3 # ====================================================================
4 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
5 # project. The module is, however, dual licensed under OpenSSL and
6 # CRYPTOGAMS licenses depending on where you obtain it. For further
7 # details see http://www.openssl.org/~appro/cryptogams/.
8 # ====================================================================
10 # sha1_block procedure for x86_64.
12 # It was brought to my attention that on EM64T compiler-generated code
13 # was far behind 32-bit assembler implementation. This is unlike on
14 # Opteron where compiler-generated code was only 15% behind 32-bit
15 # assembler, which originally made it hard to motivate the effort.
16 # There was suggestion to mechanically translate 32-bit code, but I
17 # dismissed it, reasoning that x86_64 offers enough register bank
18 # capacity to fully utilize SHA-1 parallelism. Therefore this fresh
19 # implementation:-) However! While 64-bit code does perform better
20 # on Opteron, I failed to beat 32-bit assembler on EM64T core. Well,
21 # x86_64 does offer larger *addressable* bank, but out-of-order core
22 # reaches for even more registers through dynamic aliasing, and EM64T
23 # core must have managed to run-time optimize even 32-bit code just as
24 # good as 64-bit one. Performance improvement is summarized in the
27 # gcc 3.4 32-bit asm cycles/byte
28 # Opteron +45% +20% 6.8
29 # Xeon P4 +65% +0% 9.9
34 # The code was revised to minimize code size and to maximize
35 # "distance" between instructions producing input to 'lea'
36 # instruction and the 'lea' instruction itself, which is essential
37 # for Intel Atom core.
41 # Add SSSE3, Supplemental[!] SSE3, implementation. The idea behind it
42 # is to offload message schedule denoted by Wt in NIST specification,
43 # or Xupdate in OpenSSL source, to SIMD unit. See sha1-586.pl module
44 # for background and implementation details. The only difference from
45 # 32-bit code is that 64-bit code doesn't have to spill @X[] elements
46 # to free temporary registers.
50 # Add AVX code path. See sha1-586.pl for further information.
54 # Add AVX2+BMI code path. Initial attempt (utilizing BMI instructions
55 # and loading pair of consecutive blocks to 256-bit %ymm registers)
56 # did not provide impressive performance improvement till a crucial
57 # hint regarding the number of Xupdate iterations to pre-compute in
58 # advance was provided by Ilya Albrekht of Intel Corp.
60 ######################################################################
61 # Current performance is summarized in following table. Numbers are
62 # CPU clock cycles spent to process single byte (less is better).
67 # Core2 6.55 6.05/+8% -
68 # Westmere 6.73 5.30/+27% -
69 # Sandy Bridge 7.70 6.10/+26% 4.99/+54%
70 # Ivy Bridge 6.06 4.67/+30% 4.60/+32%
71 # Haswell 5.45 4.15/+31% 3.57/+53%
72 # Bulldozer 9.11 5.95/+53%
73 # VIA Nano 9.32 7.15/+30%
74 # Atom [10.5?] [9.23?]/+14%
75 # Silvermont 13.1(*) 9.37/+40%
77 # (*) obviously suboptimal result, nothing was done about it,
78 # because SSSE3 code is compiled unconditionally;
82 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
84 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
86 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
87 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
88 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
89 die "can't locate x86_64-xlate.pl";
91 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
92 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
93 $avx = ($1>=2.19) + ($1>=2.22);
96 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
97 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
98 $avx = ($1>=2.09) + ($1>=2.10);
101 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
102 `ml64 2>&1` =~ /Version ([0-9]+)\./) {
103 $avx = ($1>=10) + ($1>=11);
106 open OUT,"| \"$^X\" $xlate $flavour $output";
109 $ctx="%rdi"; # 1st arg
110 $inp="%rsi"; # 2nd arg
111 $num="%rdx"; # 3rd arg
113 # reassign arguments in order to produce more compact code
121 @xi=("%edx","%ebp","%r14d");
131 my ($i,$a,$b,$c,$d,$e)=@_;
133 $code.=<<___ if ($i==0);
134 mov `4*$i`($inp),$xi[0]
137 $code.=<<___ if ($i<15);
138 mov `4*$j`($inp),$xi[1]
140 mov $xi[0],`4*$i`(%rsp)
146 lea 0x5a827999($xi[0],$e),$e
152 $code.=<<___ if ($i>=15);
153 xor `4*($j%16)`(%rsp),$xi[1]
155 mov $xi[0],`4*($i%16)`(%rsp)
157 xor `4*(($j+2)%16)`(%rsp),$xi[1]
160 xor `4*(($j+8)%16)`(%rsp),$xi[1]
162 lea 0x5a827999($xi[0],$e),$e
169 push(@xi,shift(@xi));
173 my ($i,$a,$b,$c,$d,$e)=@_;
175 my $K=($i<40)?0x6ed9eba1:0xca62c1d6;
176 $code.=<<___ if ($i<79);
177 xor `4*($j%16)`(%rsp),$xi[1]
179 `"mov $xi[0],".4*($i%16)."(%rsp)" if ($i<72)`
181 xor `4*(($j+2)%16)`(%rsp),$xi[1]
184 xor `4*(($j+8)%16)`(%rsp),$xi[1]
192 $code.=<<___ if ($i==79);
203 push(@xi,shift(@xi));
207 my ($i,$a,$b,$c,$d,$e)=@_;
210 xor `4*($j%16)`(%rsp),$xi[1]
212 mov $xi[0],`4*($i%16)`(%rsp)
214 xor `4*(($j+2)%16)`(%rsp),$xi[1]
