2 # Copyright 2006-2016 The OpenSSL Project Authors. All Rights Reserved.
4 # Licensed under the Apache License 2.0 (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22 # assembler, which originally made it hard to motivate the effort.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 # Ryzen 4.75 3.80/+24% 1.93/+150%(**)
86 # VIA Nano 9.32 7.15/+30%
88 # Silvermont 13.1(*) 9.37/+40%
89 # Knights L 13.2(*) 9.68/+36% 8.30/+59%
90 # Goldmont 8.13 6.42/+27% 1.70/+380%(**)
92 # (*) obviously suboptimal result, nothing was done about it,
93 # because SSSE3 code is compiled unconditionally;
96 # $output is the last argument if it looks like a file (it has an extension)
97 # $flavour is the first argument if it doesn't look like a file
98 $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
99 $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
101 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
103 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
104 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
105 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
106 die "can't locate x86_64-xlate.pl";
108 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
109 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
110 $avx = ($1>=2.19) + ($1>=2.22);
113 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
114 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
115 $avx = ($1>=2.09) + ($1>=2.10);
118 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
119 `ml64 2>&1` =~ /Version ([0-9]+)\./) {
120 $avx = ($1>=10) + ($1>=11);
123 if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([0-9]+\.[0-9]+)/) {
124 $avx = ($2>=3.0) + ($2>3.0);
127 $shaext=1; ### set to zero if compiling for 1.0.1
128 $avx=1 if (!$shaext && $avx);
130 open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\""
131 or die "can't call $xlate: $!";
134 $ctx="%rdi"; # 1st arg
135 $inp="%rsi"; # 2nd arg
136 $num="%rdx"; # 3rd arg
138 # reassign arguments in order to produce more compact code
146 @xi=("%edx","%ebp","%r14d");
156 my ($i,$a,$b,$c,$d,$e)=@_;
158 $code.=<<___ if ($i==0);
159 mov `4*$i`($inp),$xi[0]
162 $code.=<<___ if ($i<15);
163 mov `4*$j`($inp),$xi[1]
165 mov $xi[0],`4*$i`(%rsp)
171 lea 0x5a827999($xi[0],$e),$e
177 $code.=<<___ if ($i>=15);
178 xor `4*($j%16)`(%rsp),$xi[1]
180 mov $xi[0],`4*($i%16)`(%rsp)
182 xor `4*(($j+2)%16)`(%rsp),$xi[1]
185 xor `4*(($j+8)%16)`(%rsp),$xi[1]
187 lea 0x5a827999($xi[0],$e),$e
194 push(@xi,shift(@xi));
198 my ($i,$a,$b,$c,$d,$e)=@_;
200 my $K=($i<40)?0x6ed9eba1:0xca62c1d6;
201 $code.=<<___ if ($i<79);
202 xor `4*($j%16)`(%rsp),$xi[1]
204 `"mov $xi[0],".4*($i%16)."(%rsp)" if ($i<72)`
206 xor `4*(($j+2)%16)`(%rsp),$xi[1]
209 xor `4*(($j+8)%16)`(%rsp),$xi[1]
217 $code.=<<___ if ($i==79);
228 push(@xi,shift(@xi));
232 my ($i,$a,$b,$c,$d,$e)=@_;
235 xor `4*($j%16)`(%rsp),$xi[1]
237 mov $xi[0],`4*($i%16)`(%rsp)
239 xor `4*(($j+2)%16)`(%rsp),$xi[1]
242 xor `4*(($j+8)%16)`(%rsp),$xi[1]
243 lea 0x8f1bbcdc($xi[0],$e),$e
253 push(@xi,shift(@xi));
258 .extern OPENSSL_ia32cap_P
260 .globl sha1_block_data_order
261 .type sha1_block_data_order,\@function,3
263 sha1_block_data_order:
265 mov OPENSSL_ia32cap_P+0(%rip),%r9d
266 mov OPENSSL_ia32cap_P+4(%rip),%r8d
267 mov OPENSSL_ia32cap_P+8(%rip),%r10d
268 test \$`1<<9`,%r8d # check SSSE3 bit
271 $code.=<<___ if ($shaext);
272 test \$`1<<29`,%r10d # check SHA bit
275 $code.=<<___ if ($avx>1);
276 and \$`1<<3|1<<5|1<<8`,%r10d # check AVX2+BMI1+BMI2
277 cmp \$`1<<3|1<<5|1<<8`,%r10d
280 $code.=<<___ if ($avx);
281 and \$`1<<28`,%r8d # mask AVX bit
282 and \$`1<<30`,%r9d # mask "Intel CPU" bit
284 cmp \$`1<<28|1<<30`,%r8d
293 .cfi_def_cfa_register %rax
304 mov %rdi,$ctx # reassigned argument
306 mov %rsi,$inp # reassigned argument
308 mov %rdx,$num # reassigned argument
309 mov %rax,`16*4`(%rsp)
310 .cfi_cfa_expression %rsp+64,deref,+8
323 for($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); }
324 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
325 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
326 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
340 lea `16*4`($inp),$inp
343 mov `16*4`(%rsp),%rsi
356 .cfi_def_cfa_register %rsp
360 .size sha1_block_data_order,.-sha1_block_data_order
363 ######################################################################
364 # Intel SHA Extensions implementation of SHA1 update function.
366 my ($ctx,$inp,$num)=("%rdi","%rsi","%rdx");
367 my ($ABCD,$E,$E_,$BSWAP,$ABCD_SAVE,$E_SAVE)=map("%xmm$_",(0..3,8,9));
368 my @MSG=map("%xmm$_",(4..7));
371 .type sha1_block_data_order_shaext,\@function,3
373 sha1_block_data_order_shaext:
377 $code.=<<___ if ($win64);
378 lea `-8-4*16`(%rsp),%rsp
379 movaps %xmm6,-8-4*16(%rax)
380 movaps %xmm7,-8-3*16(%rax)
381 movaps %xmm8,-8-2*16(%rax)
382 movaps %xmm9,-8-1*16(%rax)
388 movdqa K_XX_XX+0xa0(%rip),$BSWAP # byte-n-word swap
390 movdqu ($inp),@MSG[0]
391 pshufd \$0b00011011,$ABCD,$ABCD # flip word order
392 movdqu 0x10($inp),@MSG[1]
393 pshufd \$0b00011011,$E,$E # flip word order
394 movdqu 0x20($inp),@MSG[2]
395 pshufb $BSWAP,@MSG[0]
396 movdqu 0x30($inp),@MSG[3]
397 pshufb $BSWAP,@MSG[1]
398 pshufb $BSWAP,@MSG[2]
399 movdqa $E,$E_SAVE # offload $E
400 pshufb $BSWAP,@MSG[3]
406 lea 0x40($inp),%r8 # next input block
409 movdqa $ABCD,$ABCD_SAVE # offload $ABCD
411 for($i=0;$i<20-4;$i+=2) {
413 sha1msg1 @MSG[1],@MSG[0]
415 sha1rnds4 \$`int($i/5)`,$E,$ABCD # 0-3...
