mov @ROTX[1],$F # F=e
1739 mov @ROTX[5],16($ctx)
1741 mov @ROTX[5],$E # E=c
1743 #xchg $F,$E # E=c, F=e
1751 $code.=<<___ if ($win64);
1752 movaps -40-6*16(%r14),%xmm6
1753 movaps -40-5*16(%r14),%xmm7
1754 movaps -40-4*16(%r14),%xmm8
1755 movaps -40-3*16(%r14),%xmm9
1756 movaps -40-2*16(%r14),%xmm10
1757 movaps -40-1*16(%r14),%xmm11
1769 .size sha1_block_data_order_avx2,.-sha1_block_data_order_avx2
1776 .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19
1777 .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19
1778 .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39
1779 .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39
1780 .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59
1781 .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59
1782 .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79
1783 .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79
1784 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask
1785 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask
1786 .byte 0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0
1790 .asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1794 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1795 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
1803 .extern __imp_RtlVirtualUnwind
1804 .type se_handler,\@abi-omnipotent
1818 mov 120($context),%rax # pull context->Rax
1819 mov 248($context),%rbx # pull context->Rip
1821 lea .Lprologue(%rip),%r10
1822 cmp %r10,%rbx # context->Rip<.Lprologue
1823 jb .Lcommon_seh_tail
1825 mov 152($context),%rax # pull context->Rsp
1827 lea .Lepilogue(%rip),%r10
1828 cmp %r10,%rbx # context->Rip>=.Lepilogue
1829 jae .Lcommon_seh_tail
1831 mov `16*4`(%rax),%rax # pull saved stack pointer
1838 mov %rbx,144($context) # restore context->Rbx
1839 mov %rbp,160($context) # restore context->Rbp
1840 mov %r12,216($context) # restore context->R12
1841 mov %r13,224($context) # restore context->R13
1842 mov %r14,232($context) # restore context->R14
1844 jmp .Lcommon_seh_tail
1845 .size se_handler,.-se_handler
1848 $code.=<<___ if ($shaext);
1849 .type shaext_handler,\@abi-omnipotent
1863 mov 120($context),%rax # pull context->Rax
1864 mov 248($context),%rbx # pull context->Rip
1866 lea .Lprologue_shaext(%rip),%r10
1867 cmp %r10,%rbx # context->Rip<.Lprologue
1868 jb .Lcommon_seh_tail
1870 lea .Lepilogue_shaext(%rip),%r10
1871 cmp %r10,%rbx # context->Rip>=.Lepilogue
1872 jae .Lcommon_seh_tail
1874 lea -8-4*16(%rax),%rsi
1875 lea 512($context),%rdi # &context.Xmm6
1877 .long 0xa548f3fc # cld; rep movsq
1879 jmp .Lcommon_seh_tail
1880 .size shaext_handler,.-shaext_handler
1884 .type ssse3_handler,\@abi-omnipotent
1898 mov 120($context),%rax # pull context->Rax
1899 mov 248($context),%rbx # pull context->Rip
1901 mov 8($disp),%rsi # disp->ImageBase
1902 mov 56($disp),%r11 # disp->HandlerData
1904 mov 0(%r11),%r10d # HandlerData[0]
1905 lea (%rsi,%r10),%r10 # prologue label
1906 cmp %r10,%rbx # context->Rip<prologue label
1907 jb .Lcommon_seh_tail
1909 mov 152($context),%rax # pull context->Rsp
1911 mov 4(%r11),%r10d # HandlerData[1]
1912 lea (%rsi,%r10),%r10 # epilogue label
1913 cmp %r10,%rbx # context->Rip>=epilogue label
1914 jae .Lcommon_seh_tail
1916 mov 232($context),%rax # pull context->R14
1918 lea -40-6*16(%rax),%rsi
1919 lea 512($context),%rdi # &context.Xmm6
1921 .long 0xa548f3fc # cld; rep movsq