217 xor `4*(($j+8)%16)`(%rsp),$xi[1]
218 lea 0x8f1bbcdc($xi[0],$e),$e
228 push(@xi,shift(@xi));
233 .extern OPENSSL_ia32cap_P
235 .globl sha1_block_data_order
236 .type sha1_block_data_order,\@function,3
238 sha1_block_data_order:
239 mov OPENSSL_ia32cap_P+0(%rip),%r9d
240 mov OPENSSL_ia32cap_P+4(%rip),%r8d
241 mov OPENSSL_ia32cap_P+8(%rip),%r10d
242 test \$`1<<9`,%r8d # check SSSE3 bit
245 $code.=<<___ if ($avx>1);
246 and \$`1<<3|1<<5|1<<8`,%r10d # check AVX2+BMI1+BMI2
247 cmp \$`1<<3|1<<5|1<<8`,%r10d
250 $code.=<<___ if ($avx);
251 and \$`1<<28`,%r8d # mask AVX bit
252 and \$`1<<30`,%r9d # mask "Intel CPU" bit
254 cmp \$`1<<28|1<<30`,%r8d
268 mov %rdi,$ctx # reassigned argument
270 mov %rsi,$inp # reassigned argument
272 mov %rdx,$num # reassigned argument
273 mov %rax,`16*4`(%rsp)
286 for($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); }
287 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
288 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
289 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
303 lea `16*4`($inp),$inp
306 mov `16*4`(%rsp),%rsi
315 .size sha1_block_data_order,.-sha1_block_data_order
319 my @X=map("%xmm$_",(4..7,0..3));
320 my @Tx=map("%xmm$_",(8..10));
322 my @V=($A,$B,$C,$D,$E)=("%eax","%ebx","%ecx","%edx","%ebp"); # size optimization
323 my @T=("%esi","%edi");
328 my $_rol=sub { &rol(@_) };
329 my $_ror=sub { &ror(@_) };
335 jmp .Lalign32_$sn # see "Decoded ICache" in manual
343 .type sha1_block_data_order_ssse3,\@function,3
345 sha1_block_data_order_ssse3:
351 push %r13 # redundant, done to share Win64 SE handler
353 lea `-64-($win64?6*16:0)`(%rsp),%rsp
355 $code.=<<___ if ($win64);
356 movaps %xmm6,-40-6*16(%rax)
357 movaps %xmm7,-40-5*16(%rax)
358 movaps %xmm8,-40-4*16(%rax)
359 movaps %xmm9,-40-3*16(%rax)
360 movaps %xmm10,-40-2*16(%rax)
361 movaps %xmm11,-40-1*16(%rax)
365 mov %rax,%r14 # original %rsp
367 mov %rdi,$ctx # reassigned argument
368 mov %rsi,$inp # reassigned argument
369 mov %rdx,$num # reassigned argument
373 lea K_XX_XX+64(%rip),$K_XX_XX
375 mov 0($ctx),$A # load context
379 mov $B,@T[0] # magic seed
385 movdqa 64($K_XX_XX),@X[2] # pbswap mask
386 movdqa -64($K_XX_XX),@Tx[1] # K_00_19
387 movdqu 0($inp),@X[-4&7] # load input to %xmm[0-3]
388 movdqu 16($inp),@X[-3&7]
389 movdqu 32($inp),@X[-2&7]
390 movdqu 48($inp),@X[-1&7]
391 pshufb @X[2],@X[-4&7] # byte swap
392 pshufb @X[2],@X[-3&7]
393 pshufb @X[2],@X[-2&7]
395 paddd @Tx[1],@X[-4&7] # add K_00_19
396 pshufb @X[2],@X[-1&7]
397 paddd @Tx[1],@X[-3&7]
398 paddd @Tx[1],@X[-2&7]
399 movdqa @X[-4&7],0(%rsp) # X[]+K xfer to IALU
400 psubd @Tx[1],@X[-4&7] # restore X[]
401 movdqa @X[-3&7],16(%rsp)
402 psubd @Tx[1],@X[-3&7]
403 movdqa @X[-2&7],32(%rsp)
404 psubd @Tx[1],@X[-2&7]
408 sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm
409 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
411 $arg = "\$$arg" if ($arg*1 eq $arg);
412 $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
415 sub Xupdate_ssse3_16_31() # recall that $Xi starts wtih 4
418 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
421 eval(shift(@insns)); # ror
422 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
424 &movdqa (@Tx[0],@X[-1&7]);
425 &paddd (@Tx[1],@X[-1&7]);
429 &punpcklqdq(@X[0],@X[-3&7]); # compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8);
431 eval(shift(@insns)); # rol
433 &psrldq (@Tx[0],4); # "X[-3]", 3 dwords
437 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
439 eval(shift(@insns)); # ror
440 &pxor (@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
445 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
447 eval(shift(@insns)); # rol
448 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
452 &movdqa (@Tx[2],@X[0]);
455 eval(shift(@insns)); # ror
456 &movdqa (@Tx[0],@X[0]);
459 &pslldq (@Tx[2],12); # "X[0]"<<96, extract one dword
460 &paddd (@X[0],@X[0]);
466 eval(shift(@insns)); # rol
468 &movdqa (@Tx[1],@Tx[2]);
474 eval(shift(@insns)); # ror
475 &por (@X[0],@Tx[0]); # "X[0]"<<<=1
481 &pxor (@X[0],@Tx[2]);
483 &movdqa (@Tx[2],eval(2*16*(($Xi)/5)-64)."($K_XX_XX)"); # K_XX_XX
484 eval(shift(@insns)); # rol
488 &pxor (@X[0],@Tx[1]); # "X[0]"^=("X[0]">>96)<<<2
489 &pshufd (@Tx[1],@X[-1&7],0xee) if ($Xi==7); # was &movdqa (@Tx[0],@X[-1&7]) in Xupdate_ssse3_32_79
491 foreach (@insns) { eval; } # remaining instructions [if any]
493 $Xi++; push(@X,shift(@X)); # "rotate" X[]
494 push(@Tx,shift(@Tx));
497 sub Xupdate_ssse3_32_79()
500 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
503 eval(shift(@insns)) if ($Xi==8);
504 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
505 eval(shift(@insns)) if ($Xi==8);
506 eval(shift(@insns)); # body_20_39
508 eval(shift(@insns)) if (@insns[1] =~ /_ror/);
509 eval(shift(@insns)) if (@insns[0] =~ /_ror/);
510 &punpcklqdq(@Tx[0],@X[-1&7]); # compose "X[-6]", was &palignr(@Tx[0],@X[-2&7],8);
512 eval(shift(@insns)); # rol
514 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
518 &movdqa (@Tx[2],@Tx[1]);# "perpetuate" K_XX_XX...