416 sha1nexte @MSG[1],$E_
418 sha1msg1 @MSG[2],@MSG[1]
419 sha1msg2 @MSG[3],@MSG[0]
422 sha1rnds4 \$`int(($i+1)/5)`,$E_,$ABCD
425 sha1msg2 @MSG[0],@MSG[1]
427 push(@MSG,shift(@MSG)); push(@MSG,shift(@MSG));
430 movdqu ($inp),@MSG[0]
432 sha1rnds4 \$3,$E,$ABCD # 64-67
433 sha1nexte @MSG[1],$E_
434 movdqu 0x10($inp),@MSG[1]
435 pshufb $BSWAP,@MSG[0]
438 sha1rnds4 \$3,$E_,$ABCD # 68-71
440 movdqu 0x20($inp),@MSG[2]
441 pshufb $BSWAP,@MSG[1]
444 sha1rnds4 \$3,$E,$ABCD # 72-75
445 sha1nexte @MSG[3],$E_
446 movdqu 0x30($inp),@MSG[3]
447 pshufb $BSWAP,@MSG[2]
450 sha1rnds4 \$3,$E_,$ABCD # 76-79
452 pshufb $BSWAP,@MSG[3]
454 paddd $ABCD_SAVE,$ABCD
455 movdqa $E,$E_SAVE # offload $E
459 pshufd \$0b00011011,$ABCD,$ABCD
460 pshufd \$0b00011011,$E,$E
464 $code.=<<___ if ($win64);
465 movaps -8-4*16(%rax),%xmm6
466 movaps -8-3*16(%rax),%xmm7
467 movaps -8-2*16(%rax),%xmm8
468 movaps -8-1*16(%rax),%xmm9
475 .size sha1_block_data_order_shaext,.-sha1_block_data_order_shaext
480 my @X=map("%xmm$_",(4..7,0..3));
481 my @Tx=map("%xmm$_",(8..10));
483 my @V=($A,$B,$C,$D,$E)=("%eax","%ebx","%ecx","%edx","%ebp"); # size optimization
484 my @T=("%esi","%edi");
490 my $_rol=sub { &rol(@_) };
491 my $_ror=sub { &ror(@_) };
497 jmp .Lalign32_$sn # see "Decoded ICache" in manual
505 .type sha1_block_data_order_ssse3,\@function,3
507 sha1_block_data_order_ssse3:
510 mov %rsp,$fp # frame pointer
511 .cfi_def_cfa_register $fp
518 push %r13 # redundant, done to share Win64 SE handler
522 lea `-64-($win64?6*16:0)`(%rsp),%rsp
524 $code.=<<___ if ($win64);
525 movaps %xmm6,-40-6*16($fp)
526 movaps %xmm7,-40-5*16($fp)
527 movaps %xmm8,-40-4*16($fp)
528 movaps %xmm9,-40-3*16($fp)
529 movaps %xmm10,-40-2*16($fp)
530 movaps %xmm11,-40-1*16($fp)
535 mov %rdi,$ctx # reassigned argument
536 mov %rsi,$inp # reassigned argument
537 mov %rdx,$num # reassigned argument
541 lea K_XX_XX+64(%rip),$K_XX_XX
543 mov 0($ctx),$A # load context
547 mov $B,@T[0] # magic seed
553 movdqa 64($K_XX_XX),@X[2] # pbswap mask
554 movdqa -64($K_XX_XX),@Tx[1] # K_00_19
555 movdqu 0($inp),@X[-4&7] # load input to %xmm[0-3]
556 movdqu 16($inp),@X[-3&7]
557 movdqu 32($inp),@X[-2&7]
558 movdqu 48($inp),@X[-1&7]
559 pshufb @X[2],@X[-4&7] # byte swap
560 pshufb @X[2],@X[-3&7]
561 pshufb @X[2],@X[-2&7]
563 paddd @Tx[1],@X[-4&7] # add K_00_19
564 pshufb @X[2],@X[-1&7]
565 paddd @Tx[1],@X[-3&7]
566 paddd @Tx[1],@X[-2&7]
567 movdqa @X[-4&7],0(%rsp) # X[]+K xfer to IALU
568 psubd @Tx[1],@X[-4&7] # restore X[]
569 movdqa @X[-3&7],16(%rsp)
570 psubd @Tx[1],@X[-3&7]
571 movdqa @X[-2&7],32(%rsp)
572 psubd @Tx[1],@X[-2&7]
576 sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm
577 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
579 $arg = "\$$arg" if ($arg*1 eq $arg);
580 $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
583 sub Xupdate_ssse3_16_31() # recall that $Xi starts with 4
586 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
589 eval(shift(@insns)); # ror
590 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
592 &movdqa (@Tx[0],@X[-1&7]);
593 &paddd (@Tx[1],@X[-1&7]);
597 &punpcklqdq(@X[0],@X[-3&7]); # compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8);
599 eval(shift(@insns)); # rol
601 &psrldq (@Tx[0],4); # "X[-3]", 3 dwords
605 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
607 eval(shift(@insns)); # ror
608 &pxor (@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
613 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
615 eval(shift(@insns)); # rol
616 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
620 &movdqa (@Tx[2],@X[0]);
623 eval(shift(@insns)); # ror
624 &movdqa (@Tx[0],@X[0]);
627 &pslldq (@Tx[2],12); # "X[0]"<<96, extract one dword
628 &paddd (@X[0],@X[0]);
634 eval(shift(@insns)); # rol
636 &movdqa (@Tx[1],@Tx[2]);
642 eval(shift(@insns)); # ror
643 &por (@X[0],@Tx[0]); # "X[0]"<<<=1
649 &pxor (@X[0],@Tx[2]);
651 &movdqa (@Tx[2],eval(2*16*(($Xi)/5)-64)."($K_XX_XX)"); # K_XX_XX
652 eval(shift(@insns)); # rol
656 &pxor (@X[0],@Tx[1]); # "X[0]"^=("X[0]">>96)<<<2
657 &pshufd (@Tx[1],@X[-1&7],0xee) if ($Xi==7); # was &movdqa (@Tx[0],@X[-1&7]) in Xupdate_ssse3_32_79
659 foreach (@insns) { eval; } # remaining instructions [if any]
661 $Xi++; push(@X,shift(@X)); # "rotate" X[]
662 push(@Tx,shift(@Tx));
665 sub Xupdate_ssse3_32_79()
668 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
671 eval(shift(@insns)) if ($Xi==8);
672 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
673 eval(shift(@insns)) if ($Xi==8);
674 eval(shift(@insns)); # body_20_39
676 eval(shift(@insns)) if (@insns[1] =~ /_ror/);
677 eval(shift(@insns)) if (@insns[0] =~ /_ror/);
678 &punpcklqdq(@Tx[0],@X[-1&7]); # compose "X[-6]", was &palignr(@Tx[0],@X[-2&7],8);
680 eval(shift(@insns)); # rol
682 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
686 &movdqa (@Tx[2],@Tx[1]);# "perpetuate" K_XX_XX...