1928 mov %rbx,144($context) # restore context->Rbx
1929 mov %rbp,160($context) # restore context->Rbp
1930 mov %r12,216($context) # restore cotnext->R12
1931 mov %r13,224($context) # restore cotnext->R13
1932 mov %r14,232($context) # restore cotnext->R14
1937 mov %rax,152($context) # restore context->Rsp
1938 mov %rsi,168($context) # restore context->Rsi
1939 mov %rdi,176($context) # restore context->Rdi
1941 mov 40($disp),%rdi # disp->ContextRecord
1942 mov $context,%rsi # context
1943 mov \$154,%ecx # sizeof(CONTEXT)
1944 .long 0xa548f3fc # cld; rep movsq
1947 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
1948 mov 8(%rsi),%rdx # arg2, disp->ImageBase
1949 mov 0(%rsi),%r8 # arg3, disp->ControlPc
1950 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
1951 mov 40(%rsi),%r10 # disp->ContextRecord
1952 lea 56(%rsi),%r11 # &disp->HandlerData
1953 lea 24(%rsi),%r12 # &disp->EstablisherFrame
1954 mov %r10,32(%rsp) # arg5
1955 mov %r11,40(%rsp) # arg6
1956 mov %r12,48(%rsp) # arg7
1957 mov %rcx,56(%rsp) # arg8, (NULL)
1958 call *__imp_RtlVirtualUnwind(%rip)
1960 mov \$1,%eax # ExceptionContinueSearch
1972 .size ssse3_handler,.-ssse3_handler
1976 .rva .LSEH_begin_sha1_block_data_order
1977 .rva .LSEH_end_sha1_block_data_order
1978 .rva .LSEH_info_sha1_block_data_order
1980 $code.=<<___ if ($shaext);
1981 .rva .LSEH_begin_sha1_block_data_order_shaext
1982 .rva .LSEH_end_sha1_block_data_order_shaext
1983 .rva .LSEH_info_sha1_block_data_order_shaext
1986 .rva .LSEH_begin_sha1_block_data_order_ssse3
1987 .rva .LSEH_end_sha1_block_data_order_ssse3
1988 .rva .LSEH_info_sha1_block_data_order_ssse3
1990 $code.=<<___ if ($avx);
1991 .rva .LSEH_begin_sha1_block_data_order_avx
1992 .rva .LSEH_end_sha1_block_data_order_avx
1993 .rva .LSEH_info_sha1_block_data_order_avx
1995 $code.=<<___ if ($avx>1);
1996 .rva .LSEH_begin_sha1_block_data_order_avx2
1997 .rva .LSEH_end_sha1_block_data_order_avx2
1998 .rva .LSEH_info_sha1_block_data_order_avx2
2003 .LSEH_info_sha1_block_data_order:
2007 $code.=<<___ if ($shaext);
2008 .LSEH_info_sha1_block_data_order_shaext:
2013 .LSEH_info_sha1_block_data_order_ssse3:
2016 .rva .Lprologue_ssse3,.Lepilogue_ssse3 # HandlerData[]
2018 $code.=<<___ if ($avx);
2019 .LSEH_info_sha1_block_data_order_avx:
2022 .rva .Lprologue_avx,.Lepilogue_avx # HandlerData[]
2024 $code.=<<___ if ($avx>1);
2025 .LSEH_info_sha1_block_data_order_avx2:
2028 .rva .Lprologue_avx2,.Lepilogue_avx2 # HandlerData[]
2032 ####################################################################
2035 if (@_[0] =~ /\$([x0-9a-f]+),\s*%xmm([0-7]),\s*%xmm([0-7])/) {
2036 my @opcode=(0x0f,0x3a,0xcc);
2037 push @opcode,0xc0|($2&7)|(($3&7)<<3); # ModR/M
2039 push @opcode,$c=~/^0/?oct($c):$c;
2040 return ".byte\t".join(',',@opcode);
2042 return "sha1rnds4\t".@_[0];
2049 "sha1nexte" => 0xc8,
2051 "sha1msg2" => 0xca );
2053 if (defined($opcodelet{$instr}) && @_[0] =~ /%xmm([0-9]+),\s*%xmm([0-9]+)/) {
2054 my @opcode=(0x0f,0x38);
2056 $rex|=0x04 if ($2>=8);
2057 $rex|=0x01 if ($1>=8);
2058 unshift @opcode,0x40|$rex if ($rex);
2059 push @opcode,$opcodelet{$instr};
2060 push @opcode,0xc0|($1&7)|(($2&7)<<3); # ModR/M
2061 return ".byte\t".join(',',@opcode);
2063 return $instr."\t".@_[0];
2067 foreach (split("\n",$code)) {
2068 s/\`([^\`]*)\`/eval $1/geo;
2070 s/\b(sha1rnds4)\s+(.*)/sha1rnds4($2)/geo or
2071 s/\b(sha1[^\s]*)\s+(.*)/sha1op38($1,$2)/geo;