519 } else { # ... or load next one
520 &movdqa (@Tx[2],eval(2*16*($Xi/5)-64)."($K_XX_XX)");
522 eval(shift(@insns)); # ror
523 &paddd (@Tx[1],@X[-1&7]);
526 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-6]"
527 eval(shift(@insns)); # body_20_39
530 eval(shift(@insns)); # rol
531 eval(shift(@insns)) if (@insns[0] =~ /_ror/);
533 &movdqa (@Tx[0],@X[0]);
536 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
537 eval(shift(@insns)); # ror
539 eval(shift(@insns)); # body_20_39
545 eval(shift(@insns)) if (@insns[0] =~ /_rol/);# rol
548 eval(shift(@insns)); # ror
550 &por (@X[0],@Tx[0]); # "X[0]"<<<=2
552 eval(shift(@insns)); # body_20_39
553 eval(shift(@insns)) if (@insns[1] =~ /_rol/);
554 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
555 &pshufd(@Tx[1],@X[-1&7],0xee) if ($Xi<19); # was &movdqa (@Tx[1],@X[0])
557 eval(shift(@insns)); # rol
560 eval(shift(@insns)); # rol
563 foreach (@insns) { eval; } # remaining instructions
565 $Xi++; push(@X,shift(@X)); # "rotate" X[]
566 push(@Tx,shift(@Tx));
569 sub Xuplast_ssse3_80()
572 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
579 &paddd (@Tx[1],@X[-1&7]);
583 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU
585 foreach (@insns) { eval; } # remaining instructions
588 &je (".Ldone_ssse3");
590 unshift(@Tx,pop(@Tx));
592 &movdqa (@X[2],"64($K_XX_XX)"); # pbswap mask
593 &movdqa (@Tx[1],"-64($K_XX_XX)"); # K_00_19
594 &movdqu (@X[-4&7],"0($inp)"); # load input
595 &movdqu (@X[-3&7],"16($inp)");
596 &movdqu (@X[-2&7],"32($inp)");
597 &movdqu (@X[-1&7],"48($inp)");
598 &pshufb (@X[-4&7],@X[2]); # byte swap
607 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
613 &pshufb (@X[($Xi-3)&7],@X[2]);
618 &paddd (@X[($Xi-4)&7],@Tx[1]);
623 &movdqa (eval(16*$Xi)."(%rsp)",@X[($Xi-4)&7]); # X[]+K xfer to IALU
628 &psubd (@X[($Xi-4)&7],@Tx[1]);
630 foreach (@insns) { eval; }
637 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
640 foreach (@insns) { eval; }
643 sub body_00_19 () { # ((c^d)&b)^d
644 # on start @T[0]=(c^d)&b
645 return &body_20_39() if ($rx==19); $rx++;
647 '($a,$b,$c,$d,$e)=@V;'.
648 '&$_ror ($b,$j?7:2)', # $b>>>2
650 '&mov (@T[1],$a)', # $b for next round
652 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
653 '&xor ($b,$c)', # $c^$d for next round
657 '&and (@T[1],$b)', # ($b&($c^$d)) for next round
659 '&xor ($b,$c)', # restore $b
660 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
664 sub body_20_39 () { # b^d^c
666 return &body_40_59() if ($rx==39); $rx++;
668 '($a,$b,$c,$d,$e)=@V;'.
669 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
670 '&xor (@T[0],$d) if($j==19);'.
671 '&xor (@T[0],$c) if($j> 19)', # ($b^$d^$c)
672 '&mov (@T[1],$a)', # $b for next round
676 '&xor (@T[1],$c) if ($j< 79)', # $b^$d for next round
678 '&$_ror ($b,7)', # $b>>>2
679 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
683 sub body_40_59 () { # ((b^c)&(c^d))^c
684 # on entry @T[0]=(b^c), (c^=d)
687 '($a,$b,$c,$d,$e)=@V;'.
688 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
689 '&and (@T[0],$c) if ($j>=40)', # (b^c)&(c^d)
690 '&xor ($c,$d) if ($j>=40)', # restore $c
692 '&$_ror ($b,7)', # $b>>>2
693 '&mov (@T[1],$a)', # $b for next round
698 '&xor (@T[1],$c) if ($j==59);'.