687 } else { # ... or load next one
688 &movdqa (@Tx[2],eval(2*16*($Xi/5)-64)."($K_XX_XX)");
690 eval(shift(@insns)); # ror
691 &paddd (@Tx[1],@X[-1&7]);
694 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-6]"
695 eval(shift(@insns)); # body_20_39
698 eval(shift(@insns)); # rol
699 eval(shift(@insns)) if (@insns[0] =~ /_ror/);
701 &movdqa (@Tx[0],@X[0]);
704 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
705 eval(shift(@insns)); # ror
707 eval(shift(@insns)); # body_20_39
713 eval(shift(@insns)) if (@insns[0] =~ /_rol/);# rol
716 eval(shift(@insns)); # ror
718 &por (@X[0],@Tx[0]); # "X[0]"<<<=2
720 eval(shift(@insns)); # body_20_39
721 eval(shift(@insns)) if (@insns[1] =~ /_rol/);
722 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
723 &pshufd(@Tx[1],@X[-1&7],0xee) if ($Xi<19); # was &movdqa (@Tx[1],@X[0])
725 eval(shift(@insns)); # rol
728 eval(shift(@insns)); # rol
731 foreach (@insns) { eval; } # remaining instructions
733 $Xi++; push(@X,shift(@X)); # "rotate" X[]
734 push(@Tx,shift(@Tx));
737 sub Xuplast_ssse3_80()
740 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
747 &paddd (@Tx[1],@X[-1&7]);
751 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU
753 foreach (@insns) { eval; } # remaining instructions
756 &je (".Ldone_ssse3");
758 unshift(@Tx,pop(@Tx));
760 &movdqa (@X[2],"64($K_XX_XX)"); # pbswap mask
761 &movdqa (@Tx[1],"-64($K_XX_XX)"); # K_00_19
762 &movdqu (@X[-4&7],"0($inp)"); # load input
763 &movdqu (@X[-3&7],"16($inp)");
764 &movdqu (@X[-2&7],"32($inp)");
765 &movdqu (@X[-1&7],"48($inp)");
766 &pshufb (@X[-4&7],@X[2]); # byte swap
775 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
781 &pshufb (@X[($Xi-3)&7],@X[2]);
786 &paddd (@X[($Xi-4)&7],@Tx[1]);
791 &movdqa (eval(16*$Xi)."(%rsp)",@X[($Xi-4)&7]); # X[]+K xfer to IALU
796 &psubd (@X[($Xi-4)&7],@Tx[1]);
798 foreach (@insns) { eval; }
805 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
808 foreach (@insns) { eval; }
811 sub body_00_19 () { # ((c^d)&b)^d
812 # on start @T[0]=(c^d)&b
813 return &body_20_39() if ($rx==19); $rx++;
815 '($a,$b,$c,$d,$e)=@V;'.
816 '&$_ror ($b,$j?7:2)', # $b>>>2
818 '&mov (@T[1],$a)', # $b for next round
820 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
821 '&xor ($b,$c)', # $c^$d for next round
825 '&and (@T[1],$b)', # ($b&($c^$d)) for next round
827 '&xor ($b,$c)', # restore $b
828 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
832 sub body_20_39 () { # b^d^c
834 return &body_40_59() if ($rx==39); $rx++;
836 '($a,$b,$c,$d,$e)=@V;'.
837 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
838 '&xor (@T[0],$d) if($j==19);'.
839 '&xor (@T[0],$c) if($j> 19)', # ($b^$d^$c)
840 '&mov (@T[1],$a)', # $b for next round
844 '&xor (@T[1],$c) if ($j< 79)', # $b^$d for next round
846 '&$_ror ($b,7)', # $b>>>2
847 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
851 sub body_40_59 () { # ((b^c)&(c^d))^c
852 # on entry @T[0]=(b^c), (c^=d)
855 '($a,$b,$c,$d,$e)=@V;'.
856 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
857 '&and (@T[0],$c) if ($j>=40)', # (b^c)&(c^d)
858 '&xor ($c,$d) if ($j>=40)', # restore $c
860 '&$_ror ($b,7)', # $b>>>2
861 '&mov (@T[1],$a)', # $b for next round
866 '&xor (@T[1],$c) if ($j==59);'.
867 '&xor (@T[1],$b) if ($j< 59)', # b^c for next round
869 '&xor ($b,$c) if ($j< 59)', # c^d for next round
870 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
877 &Xupdate_ssse3_16_31(\&body_00_19);
878 &Xupdate_ssse3_16_31(\&body_00_19);
879 &Xupdate_ssse3_16_31(\&body_00_19);
880 &Xupdate_ssse3_16_31(\&body_00_19);
881 &Xupdate_ssse3_32_79(\&body_00_19);
882 &Xupdate_ssse3_32_79(\&body_20_39);
883 &Xupdate_ssse3_32_79(\&body_20_39);
884 &Xupdate_ssse3_32_79(\&body_20_39);
885 &Xupdate_ssse3_32_79(\&body_20_39);
886 &Xupdate_ssse3_32_79(\&body_20_39);
887 &Xupdate_ssse3_32_79(\&body_40_59);
888 &Xupdate_ssse3_32_79(\&body_40_59);
889 &Xupdate_ssse3_32_79(\&body_40_59);
890 &Xupdate_ssse3_32_79(\&body_40_59);
891 &Xupdate_ssse3_32_79(\&body_40_59);
892 &Xupdate_ssse3_32_79(\&body_20_39);
893 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done"
895 $saved_j=$j; @saved_V=@V;
897 &Xloop_ssse3(\&body_20_39);
898 &Xloop_ssse3(\&body_20_39);
899 &Xloop_ssse3(\&body_20_39);
902 add 0($ctx),$A # update context
909 mov @T[0],$B # magic seed
921 $j=$saved_j; @V=@saved_V;
923 &Xtail_ssse3(\&body_20_39);
924 &Xtail_ssse3(\&body_20_39);
925 &Xtail_ssse3(\&body_20_39);
928 add 0($ctx),$A # update context
939 $code.=<<___ if ($win64);
940 movaps -40-6*16($fp),%xmm6
941 movaps -40-5*16($fp),%xmm7
942 movaps -40-4*16($fp),%xmm8
943 movaps -40-3*16($fp),%xmm9
944 movaps -40-2*16($fp),%xmm10
945 movaps -40-1*16($fp),%xmm11
959 .cfi_def_cfa_register %rsp
963 .size sha1_block_data_order_ssse3,.-sha1_block_data_order_ssse3
967 $Xi=4; # reset variables
968 @X=map("%xmm$_",(4..7,0..3));
969 @Tx=map("%xmm$_",(8..10));
973 my $done_avx_label=".Ldone_avx";
975 my $_rol=sub { &shld(@_[0],@_) };
976 my $_ror=sub { &shrd(@_[0],@_) };
979 .type sha1_block_data_order_avx,\@function,3
981 sha1_block_data_order_avx:
985 .cfi_def_cfa_register $fp
992 push %r13 # redundant, done to share Win64 SE handler
996 lea `-64-($win64?6*16:0)`(%rsp),%rsp