699 '&xor (@T[1],$b) if ($j< 59)', # b^c for next round
701 '&xor ($b,$c) if ($j< 59)', # c^d for next round
702 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
709 &Xupdate_ssse3_16_31(\&body_00_19);
710 &Xupdate_ssse3_16_31(\&body_00_19);
711 &Xupdate_ssse3_16_31(\&body_00_19);
712 &Xupdate_ssse3_16_31(\&body_00_19);
713 &Xupdate_ssse3_32_79(\&body_00_19);
714 &Xupdate_ssse3_32_79(\&body_20_39);
715 &Xupdate_ssse3_32_79(\&body_20_39);
716 &Xupdate_ssse3_32_79(\&body_20_39);
717 &Xupdate_ssse3_32_79(\&body_20_39);
718 &Xupdate_ssse3_32_79(\&body_20_39);
719 &Xupdate_ssse3_32_79(\&body_40_59);
720 &Xupdate_ssse3_32_79(\&body_40_59);
721 &Xupdate_ssse3_32_79(\&body_40_59);
722 &Xupdate_ssse3_32_79(\&body_40_59);
723 &Xupdate_ssse3_32_79(\&body_40_59);
724 &Xupdate_ssse3_32_79(\&body_20_39);
725 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done"
727 $saved_j=$j; @saved_V=@V;
729 &Xloop_ssse3(\&body_20_39);
730 &Xloop_ssse3(\&body_20_39);
731 &Xloop_ssse3(\&body_20_39);
734 add 0($ctx),$A # update context
741 mov @T[0],$B # magic seed
753 $j=$saved_j; @V=@saved_V;
755 &Xtail_ssse3(\&body_20_39);
756 &Xtail_ssse3(\&body_20_39);
757 &Xtail_ssse3(\&body_20_39);
760 add 0($ctx),$A # update context
771 $code.=<<___ if ($win64);
772 movaps -40-6*16(%r14),%xmm6
773 movaps -40-5*16(%r14),%xmm7
774 movaps -40-4*16(%r14),%xmm8
775 movaps -40-3*16(%r14),%xmm9
776 movaps -40-2*16(%r14),%xmm10
777 movaps -40-1*16(%r14),%xmm11
789 .size sha1_block_data_order_ssse3,.-sha1_block_data_order_ssse3
793 $Xi=4; # reset variables
794 @X=map("%xmm$_",(4..7,0..3));
795 @Tx=map("%xmm$_",(8..10));
799 my $done_avx_label=".Ldone_avx";
801 my $_rol=sub { &shld(@_[0],@_) };
802 my $_ror=sub { &shrd(@_[0],@_) };
805 .type sha1_block_data_order_avx,\@function,3
807 sha1_block_data_order_avx:
813 push %r13 # redundant, done to share Win64 SE handler
815 lea `-64-($win64?6*16:0)`(%rsp),%rsp
818 $code.=<<___ if ($win64);
819 vmovaps %xmm6,-40-6*16(%rax)
820 vmovaps %xmm7,-40-5*16(%rax)
821 vmovaps %xmm8,-40-4*16(%rax)
822 vmovaps %xmm9,-40-3*16(%rax)
823 vmovaps %xmm10,-40-2*16(%rax)
824 vmovaps %xmm11,-40-1*16(%rax)
828 mov %rax,%r14 # original %rsp
830 mov %rdi,$ctx # reassigned argument
831 mov %rsi,$inp # reassigned argument
832 mov %rdx,$num # reassigned argument
836 lea K_XX_XX+64(%rip),$K_XX_XX
838 mov 0($ctx),$A # load context
842 mov $B,@T[0] # magic seed
848 vmovdqa 64($K_XX_XX),@X[2] # pbswap mask
849 vmovdqa -64($K_XX_XX),$Kx # K_00_19
850 vmovdqu 0($inp),@X[-4&7] # load input to %xmm[0-3]
851 vmovdqu 16($inp),@X[-3&7]
852 vmovdqu 32($inp),@X[-2&7]
853 vmovdqu 48($inp),@X[-1&7]
854 vpshufb @X[2],@X[-4&7],@X[-4&7] # byte swap
856 vpshufb @X[2],@X[-3&7],@X[-3&7]
857 vpshufb @X[2],@X[-2&7],@X[-2&7]
858 vpshufb @X[2],@X[-1&7],@X[-1&7]
859 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19
860 vpaddd $Kx,@X[-3&7],@X[1]
861 vpaddd $Kx,@X[-2&7],@X[2]
862 vmovdqa @X[0],0(%rsp) # X[]+K xfer to IALU
863 vmovdqa @X[1],16(%rsp)
864 vmovdqa @X[2],32(%rsp)
868 sub Xupdate_avx_16_31() # recall that $Xi starts wtih 4
871 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
876 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
880 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
883 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
886 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
890 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
896 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
899 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
903 &vpsrld (@Tx[0],@X[0],31);
909 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
910 &vpaddd (@X[0],@X[0],@X[0]);
916 &vpsrld (@Tx[1],@Tx[2],30);
917 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
923 &vpslld (@Tx[2],@Tx[2],2);
924 &vpxor (@X[0],@X[0],@Tx[1]);
930 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
933 &vmovdqa ($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
938 foreach (@insns) { eval; } # remaining instructions [if any]
940 $Xi++; push(@X,shift(@X)); # "rotate" X[]
943 sub Xupdate_avx_32_79()
946 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
949 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
950 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
951 eval(shift(@insns)); # body_20_39
954 eval(shift(@insns)); # rol
956 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
958 eval(shift(@insns)) if (@insns[0] !~ /&ro[rl]/);
959 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
960 &vmovdqa ($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0);
961 eval(shift(@insns)); # ror
964 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]"
965 eval(shift(@insns)); # body_20_39
968 eval(shift(@insns)); # rol
970 &vpsrld (@Tx[0],@X[0],30);
971 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
974 eval(shift(@insns)); # ror
977 &vpslld (@X[0],@X[0],2);
978 eval(shift(@insns)); # body_20_39