999 $code.=<<___ if ($win64);
1000 vmovaps %xmm6,-40-6*16($fp)
1001 vmovaps %xmm7,-40-5*16($fp)
1002 vmovaps %xmm8,-40-4*16($fp)
1003 vmovaps %xmm9,-40-3*16($fp)
1004 vmovaps %xmm10,-40-2*16($fp)
1005 vmovaps %xmm11,-40-1*16($fp)
1010 mov %rdi,$ctx # reassigned argument
1011 mov %rsi,$inp # reassigned argument
1012 mov %rdx,$num # reassigned argument
1016 lea K_XX_XX+64(%rip),$K_XX_XX
1018 mov 0($ctx),$A # load context
1022 mov $B,@T[0] # magic seed
1028 vmovdqa 64($K_XX_XX),@X[2] # pbswap mask
1029 vmovdqa -64($K_XX_XX),$Kx # K_00_19
1030 vmovdqu 0($inp),@X[-4&7] # load input to %xmm[0-3]
1031 vmovdqu 16($inp),@X[-3&7]
1032 vmovdqu 32($inp),@X[-2&7]
1033 vmovdqu 48($inp),@X[-1&7]
1034 vpshufb @X[2],@X[-4&7],@X[-4&7] # byte swap
1036 vpshufb @X[2],@X[-3&7],@X[-3&7]
1037 vpshufb @X[2],@X[-2&7],@X[-2&7]
1038 vpshufb @X[2],@X[-1&7],@X[-1&7]
1039 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19
1040 vpaddd $Kx,@X[-3&7],@X[1]
1041 vpaddd $Kx,@X[-2&7],@X[2]
1042 vmovdqa @X[0],0(%rsp) # X[]+K xfer to IALU
1043 vmovdqa @X[1],16(%rsp)
1044 vmovdqa @X[2],32(%rsp)
1048 sub Xupdate_avx_16_31() # recall that $Xi starts with 4
1051 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
1052 my ($a,$b,$c,$d,$e);
1054 eval(shift(@insns));
1055 eval(shift(@insns));
1056 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1057 eval(shift(@insns));
1058 eval(shift(@insns));
1060 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1061 eval(shift(@insns));
1062 eval(shift(@insns));
1063 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1064 eval(shift(@insns));
1065 eval(shift(@insns));
1066 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1067 eval(shift(@insns));
1068 eval(shift(@insns));
1070 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1071 eval(shift(@insns));
1072 eval(shift(@insns));
1073 eval(shift(@insns));
1074 eval(shift(@insns));
1076 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1077 eval(shift(@insns));
1078 eval(shift(@insns));
1079 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1080 eval(shift(@insns));
1081 eval(shift(@insns));
1083 &vpsrld (@Tx[0],@X[0],31);
1084 eval(shift(@insns));
1085 eval(shift(@insns));
1086 eval(shift(@insns));
1087 eval(shift(@insns));
1089 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1090 &vpaddd (@X[0],@X[0],@X[0]);
1091 eval(shift(@insns));
1092 eval(shift(@insns));
1093 eval(shift(@insns));
1094 eval(shift(@insns));
1096 &vpsrld (@Tx[1],@Tx[2],30);
1097 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1098 eval(shift(@insns));
1099 eval(shift(@insns));
1100 eval(shift(@insns));
1101 eval(shift(@insns));
1103 &vpslld (@Tx[2],@Tx[2],2);
1104 &vpxor (@X[0],@X[0],@Tx[1]);
1105 eval(shift(@insns));
1106 eval(shift(@insns));
1107 eval(shift(@insns));
1108 eval(shift(@insns));
1110 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1111 eval(shift(@insns));
1112 eval(shift(@insns));
1113 &vmovdqa ($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1114 eval(shift(@insns));
1115 eval(shift(@insns));
1118 foreach (@insns) { eval; } # remaining instructions [if any]
1120 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1123 sub Xupdate_avx_32_79()
1126 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
1127 my ($a,$b,$c,$d,$e);
1129 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1130 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1131 eval(shift(@insns)); # body_20_39
1132 eval(shift(@insns));
1133 eval(shift(@insns));
1134 eval(shift(@insns)); # rol
1136 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1137 eval(shift(@insns));
1138 eval(shift(@insns)) if (@insns[0] !~ /&ro[rl]/);
1139 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1140 &vmovdqa ($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0);
1141 eval(shift(@insns)); # ror
1142 eval(shift(@insns));
1144 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]"
1145 eval(shift(@insns)); # body_20_39
1146 eval(shift(@insns));
1147 eval(shift(@insns));
1148 eval(shift(@insns)); # rol
1150 &vpsrld (@Tx[0],@X[0],30);
1151 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1152 eval(shift(@insns));
1153 eval(shift(@insns));
1154 eval(shift(@insns)); # ror
1155 eval(shift(@insns));
1157 &vpslld (@X[0],@X[0],2);
1158 eval(shift(@insns)); # body_20_39
1159 eval(shift(@insns));
1160 eval(shift(@insns));
1161 eval(shift(@insns)); # rol
1162 eval(shift(@insns));
1163 eval(shift(@insns));
1164 eval(shift(@insns)); # ror
1165 eval(shift(@insns));
1167 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2
1168 eval(shift(@insns)); # body_20_39
1169 eval(shift(@insns));
1170 eval(shift(@insns));
1171 eval(shift(@insns)); # rol
1172 eval(shift(@insns));
1173 eval(shift(@insns));
1174 eval(shift(@insns)); # rol
1175 eval(shift(@insns));
1177 foreach (@insns) { eval; } # remaining instructions
1179 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1182 sub Xuplast_avx_80()
1185 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1186 my ($a,$b,$c,$d,$e);
1188 eval(shift(@insns));
1189 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1190 eval(shift(@insns));
1191 eval(shift(@insns));
1192 eval(shift(@insns));