981 eval(shift(@insns)); # rol
984 eval(shift(@insns)); # ror
987 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2
988 eval(shift(@insns)); # body_20_39
991 eval(shift(@insns)); # rol
994 eval(shift(@insns)); # rol
997 foreach (@insns) { eval; } # remaining instructions
999 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1002 sub Xuplast_avx_80()
1005 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1006 my ($a,$b,$c,$d,$e);
1008 eval(shift(@insns));
1009 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1010 eval(shift(@insns));
1011 eval(shift(@insns));
1012 eval(shift(@insns));
1013 eval(shift(@insns));
1015 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU
1017 foreach (@insns) { eval; } # remaining instructions
1020 &je ($done_avx_label);
1022 &vmovdqa(@X[2],"64($K_XX_XX)"); # pbswap mask
1023 &vmovdqa($Kx,"-64($K_XX_XX)"); # K_00_19
1024 &vmovdqu(@X[-4&7],"0($inp)"); # load input
1025 &vmovdqu(@X[-3&7],"16($inp)");
1026 &vmovdqu(@X[-2&7],"32($inp)");
1027 &vmovdqu(@X[-1&7],"48($inp)");
1028 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1037 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1038 my ($a,$b,$c,$d,$e);
1040 eval(shift(@insns));
1041 eval(shift(@insns));
1042 &vpshufb(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1043 eval(shift(@insns));
1044 eval(shift(@insns));
1045 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],$Kx);
1046 eval(shift(@insns));
1047 eval(shift(@insns));
1048 eval(shift(@insns));
1049 eval(shift(@insns));
1050 &vmovdqa(eval(16*$Xi)."(%rsp)",@X[$Xi&7]); # X[]+K xfer to IALU
1051 eval(shift(@insns));
1052 eval(shift(@insns));
1054 foreach (@insns) { eval; }
1061 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1062 my ($a,$b,$c,$d,$e);
1064 foreach (@insns) { eval; }
1071 &Xupdate_avx_16_31(\&body_00_19);
1072 &Xupdate_avx_16_31(\&body_00_19);
1073 &Xupdate_avx_16_31(\&body_00_19);
1074 &Xupdate_avx_16_31(\&body_00_19);
1075 &Xupdate_avx_32_79(\&body_00_19);
1076 &Xupdate_avx_32_79(\&body_20_39);
1077 &Xupdate_avx_32_79(\&body_20_39);
1078 &Xupdate_avx_32_79(\&body_20_39);
1079 &Xupdate_avx_32_79(\&body_20_39);
1080 &Xupdate_avx_32_79(\&body_20_39);
1081 &Xupdate_avx_32_79(\&body_40_59);
1082 &Xupdate_avx_32_79(\&body_40_59);
1083 &Xupdate_avx_32_79(\&body_40_59);
1084 &Xupdate_avx_32_79(\&body_40_59);
1085 &Xupdate_avx_32_79(\&body_40_59);
1086 &Xupdate_avx_32_79(\&body_20_39);
1087 &Xuplast_avx_80(\&body_20_39); # can jump to "done"
1089 $saved_j=$j; @saved_V=@V;
1091 &Xloop_avx(\&body_20_39);
1092 &Xloop_avx(\&body_20_39);
1093 &Xloop_avx(\&body_20_39);
1096 add 0($ctx),$A # update context
1103 mov @T[0],$B # magic seed
1115 $j=$saved_j; @V=@saved_V;
1117 &Xtail_avx(\&body_20_39);
1118 &Xtail_avx(\&body_20_39);
1119 &Xtail_avx(\&body_20_39);
1124 add 0($ctx),$A # update context
1135 $code.=<<___ if ($win64);
1136 movaps -40-6*16(%r14),%xmm6
1137 movaps -40-5*16(%r14),%xmm7
1138 movaps -40-4*16(%r14),%xmm8
1139 movaps -40-3*16(%r14),%xmm9
1140 movaps -40-2*16(%r14),%xmm10
1141 movaps -40-1*16(%r14),%xmm11
1153 .size sha1_block_data_order_avx,.-sha1_block_data_order_avx
1158 $Xi=4; # reset variables
1159 @X=map("%ymm$_",(4..7,0..3));
1160 @Tx=map("%ymm$_",(8..10));
1164 my @ROTX=("%eax","%ebp","%ebx","%ecx","%edx","%esi");
1165 my ($a5,$t0)=("%r12d","%edi");
1167 my ($A,$F,$B,$C,$D,$E)=@ROTX;
1172 .type sha1_block_data_order_avx2,\@function,3
1174 sha1_block_data_order_avx2:
1184 $code.=<<___ if ($win64);
1185 lea -6*16(%rsp),%rsp
1186 vmovaps %xmm6,-40-6*16(%rax)
1187 vmovaps %xmm7,-40-5*16(%rax)
1188 vmovaps %xmm8,-40-4*16(%rax)
1189 vmovaps %xmm9,-40-3*16(%rax)
1190 vmovaps %xmm10,-40-2*16(%rax)
1191 vmovaps %xmm11,-40-1*16(%rax)
1195 mov %rax,%r14 # original %rsp
1196 mov %rdi,$ctx # reassigned argument
1197 mov %rsi,$inp # reassigned argument
1198 mov %rdx,$num # reassigned argument
1205 lea K_XX_XX+64(%rip),$K_XX_XX
1207 mov 0($ctx),$A # load context
1209 cmovae $inp,$frame # next or same block
1214 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask
1216 vmovdqu ($inp),%xmm0
1217 vmovdqu 16($inp),%xmm1
1218 vmovdqu 32($inp),%xmm2
1219 vmovdqu 48($inp),%xmm3
1221 vinserti128 \$1,($frame),@X[-4&7],@X[-4&7]
1222 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7]
1223 vpshufb @X[2],@X[-4&7],@X[-4&7]
1224 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7]
1225 vpshufb @X[2],@X[-3&7],@X[-3&7]
1226 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7]
1227 vpshufb @X[2],@X[-2&7],@X[-2&7]
1228 vmovdqu -64($K_XX_XX),$Kx # K_00_19
1229 vpshufb @X[2],@X[-1&7],@X[-1&7]
1231 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19
1232 vpaddd $Kx,@X[-3&7],@X[1]
1233 vmovdqu @X[0],0(%rsp) # X[]+K xfer to IALU
1234 vpaddd $Kx,@X[-2&7],@X[2]
1235 vmovdqu @X[1],32(%rsp)
1236 vpaddd $Kx,@X[-1&7],@X[3]
1237 vmovdqu @X[2],64(%rsp)
1238 vmovdqu @X[3],96(%rsp)
1240 for (;$Xi<8;$Xi++) { # Xupdate_avx2_16_31