1193 eval(shift(@insns));
1195 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU
1197 foreach (@insns) { eval; } # remaining instructions
1200 &je ($done_avx_label);
1202 &vmovdqa(@X[2],"64($K_XX_XX)"); # pbswap mask
1203 &vmovdqa($Kx,"-64($K_XX_XX)"); # K_00_19
1204 &vmovdqu(@X[-4&7],"0($inp)"); # load input
1205 &vmovdqu(@X[-3&7],"16($inp)");
1206 &vmovdqu(@X[-2&7],"32($inp)");
1207 &vmovdqu(@X[-1&7],"48($inp)");
1208 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1217 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1218 my ($a,$b,$c,$d,$e);
1220 eval(shift(@insns));
1221 eval(shift(@insns));
1222 &vpshufb(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1223 eval(shift(@insns));
1224 eval(shift(@insns));
1225 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],$Kx);
1226 eval(shift(@insns));
1227 eval(shift(@insns));
1228 eval(shift(@insns));
1229 eval(shift(@insns));
1230 &vmovdqa(eval(16*$Xi)."(%rsp)",@X[$Xi&7]); # X[]+K xfer to IALU
1231 eval(shift(@insns));
1232 eval(shift(@insns));
1234 foreach (@insns) { eval; }
1241 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1242 my ($a,$b,$c,$d,$e);
1244 foreach (@insns) { eval; }
1251 &Xupdate_avx_16_31(\&body_00_19);
1252 &Xupdate_avx_16_31(\&body_00_19);
1253 &Xupdate_avx_16_31(\&body_00_19);
1254 &Xupdate_avx_16_31(\&body_00_19);
1255 &Xupdate_avx_32_79(\&body_00_19);
1256 &Xupdate_avx_32_79(\&body_20_39);
1257 &Xupdate_avx_32_79(\&body_20_39);
1258 &Xupdate_avx_32_79(\&body_20_39);
1259 &Xupdate_avx_32_79(\&body_20_39);
1260 &Xupdate_avx_32_79(\&body_20_39);
1261 &Xupdate_avx_32_79(\&body_40_59);
1262 &Xupdate_avx_32_79(\&body_40_59);
1263 &Xupdate_avx_32_79(\&body_40_59);
1264 &Xupdate_avx_32_79(\&body_40_59);
1265 &Xupdate_avx_32_79(\&body_40_59);
1266 &Xupdate_avx_32_79(\&body_20_39);
1267 &Xuplast_avx_80(\&body_20_39); # can jump to "done"
1269 $saved_j=$j; @saved_V=@V;
1271 &Xloop_avx(\&body_20_39);
1272 &Xloop_avx(\&body_20_39);
1273 &Xloop_avx(\&body_20_39);
1276 add 0($ctx),$A # update context
1283 mov @T[0],$B # magic seed
1295 $j=$saved_j; @V=@saved_V;
1297 &Xtail_avx(\&body_20_39);
1298 &Xtail_avx(\&body_20_39);
1299 &Xtail_avx(\&body_20_39);
1304 add 0($ctx),$A # update context
1315 $code.=<<___ if ($win64);
1316 movaps -40-6*16($fp),%xmm6
1317 movaps -40-5*16($fp),%xmm7
1318 movaps -40-4*16($fp),%xmm8
1319 movaps -40-3*16($fp),%xmm9
1320 movaps -40-2*16($fp),%xmm10
1321 movaps -40-1*16($fp),%xmm11
1335 .cfi_def_cfa_register %rsp
1339 .size sha1_block_data_order_avx,.-sha1_block_data_order_avx
1344 $Xi=4; # reset variables
1345 @X=map("%ymm$_",(4..7,0..3));
1346 @Tx=map("%ymm$_",(8..10));
1350 my @ROTX=("%eax","%ebp","%ebx","%ecx","%edx","%esi");
1351 my ($a5,$t0)=("%r12d","%edi");
1353 my ($A,$F,$B,$C,$D,$E)=@ROTX;
1358 .type sha1_block_data_order_avx2,\@function,3
1360 sha1_block_data_order_avx2:
1364 .cfi_def_cfa_register $fp
1377 $code.=<<___ if ($win64);
1378 lea -6*16(%rsp),%rsp
1379 vmovaps %xmm6,-40-6*16($fp)
1380 vmovaps %xmm7,-40-5*16($fp)
1381 vmovaps %xmm8,-40-4*16($fp)
1382 vmovaps %xmm9,-40-3*16($fp)
1383 vmovaps %xmm10,-40-2*16($fp)
1384 vmovaps %xmm11,-40-1*16($fp)
1388 mov %rdi,$ctx # reassigned argument
1389 mov %rsi,$inp # reassigned argument
1390 mov %rdx,$num # reassigned argument
1397 lea K_XX_XX+64(%rip),$K_XX_XX
1399 mov 0($ctx),$A # load context
1401 cmovae $inp,$frame # next or same block
1406 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask
1408 vmovdqu ($inp),%xmm0
1409 vmovdqu 16($inp),%xmm1
1410 vmovdqu 32($inp),%xmm2
1411 vmovdqu 48($inp),%xmm3
1413 vinserti128 \$1,($frame),@X[-4&7],@X[-4&7]
1414 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7]
1415 vpshufb @X[2],@X[-4&7],@X[-4&7]
1416 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7]
1417 vpshufb @X[2],@X[-3&7],@X[-3&7]
1418 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7]
1419 vpshufb @X[2],@X[-2&7],@X[-2&7]
1420 vmovdqu -64($K_XX_XX),$Kx # K_00_19
1421 vpshufb @X[2],@X[-1&7],@X[-1&7]
1423 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19
1424 vpaddd $Kx,@X[-3&7],@X[1]
1425 vmovdqu @X[0],0(%rsp) # X[]+K xfer to IALU
1426 vpaddd $Kx,@X[-2&7],@X[2]
1427 vmovdqu @X[1],32(%rsp)
1428 vpaddd $Kx,@X[-1&7],@X[3]
1429 vmovdqu @X[2],64(%rsp)
1430 vmovdqu @X[3],96(%rsp)
1432 for (;$Xi<8;$Xi++) { # Xupdate_avx2_16_31
1435 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1436 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1437 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1438 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1439 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1440 &vpsrld (@Tx[0],@X[0],31);
1441 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1442 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1443 &vpaddd (@X[0],@X[0],@X[0]);
1444 &vpsrld (@Tx[1],@Tx[2],30);
1445 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1446 &vpslld (@Tx[2],@Tx[2],2);
1447 &vpxor (@X[0],@X[0],@Tx[1]);
1448 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1449 &vpaddd (@Tx[1],@X[0],$Kx);
1450 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1452 push(@X,shift(@X)); # "rotate" X[]
1455 lea 128(%rsp),$frame
1464 sub bodyx_00_19 () { # 8 instructions, 3 cycles critical path
1465 # at start $f=(b&c)^(~b&d), $b>>>=2
1466 return &bodyx_20_39() if ($rx==19); $rx++;
1468 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1470 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1471 '&lea ($frame,"256($frame)") if ($j%32==31);',
1472 '&andn ($t0,$a,$c)', # ~b&d for next round
1474 '&add ($e,$f)', # e+=(b&c)^(~b&d)
1475 '&rorx ($a5,$a,27)', # a<<<5
1476 '&rorx ($f,$a,2)', # b>>>2 for next round
1477 '&and ($a,$b)', # b&c for next round
1479 '&add ($e,$a5)', # e+=a<<<5
1480 '&xor ($a,$t0);'. # f=(b&c)^(~b&d) for next round
1482 'unshift(@ROTX,pop(@ROTX)); $j++;'
1486 sub bodyx_20_39 () { # 7 instructions, 2 cycles critical path
1487 # on entry $f=b^c^d, $b>>>=2
1488 return &bodyx_40_59() if ($rx==39); $rx++;
1490 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1492 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1493 '&lea ($frame,"256($frame)") if ($j%32==31);',
1495 '&lea ($e,"($e,$f)")', # e+=b^c^d
1496 '&rorx ($a5,$a,27)', # a<<<5
1497 '&rorx ($f,$a,2) if ($j<79)', # b>>>2 in next round
1498 '&xor ($a,$b) if ($j<79)', # b^c for next round
1500 '&add ($e,$a5)', # e+=a<<<5
1501 '&xor ($a,$c) if ($j<79);'. # f=b^c^d for next round
1503 'unshift(@ROTX,pop(@ROTX)); $j++;'
1507 sub bodyx_40_59 () { # 10 instructions, 3 cycles critical path
1508 # on entry $f=((b^c)&(c^d)), $b>>>=2
1511 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1513 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1514 '&lea ($frame,"256($frame)") if ($j%32==31);',
1515 '&xor ($f,$c) if ($j>39)', # (b^c)&(c^d)^c
1516 '&mov ($t0,$b) if ($j<59)', # count on zero latency
1517 '&xor ($t0,$c) if ($j<59)', # c^d for next round
1519 '&lea ($e,"($e,$f)")', # e+=(b^c)&(c^d)^c
1520 '&rorx ($a5,$a,27)', # a<<<5
1521 '&rorx ($f,$a,2)', # b>>>2 in next round
1522 '&xor ($a,$b)', # b^c for next round
1524 '&add ($e,$a5)', # e+=a<<<5
1525 '&and ($a,$t0) if ($j< 59);'. # f=(b^c)&(c^d) for next round
1526 '&xor ($a,$c) if ($j==59);'. # f=b^c^d for next round
1528 'unshift(@ROTX,pop(@ROTX)); $j++;'
1532 sub Xupdate_avx2_16_31() # recall that $Xi starts with 4
1535 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 instructions
1536 my ($a,$b,$c,$d,$e);
1538 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1539 eval(shift(@insns));
1540 eval(shift(@insns));
1541 eval(shift(@insns));
1542 eval(shift(@insns));
1544 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1545 eval(shift(@insns));
1546 eval(shift(@insns));
1547 eval(shift(@insns));
1549 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1550 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1551 eval(shift(@insns));
1552 eval(shift(@insns));
1554 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1555 eval(shift(@insns));
1556 eval(shift(@insns));
1557 eval(shift(@insns));
1558 eval(shift(@insns));
1560 &vpsrld (@Tx[0],@X[0],31);
1561 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1562 eval(shift(@insns));
1563 eval(shift(@insns));
1564 eval(shift(@insns));
1566 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1567 &vpaddd (@X[0],@X[0],@X[0]);
1568 eval(shift(@insns));
1569 eval(shift(@insns));
1571 &vpsrld (@Tx[1],@Tx[2],30);
1572 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1573 eval(shift(@insns));
1574 eval(shift(@insns));
1576 &vpslld (@Tx[2],@Tx[2],2);
1577 &vpxor (@X[0],@X[0],@Tx[1]);
1578 eval(shift(@insns));
1579 eval(shift(@insns));
1581 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1582 eval(shift(@insns));
1583 eval(shift(@insns));
1584 eval(shift(@insns));
1586 &vpaddd (@Tx[1],@X[0],$Kx);
1587 eval(shift(@insns));
1588 eval(shift(@insns));
1589 eval(shift(@insns));
1590 &vmovdqu(eval(32*($Xi))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1592 foreach (@insns) { eval; } # remaining instructions [if any]
1595 push(@X,shift(@X)); # "rotate" X[]
1598 sub Xupdate_avx2_32_79()
1601 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 to 50 instructions
1602 my ($a,$b,$c,$d,$e);
1604 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1605 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1606 eval(shift(@insns));
1607 eval(shift(@insns));
1609 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1610 &vmovdqu($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0);
1611 eval(shift(@insns));
1612 eval(shift(@insns));
1613 eval(shift(@insns));
1615 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]"
1616 eval(shift(@insns));
1617 eval(shift(@insns));
1618 eval(shift(@insns));
1620 &vpsrld (@Tx[0],@X[0],30);
1621 &vpslld (@X[0],@X[0],2);
1622 eval(shift(@insns));
1623 eval(shift(@insns));
1624 eval(shift(@insns));
1626 #&vpslld (@X[0],@X[0],2);
1627 eval(shift(@insns));
1628 eval(shift(@insns));
1629 eval(shift(@insns));
1631 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2
1632 eval(shift(@insns));
1633 eval(shift(@insns));
1634 eval(shift(@insns));
1635 eval(shift(@insns));
1637 &vpaddd (@Tx[1],@X[0],$Kx);
1638 eval(shift(@insns));
1639 eval(shift(@insns));
1640 eval(shift(@insns));
1641 eval(shift(@insns));
1643 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1645 foreach (@insns) { eval; } # remaining instructions