1243 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1244 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1245 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1246 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1247 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1248 &vpsrld (@Tx[0],@X[0],31);
1249 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1250 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1251 &vpaddd (@X[0],@X[0],@X[0]);
1252 &vpsrld (@Tx[1],@Tx[2],30);
1253 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1254 &vpslld (@Tx[2],@Tx[2],2);
1255 &vpxor (@X[0],@X[0],@Tx[1]);
1256 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1257 &vpaddd (@Tx[1],@X[0],$Kx);
1258 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1260 push(@X,shift(@X)); # "rotate" X[]
1263 lea 128(%rsp),$frame
1272 sub bodyx_00_19 () { # 8 instructions, 3 cycles critical path
1273 # at start $f=(b&c)^(~b&d), $b>>>=2
1274 return &bodyx_20_39() if ($rx==19); $rx++;
1276 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1278 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1279 '&lea ($frame,"256($frame)") if ($j%32==31);',
1280 '&andn ($t0,$a,$c)', # ~b&d for next round
1282 '&add ($e,$f)', # e+=(b&c)^(~b&d)
1283 '&rorx ($a5,$a,27)', # a<<<5
1284 '&rorx ($f,$a,2)', # b>>>2 for next round
1285 '&and ($a,$b)', # b&c for next round
1287 '&add ($e,$a5)', # e+=a<<<5
1288 '&xor ($a,$t0);'. # f=(b&c)^(~b&d) for next round
1290 'unshift(@ROTX,pop(@ROTX)); $j++;'
1294 sub bodyx_20_39 () { # 7 instructions, 2 cycles critical path
1295 # on entry $f=b^c^d, $b>>>=2
1296 return &bodyx_40_59() if ($rx==39); $rx++;
1298 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1300 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1301 '&lea ($frame,"256($frame)") if ($j%32==31);',
1303 '&lea ($e,"($e,$f)")', # e+=b^c^d
1304 '&rorx ($a5,$a,27)', # a<<<5
1305 '&rorx ($f,$a,2) if ($j<79)', # b>>>2 in next round
1306 '&xor ($a,$b) if ($j<79)', # b^c for next round
1308 '&add ($e,$a5)', # e+=a<<<5
1309 '&xor ($a,$c) if ($j<79);'. # f=b^c^d for next round
1311 'unshift(@ROTX,pop(@ROTX)); $j++;'
1315 sub bodyx_40_59 () { # 10 instructions, 3 cycles critical path
1316 # on entry $f=((b^c)&(c^d)), $b>>>=2
1319 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1321 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1322 '&lea ($frame,"256($frame)") if ($j%32==31);',
1323 '&xor ($f,$c) if ($j>39)', # (b^c)&(c^d)^c
1324 '&mov ($t0,$b) if ($j<59)', # count on zero latency
1325 '&xor ($t0,$c) if ($j<59)', # c^d for next round
1327 '&lea ($e,"($e,$f)")', # e+=(b^c)&(c^d)^c
1328 '&rorx ($a5,$a,27)', # a<<<5
1329 '&rorx ($f,$a,2)', # b>>>2 in next round
1330 '&xor ($a,$b)', # b^c for next round
1332 '&add ($e,$a5)', # e+=a<<<5
1333 '&and ($a,$t0) if ($j< 59);'. # f=(b^c)&(c^d) for next round
1334 '&xor ($a,$c) if ($j==59);'. # f=b^c^d for next round
1336 'unshift(@ROTX,pop(@ROTX)); $j++;'
1340 sub Xupdate_avx2_16_31() # recall that $Xi starts wtih 4
1343 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 instructions
1344 my ($a,$b,$c,$d,$e);
1346 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1347 eval(shift(@insns));
1348 eval(shift(@insns));
1349 eval(shift(@insns));
1350 eval(shift(@insns));
1352 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1353 eval(shift(@insns));
1354 eval(shift(@insns));
1355 eval(shift(@insns));
1357 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1358 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1359 eval(shift(@insns));
1360 eval(shift(@insns));
1362 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1363 eval(shift(@insns));
1364 eval(shift(@insns));
1365 eval(shift(@insns));
1366 eval(shift(@insns));
1368 &vpsrld (@Tx[0],@X[0],31);
1369 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1370 eval(shift(@insns));
1371 eval(shift(@insns));
1372 eval(shift(@insns));
1374 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1375 &vpaddd (@X[0],@X[0],@X[0]);
1376 eval(shift(@insns));
1377 eval(shift(@insns));
1379 &vpsrld (@Tx[1],@Tx[2],30);
1380 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1381 eval(shift(@insns));
1382 eval(shift(@insns));
1384 &vpslld (@Tx[2],@Tx[2],2);
1385 &vpxor (@X[0],@X[0],@Tx[1]);
1386 eval(shift(@insns));
1387 eval(shift(@insns));
1389 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1390 eval(shift(@insns));
1391 eval(shift(@insns));
1392 eval(shift(@insns));
1394 &vpaddd (@Tx[1],@X[0],$Kx);
1395 eval(shift(@insns));
1396 eval(shift(@insns));
1397 eval(shift(@insns));
1398 &vmovdqu(eval(32*($Xi))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1400 foreach (@insns) { eval; } # remaining instructions [if any]
1403 push(@X,shift(@X)); # "rotate" X[]
1406 sub Xupdate_avx2_32_79()
1409 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 to 50 instructions
1410 my ($a,$b,$c,$d,$e);