1648 push(@X,shift(@X)); # "rotate" X[]
1654 my @insns = (&$body,&$body,&$body,&$body,&$body); # 32 instructions
1655 my ($a,$b,$c,$d,$e);
1657 foreach (@insns) { eval; }
1661 &Xupdate_avx2_32_79(\&bodyx_00_19);
1662 &Xupdate_avx2_32_79(\&bodyx_00_19);
1663 &Xupdate_avx2_32_79(\&bodyx_00_19);
1664 &Xupdate_avx2_32_79(\&bodyx_00_19);
1666 &Xupdate_avx2_32_79(\&bodyx_20_39);
1667 &Xupdate_avx2_32_79(\&bodyx_20_39);
1668 &Xupdate_avx2_32_79(\&bodyx_20_39);
1669 &Xupdate_avx2_32_79(\&bodyx_20_39);
1672 &Xupdate_avx2_32_79(\&bodyx_40_59);
1673 &Xupdate_avx2_32_79(\&bodyx_40_59);
1674 &Xupdate_avx2_32_79(\&bodyx_40_59);
1675 &Xupdate_avx2_32_79(\&bodyx_40_59);
1677 &Xloop_avx2(\&bodyx_20_39);
1678 &Xloop_avx2(\&bodyx_20_39);
1679 &Xloop_avx2(\&bodyx_20_39);
1680 &Xloop_avx2(\&bodyx_20_39);
1683 lea 128($inp),$frame
1684 lea 128($inp),%rdi # borrow $t0
1686 cmovae $inp,$frame # next or previous block
1688 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c
1689 add 0($ctx),@ROTX[0] # update context
1690 add 4($ctx),@ROTX[1]
1691 add 8($ctx),@ROTX[3]
1692 mov @ROTX[0],0($ctx)
1693 add 12($ctx),@ROTX[4]
1694 mov @ROTX[1],4($ctx)
1695 mov @ROTX[0],$A # A=d
1696 add 16($ctx),@ROTX[5]
1698 mov @ROTX[3],8($ctx)
1699 mov @ROTX[4],$D # D=b
1700 #xchg @ROTX[5],$F # F=c, C=f
1701 mov @ROTX[4],12($ctx)
1702 mov @ROTX[1],$F # F=e
1703 mov @ROTX[5],16($ctx)
1705 mov @ROTX[5],$E # E=c
1707 #xchg $F,$E # E=c, F=e
1713 $Xi=4; # reset variables
1714 @X=map("%ymm$_",(4..7,0..3));
1717 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask
1718 cmp $num,%rdi # borrowed $t0
1721 vmovdqu -64(%rdi),%xmm0 # low part of @X[-4&7]
1722 vmovdqu -48(%rdi),%xmm1
1723 vmovdqu -32(%rdi),%xmm2
1724 vmovdqu -16(%rdi),%xmm3
1725 vinserti128 \$1,0($frame),@X[-4&7],@X[-4&7]
1726 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7]
1727 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7]
1728 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7]
1733 lea 128+16(%rsp),$frame
1740 $rx=$j=0; @ROTX=($A,$F,$B,$C,$D,$E);
1742 &Xloop_avx2 (\&bodyx_00_19);
1743 &Xloop_avx2 (\&bodyx_00_19);
1744 &Xloop_avx2 (\&bodyx_00_19);
1745 &Xloop_avx2 (\&bodyx_00_19);
1747 &Xloop_avx2 (\&bodyx_20_39);
1748 &vmovdqu ($Kx,"-64($K_XX_XX)"); # K_00_19
1749 &vpshufb (@X[-4&7],@X[-4&7],@X[2]); # byte swap
1750 &Xloop_avx2 (\&bodyx_20_39);
1751 &vpshufb (@X[-3&7],@X[-3&7],@X[2]);
1752 &vpaddd (@Tx[0],@X[-4&7],$Kx); # add K_00_19
1753 &Xloop_avx2 (\&bodyx_20_39);
1754 &vmovdqu ("0(%rsp)",@Tx[0]);
1755 &vpshufb (@X[-2&7],@X[-2&7],@X[2]);
1756 &vpaddd (@Tx[1],@X[-3&7],$Kx);
1757 &Xloop_avx2 (\&bodyx_20_39);
1758 &vmovdqu ("32(%rsp)",@Tx[1]);
1759 &vpshufb (@X[-1&7],@X[-1&7],@X[2]);
1760 &vpaddd (@X[2],@X[-2&7],$Kx);
1762 &Xloop_avx2 (\&bodyx_40_59);
1764 &vmovdqu ("64(%rsp)",@X[2]);
1765 &vpaddd (@X[3],@X[-1&7],$Kx);
1766 &Xloop_avx2 (\&bodyx_40_59);
1767 &vmovdqu ("96(%rsp)",@X[3]);
1768 &Xloop_avx2 (\&bodyx_40_59);
1769 &Xupdate_avx2_16_31(\&bodyx_40_59);
1771 &Xupdate_avx2_16_31(\&bodyx_20_39);
1772 &Xupdate_avx2_16_31(\&bodyx_20_39);
1773 &Xupdate_avx2_16_31(\&bodyx_20_39);
1774 &Xloop_avx2 (\&bodyx_20_39);
1777 lea 128(%rsp),$frame
1779 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c
1780 add 0($ctx),@ROTX[0] # update context
1781 add 4($ctx),@ROTX[1]
1782 add 8($ctx),@ROTX[3]
1783 mov @ROTX[0],0($ctx)
1784 add 12($ctx),@ROTX[4]
1785 mov @ROTX[1],4($ctx)
1786 mov @ROTX[0],$A # A=d
1787 add 16($ctx),@ROTX[5]
1789 mov @ROTX[3],8($ctx)
1790 mov @ROTX[4],$D # D=b
1791 #xchg @ROTX[5],$F # F=c, C=f
1792 mov @ROTX[4],12($ctx)
1793 mov @ROTX[1],$F # F=e
1794 mov @ROTX[5],16($ctx)
1796 mov @ROTX[5],$E # E=c
1798 #xchg $F,$E # E=c, F=e
1806 $code.=<<___ if ($win64);
1807 movaps -40-6*16($fp),%xmm6
1808 movaps -40-5*16($fp),%xmm7
1809 movaps -40-4*16($fp),%xmm8
1810 movaps -40-3*16($fp),%xmm9
1811 movaps -40-2*16($fp),%xmm10
1812 movaps -40-1*16($fp),%xmm11
1826 .cfi_def_cfa_register %rsp
1830 .size sha1_block_data_order_avx2,.-sha1_block_data_order_avx2
1837 .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19
1838 .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19
1839 .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39
1840 .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39
1841 .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59
1842 .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59
1843 .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79
1844 .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79
1845 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask
1846 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask
1847 .byte 0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0
1851 .asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1855 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1856 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
1864 .extern __imp_RtlVirtualUnwind
1865 .type se_handler,\@abi-omnipotent
1879 mov 120($context),%rax # pull context->Rax
1880 mov 248($context),%rbx # pull context->Rip
1882 lea .Lprologue(%rip),%r10
1883 cmp %r10,%rbx # context->Rip<.Lprologue
1884 jb .Lcommon_seh_tail
1886 mov 152($context),%rax # pull context->Rsp