1412 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1413 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1414 eval(shift(@insns));
1415 eval(shift(@insns));
1417 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1418 &vmovdqu($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0);
1419 eval(shift(@insns));
1420 eval(shift(@insns));
1421 eval(shift(@insns));
1423 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]"
1424 eval(shift(@insns));
1425 eval(shift(@insns));
1426 eval(shift(@insns));
1428 &vpsrld (@Tx[0],@X[0],30);
1429 &vpslld (@X[0],@X[0],2);
1430 eval(shift(@insns));
1431 eval(shift(@insns));
1432 eval(shift(@insns));
1434 #&vpslld (@X[0],@X[0],2);
1435 eval(shift(@insns));
1436 eval(shift(@insns));
1437 eval(shift(@insns));
1439 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2
1440 eval(shift(@insns));
1441 eval(shift(@insns));
1442 eval(shift(@insns));
1443 eval(shift(@insns));
1445 &vpaddd (@Tx[1],@X[0],$Kx);
1446 eval(shift(@insns));
1447 eval(shift(@insns));
1448 eval(shift(@insns));
1449 eval(shift(@insns));
1451 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1453 foreach (@insns) { eval; } # remaining instructions
1456 push(@X,shift(@X)); # "rotate" X[]
1462 my @insns = (&$body,&$body,&$body,&$body,&$body); # 32 instructions
1463 my ($a,$b,$c,$d,$e);
1465 foreach (@insns) { eval; }
1469 &Xupdate_avx2_32_79(\&bodyx_00_19);
1470 &Xupdate_avx2_32_79(\&bodyx_00_19);
1471 &Xupdate_avx2_32_79(\&bodyx_00_19);
1472 &Xupdate_avx2_32_79(\&bodyx_00_19);
1474 &Xupdate_avx2_32_79(\&bodyx_20_39);
1475 &Xupdate_avx2_32_79(\&bodyx_20_39);
1476 &Xupdate_avx2_32_79(\&bodyx_20_39);
1477 &Xupdate_avx2_32_79(\&bodyx_20_39);
1480 &Xupdate_avx2_32_79(\&bodyx_40_59);
1481 &Xupdate_avx2_32_79(\&bodyx_40_59);
1482 &Xupdate_avx2_32_79(\&bodyx_40_59);
1483 &Xupdate_avx2_32_79(\&bodyx_40_59);
1485 &Xloop_avx2(\&bodyx_20_39);
1486 &Xloop_avx2(\&bodyx_20_39);
1487 &Xloop_avx2(\&bodyx_20_39);
1488 &Xloop_avx2(\&bodyx_20_39);
1491 lea 128($inp),$frame
1492 lea 128($inp),%rdi # borrow $t0
1494 cmovae $inp,$frame # next or previous block
1496 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c
1497 add 0($ctx),@ROTX[0] # update context
1498 add 4($ctx),@ROTX[1]
1499 add 8($ctx),@ROTX[3]
1500 mov @ROTX[0],0($ctx)
1501 add 12($ctx),@ROTX[4]
1502 mov @ROTX[1],4($ctx)
1503 mov @ROTX[0],$A # A=d
1504 add 16($ctx),@ROTX[5]
1506 mov @ROTX[3],8($ctx)
1507 mov @ROTX[4],$D # D=b
1508 #xchg @ROTX[5],$F # F=c, C=f
1509 mov @ROTX[4],12($ctx)
1510 mov @ROTX[1],$F # F=e
1511 mov @ROTX[5],16($ctx)
1513 mov @ROTX[5],$E # E=c
1515 #xchg $F,$E # E=c, F=e
1521 $Xi=4; # reset variables
1522 @X=map("%ymm$_",(4..7,0..3));
1525 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask
1526 cmp $num,%rdi # borrowed $t0
1529 vmovdqu -64(%rdi),%xmm0 # low part of @X[-4&7]
1530 vmovdqu -48(%rdi),%xmm1
1531 vmovdqu -32(%rdi),%xmm2
1532 vmovdqu -16(%rdi),%xmm3
1533 vinserti128 \$1,0($frame),@X[-4&7],@X[-4&7]
1534 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7]
1535 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7]
1536 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7]
1541 lea 128+16(%rsp),$frame
1548 $rx=$j=0; @ROTX=($A,$F,$B,$C,$D,$E);
1550 &Xloop_avx2 (\&bodyx_00_19);
1551 &Xloop_avx2 (\&bodyx_00_19);
1552 &Xloop_avx2 (\&bodyx_00_19);
1553 &Xloop_avx2 (\&bodyx_00_19);
1555 &Xloop_avx2 (\&bodyx_20_39);
1556 &vmovdqu ($Kx,"-64($K_XX_XX)"); # K_00_19
1557 &vpshufb (@X[-4&7],@X[-4&7],@X[2]); # byte swap
1558 &Xloop_avx2 (\&bodyx_20_39);
1559 &vpshufb (@X[-3&7],@X[-3&7],@X[2]);
1560 &vpaddd (@Tx[0],@X[-4&7],$Kx); # add K_00_19
1561 &Xloop_avx2 (\&bodyx_20_39);
1562 &vmovdqu ("0(%rsp)",@Tx[0]);
1563 &vpshufb (@X[-2&7],@X[-2&7],@X[2]);
1564 &vpaddd (@Tx[1],@X[-3&7],$Kx);
1565 &Xloop_avx2 (\&bodyx_20_39);
1566 &vmovdqu ("32(%rsp)",@Tx[1]);
1567 &vpshufb (@X[-1&7],@X[-1&7],@X[2]);
1568 &vpaddd (@X[2],@X[-2&7],$Kx);
1570 &Xloop_avx2 (\&bodyx_40_59);
1572 &vmovdqu ("64(%rsp)",@X[2]);
1573 &vpaddd (@X[3],@X[-1&7],$Kx);
1574 &Xloop_avx2 (\&bodyx_40_59);
1575 &vmovdqu ("96(%rsp)",@X[3]);
1576 &Xloop_avx2 (\&bodyx_40_59);
1577 &Xupdate_avx2_16_31(\&bodyx_40_59);
1579 &Xupdate_avx2_16_31(\&bodyx_20_39);
1580 &Xupdate_avx2_16_31(\&bodyx_20_39);
1581 &Xupdate_avx2_16_31(\&bodyx_20_39);
1582 &Xloop_avx2 (\&bodyx_20_39);
1585 lea 128(%rsp),$frame
1587 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c
1588 add 0($ctx),@ROTX[0] # update context
1589 add 4($ctx),@ROTX[1]
1590 add 8($ctx),@ROTX[3]
1591 mov @ROTX[0],0($ctx)
1592 add 12($ctx),@ROTX[4]
1593 mov @ROTX[1],4($ctx)
1594 mov @ROTX[0],$A # A=d
1595 add 16($ctx),@ROTX[5]
1597 mov @ROTX[3],8($ctx)
1598 mov @ROTX[4],$D # D=b
1599 #xchg @ROTX[5],$F # F=c, C=f
1600 mov @ROTX[4],12($ctx)
1601 mov @ROTX[1],$F # F=e
1602 mov @ROTX[5],16($ctx)
1604 mov @ROTX[5],$E # E=c
1606 #xchg $F,$E # E=c, F=e
1614 $code.=<<___ if ($win64);
1615 movaps -40-6*16(%r14),%xmm6