1888 lea .Lepilogue(%rip),%r10
1889 cmp %r10,%rbx # context->Rip>=.Lepilogue
1890 jae .Lcommon_seh_tail
1892 mov `16*4`(%rax),%rax # pull saved stack pointer
1899 mov %rbx,144($context) # restore context->Rbx
1900 mov %rbp,160($context) # restore context->Rbp
1901 mov %r12,216($context) # restore context->R12
1902 mov %r13,224($context) # restore context->R13
1903 mov %r14,232($context) # restore context->R14
1905 jmp .Lcommon_seh_tail
1906 .size se_handler,.-se_handler
1909 $code.=<<___ if ($shaext);
1910 .type shaext_handler,\@abi-omnipotent
1924 mov 120($context),%rax # pull context->Rax
1925 mov 248($context),%rbx # pull context->Rip
1927 lea .Lprologue_shaext(%rip),%r10
1928 cmp %r10,%rbx # context->Rip<.Lprologue
1929 jb .Lcommon_seh_tail
1931 lea .Lepilogue_shaext(%rip),%r10
1932 cmp %r10,%rbx # context->Rip>=.Lepilogue
1933 jae .Lcommon_seh_tail
1935 lea -8-4*16(%rax),%rsi
1936 lea 512($context),%rdi # &context.Xmm6
1938 .long 0xa548f3fc # cld; rep movsq
1940 jmp .Lcommon_seh_tail
1941 .size shaext_handler,.-shaext_handler
1945 .type ssse3_handler,\@abi-omnipotent
1959 mov 120($context),%rax # pull context->Rax
1960 mov 248($context),%rbx # pull context->Rip
1962 mov 8($disp),%rsi # disp->ImageBase
1963 mov 56($disp),%r11 # disp->HandlerData
1965 mov 0(%r11),%r10d # HandlerData[0]
1966 lea (%rsi,%r10),%r10 # prologue label
1967 cmp %r10,%rbx # context->Rip<prologue label
1968 jb .Lcommon_seh_tail
1970 mov 208($context),%rax # pull context->R11
1972 mov 4(%r11),%r10d # HandlerData[1]
1973 lea (%rsi,%r10),%r10 # epilogue label
1974 cmp %r10,%rbx # context->Rip>=epilogue label
1975 jae .Lcommon_seh_tail
1977 lea -40-6*16(%rax),%rsi
1978 lea 512($context),%rdi # &context.Xmm6
1980 .long 0xa548f3fc # cld; rep movsq
1987 mov %rbx,144($context) # restore context->Rbx
1988 mov %rbp,160($context) # restore context->Rbp
1989 mov %r12,216($context) # restore context->R12
1990 mov %r13,224($context) # restore context->R13
1991 mov %r14,232($context) # restore context->R14
1996 mov %rax,152($context) # restore context->Rsp
1997 mov %rsi,168($context) # restore context->Rsi
1998 mov %rdi,176($context) # restore context->Rdi
2000 mov 40($disp),%rdi # disp->ContextRecord
2001 mov $context,%rsi # context
2002 mov \$154,%ecx # sizeof(CONTEXT)
2003 .long 0xa548f3fc # cld; rep movsq
2006 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
2007 mov 8(%rsi),%rdx # arg2, disp->ImageBase
2008 mov 0(%rsi),%r8 # arg3, disp->ControlPc
2009 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
2010 mov 40(%rsi),%r10 # disp->ContextRecord
2011 lea 56(%rsi),%r11 # &disp->HandlerData
2012 lea 24(%rsi),%r12 # &disp->EstablisherFrame
2013 mov %r10,32(%rsp) # arg5
2014 mov %r11,40(%rsp) # arg6
2015 mov %r12,48(%rsp) # arg7
2016 mov %rcx,56(%rsp) # arg8, (NULL)
2017 call *__imp_RtlVirtualUnwind(%rip)
2019 mov \$1,%eax # ExceptionContinueSearch
2031 .size ssse3_handler,.-ssse3_handler
2035 .rva .LSEH_begin_sha1_block_data_order
2036 .rva .LSEH_end_sha1_block_data_order
2037 .rva .LSEH_info_sha1_block_data_order
2039 $code.=<<___ if ($shaext);
2040 .rva .LSEH_begin_sha1_block_data_order_shaext
2041 .rva .LSEH_end_sha1_block_data_order_shaext
2042 .rva .LSEH_info_sha1_block_data_order_shaext
2045 .rva .LSEH_begin_sha1_block_data_order_ssse3
2046 .rva .LSEH_end_sha1_block_data_order_ssse3
2047 .rva .LSEH_info_sha1_block_data_order_ssse3
2049 $code.=<<___ if ($avx);
2050 .rva .LSEH_begin_sha1_block_data_order_avx
2051 .rva .LSEH_end_sha1_block_data_order_avx
2052 .rva .LSEH_info_sha1_block_data_order_avx
2054 $code.=<<___ if ($avx>1);
2055 .rva .LSEH_begin_sha1_block_data_order_avx2
2056 .rva .LSEH_end_sha1_block_data_order_avx2
2057 .rva .LSEH_info_sha1_block_data_order_avx2
2062 .LSEH_info_sha1_block_data_order:
2066 $code.=<<___ if ($shaext);
2067 .LSEH_info_sha1_block_data_order_shaext:
2072 .LSEH_info_sha1_block_data_order_ssse3:
2075 .rva .Lprologue_ssse3,.Lepilogue_ssse3 # HandlerData[]
2077 $code.=<<___ if ($avx);
2078 .LSEH_info_sha1_block_data_order_avx:
2081 .rva .Lprologue_avx,.Lepilogue_avx # HandlerData[]
2083 $code.=<<___ if ($avx>1);
2084 .LSEH_info_sha1_block_data_order_avx2:
2087 .rva .Lprologue_avx2,.Lepilogue_avx2 # HandlerData[]
2091 ####################################################################
2094 if (@_[0] =~ /\$([x0-9a-f]+),\s*%xmm([0-7]),\s*%xmm([0-7])/) {
2095 my @opcode=(0x0f,0x3a,0xcc);
2096 push @opcode,0xc0|($2&7)|(($3&7)<<3); # ModR/M
2098 push @opcode,$c=~/^0/?oct($c):$c;
2099 return ".byte\t".join(',',@opcode);
2101 return "sha1rnds4\t".@_[0];
2108 "sha1nexte" => 0xc8,
2110 "sha1msg2" => 0xca );
2112 if (defined($opcodelet{$instr}) && @_[0] =~ /%xmm([0-9]+),\s*%xmm([0-9]+)/) {
2113 my @opcode=(0x0f,0x38);
2115 $rex|=0x04 if ($2>=8);
2116 $rex|=0x01 if ($1>=8);
2117 unshift @opcode,0x40|$rex if ($rex);
2118 push @opcode,$opcodelet{$instr};
2119 push @opcode,0xc0|($1&7)|(($2&7)<<3); # ModR/M
2120 return ".byte\t".join(',',@opcode);
2122 return $instr."\t".@_[0];
2126 foreach (split("\n",$code)) {
2127 s/\`([^\`]*)\`/eval $1/geo;
2129 s/\b(sha1rnds4)\s+(.*)/sha1rnds4($2)/geo or
2130 s/\b(sha1[^\s]*)\s+(.*)/sha1op38($1,$2)/geo;
2134 close STDOUT or die "error closing STDOUT";