1616 movaps -40-5*16(%r14),%xmm7
1617 movaps -40-4*16(%r14),%xmm8
1618 movaps -40-3*16(%r14),%xmm9
1619 movaps -40-2*16(%r14),%xmm10
1620 movaps -40-1*16(%r14),%xmm11
1632 .size sha1_block_data_order_avx2,.-sha1_block_data_order_avx2
1639 .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19
1640 .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19
1641 .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39
1642 .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39
1643 .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59
1644 .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59
1645 .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79
1646 .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79
1647 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask
1648 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask
1652 .asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1656 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1657 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
1665 .extern __imp_RtlVirtualUnwind
1666 .type se_handler,\@abi-omnipotent
1680 mov 120($context),%rax # pull context->Rax
1681 mov 248($context),%rbx # pull context->Rip
1683 lea .Lprologue(%rip),%r10
1684 cmp %r10,%rbx # context->Rip<.Lprologue
1685 jb .Lcommon_seh_tail
1687 mov 152($context),%rax # pull context->Rsp
1689 lea .Lepilogue(%rip),%r10
1690 cmp %r10,%rbx # context->Rip>=.Lepilogue
1691 jae .Lcommon_seh_tail
1693 mov `16*4`(%rax),%rax # pull saved stack pointer
1700 mov %rbx,144($context) # restore context->Rbx
1701 mov %rbp,160($context) # restore context->Rbp
1702 mov %r12,216($context) # restore context->R12
1703 mov %r13,224($context) # restore context->R13
1704 mov %r14,232($context) # restore context->R14
1706 jmp .Lcommon_seh_tail
1707 .size se_handler,.-se_handler
1709 .type ssse3_handler,\@abi-omnipotent
1723 mov 120($context),%rax # pull context->Rax
1724 mov 248($context),%rbx # pull context->Rip
1726 mov 8($disp),%rsi # disp->ImageBase
1727 mov 56($disp),%r11 # disp->HandlerData
1729 mov 0(%r11),%r10d # HandlerData[0]
1730 lea (%rsi,%r10),%r10 # prologue label
1731 cmp %r10,%rbx # context->Rip<prologue label
1732 jb .Lcommon_seh_tail
1734 mov 152($context),%rax # pull context->Rsp
1736 mov 4(%r11),%r10d # HandlerData[1]
1737 lea (%rsi,%r10),%r10 # epilogue label
1738 cmp %r10,%rbx # context->Rip>=epilogue label
1739 jae .Lcommon_seh_tail
1741 mov 232($context),%rax # pull context->R14
1743 lea -40-6*16(%rax),%rsi
1744 lea 512($context),%rdi # &context.Xmm6
1746 .long 0xa548f3fc # cld; rep movsq
1753 mov %rbx,144($context) # restore context->Rbx
1754 mov %rbp,160($context) # restore context->Rbp
1755 mov %r12,216($context) # restore cotnext->R12
1756 mov %r13,224($context) # restore cotnext->R13
1757 mov %r14,232($context) # restore cotnext->R14
1762 mov %rax,152($context) # restore context->Rsp
1763 mov %rsi,168($context) # restore context->Rsi
1764 mov %rdi,176($context) # restore context->Rdi
1766 mov 40($disp),%rdi # disp->ContextRecord
1767 mov $context,%rsi # context
1768 mov \$154,%ecx # sizeof(CONTEXT)
1769 .long 0xa548f3fc # cld; rep movsq
1772 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
1773 mov 8(%rsi),%rdx # arg2, disp->ImageBase
1774 mov 0(%rsi),%r8 # arg3, disp->ControlPc
1775 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
1776 mov 40(%rsi),%r10 # disp->ContextRecord
1777 lea 56(%rsi),%r11 # &disp->HandlerData
1778 lea 24(%rsi),%r12 # &disp->EstablisherFrame
1779 mov %r10,32(%rsp) # arg5
1780 mov %r11,40(%rsp) # arg6
1781 mov %r12,48(%rsp) # arg7
1782 mov %rcx,56(%rsp) # arg8, (NULL)
1783 call *__imp_RtlVirtualUnwind(%rip)
1785 mov \$1,%eax # ExceptionContinueSearch
1797 .size ssse3_handler,.-ssse3_handler
1801 .rva .LSEH_begin_sha1_block_data_order
1802 .rva .LSEH_end_sha1_block_data_order
1803 .rva .LSEH_info_sha1_block_data_order
1804 .rva .LSEH_begin_sha1_block_data_order_ssse3
1805 .rva .LSEH_end_sha1_block_data_order_ssse3
1806 .rva .LSEH_info_sha1_block_data_order_ssse3
1808 $code.=<<___ if ($avx);
1809 .rva .LSEH_begin_sha1_block_data_order_avx
1810 .rva .LSEH_end_sha1_block_data_order_avx
1811 .rva .LSEH_info_sha1_block_data_order_avx
1813 $code.=<<___ if ($avx>1);
1814 .rva .LSEH_begin_sha1_block_data_order_avx2
1815 .rva .LSEH_end_sha1_block_data_order_avx2
1816 .rva .LSEH_info_sha1_block_data_order_avx2
1821 .LSEH_info_sha1_block_data_order:
1824 .LSEH_info_sha1_block_data_order_ssse3:
1827 .rva .Lprologue_ssse3,.Lepilogue_ssse3 # HandlerData[]
1829 $code.=<<___ if ($avx);
1830 .LSEH_info_sha1_block_data_order_avx:
1833 .rva .Lprologue_avx,.Lepilogue_avx # HandlerData[]
1835 $code.=<<___ if ($avx>1);
1836 .LSEH_info_sha1_block_data_order_avx2:
1839 .rva .Lprologue_avx2,.Lepilogue_avx2 # HandlerData[]
1843 ####################################################################
1845 $code =~ s/\`([^\`]*)\`/eval $1/gem;