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15 # assembler, which originally made it hard to motivate the effort.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.
62 # Add support for Intel SHA Extensions.
64 ######################################################################
65 # Current performance is summarized in following table. Numbers are
66 # CPU clock cycles spent to process single byte (less is better).
71 # Core2 6.55 6.05/+8% -
72 # Westmere 6.73 5.30/+27% -
73 # Sandy Bridge 7.70 6.10/+26% 4.99/+54%
74 # Ivy Bridge 6.06 4.67/+30% 4.60/+32%
75 # Haswell 5.45 4.15/+31% 3.57/+53%
76 # Skylake 5.18 4.06/+28% 3.54/+46%
77 # Bulldozer 9.11 5.95/+53%
78 # VIA Nano 9.32 7.15/+30%
80 # Silvermont 13.1(*) 9.37/+40%
82 # (*) obviously suboptimal result, nothing was done about it,
83 # because SSSE3 code is compiled unconditionally;
87 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
89 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
91 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
92 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
93 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
94 die "can't locate x86_64-xlate.pl";
96 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
97 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
98 $avx = ($1>=2.19) + ($1>=2.22);
101 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
102 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
103 $avx = ($1>=2.09) + ($1>=2.10);
106 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
107 `ml64 2>&1` =~ /Version ([0-9]+)\./) {
108 $avx = ($1>=10) + ($1>=11);
111 if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([2-9]\.[0-9]+)/) {
112 $avx = ($2>=3.0) + ($2>3.0);
115 $shaext=1; ### set to zero if compiling for 1.0.1
116 $avx=1 if (!$shaext && $avx);
118 open OUT,"| \"$^X\" $xlate $flavour $output";
121 $ctx="%rdi"; # 1st arg
122 $inp="%rsi"; # 2nd arg
123 $num="%rdx"; # 3rd arg
125 # reassign arguments in order to produce more compact code
133 @xi=("%edx","%ebp","%r14d");
143 my ($i,$a,$b,$c,$d,$e)=@_;
145 $code.=<<___ if ($i==0);
146 mov `4*$i`($inp),$xi[0]
149 $code.=<<___ if ($i<15);
150 mov `4*$j`($inp),$xi[1]
152 mov $xi[0],`4*$i`(%rsp)
158 lea 0x5a827999($xi[0],$e),$e
164 $code.=<<___ if ($i>=15);
165 xor `4*($j%16)`(%rsp),$xi[1]
167 mov $xi[0],`4*($i%16)`(%rsp)
169 xor `4*(($j+2)%16)`(%rsp),$xi[1]
172 xor `4*(($j+8)%16)`(%rsp),$xi[1]
174 lea 0x5a827999($xi[0],$e),$e
181 push(@xi,shift(@xi));
185 my ($i,$a,$b,$c,$d,$e)=@_;
187 my $K=($i<40)?0x6ed9eba1:0xca62c1d6;
188 $code.=<<___ if ($i<79);
189 xor `4*($j%16)`(%rsp),$xi[1]
191 `"mov $xi[0],".4*($i%16)."(%rsp)" if ($i<72)`
193 xor `4*(($j+2)%16)`(%rsp),$xi[1]
196 xor `4*(($j+8)%16)`(%rsp),$xi[1]
204 $code.=<<___ if ($i==79);
215 push(@xi,shift(@xi));
219 my ($i,$a,$b,$c,$d,$e)=@_;
222 xor `4*($j%16)`(%rsp),$xi[1]
224 mov $xi[0],`4*($i%16)`(%rsp)
226 xor `4*(($j+2)%16)`(%rsp),$xi[1]
229 xor `4*(($j+8)%16)`(%rsp),$xi[1]
230 lea 0x8f1bbcdc($xi[0],$e),$e
240 push(@xi,shift(@xi));
245 .extern OPENSSL_ia32cap_P
247 .globl sha1_block_data_order
248 .type sha1_block_data_order,\@function,3
250 sha1_block_data_order:
251 mov OPENSSL_ia32cap_P+0(%rip),%r9d
252 mov OPENSSL_ia32cap_P+4(%rip),%r8d
253 mov OPENSSL_ia32cap_P+8(%rip),%r10d
254 test \$`1<<9`,%r8d # check SSSE3 bit
257 $code.=<<___ if ($shaext);
258 test \$`1<<29`,%r10d # check SHA bit
261 $code.=<<___ if ($avx>1);
262 and \$`1<<3|1<<5|1<<8`,%r10d # check AVX2+BMI1+BMI2
263 cmp \$`1<<3|1<<5|1<<8`,%r10d
266 $code.=<<___ if ($avx);
267 and \$`1<<28`,%r8d # mask AVX bit
268 and \$`1<<30`,%r9d # mask "Intel CPU" bit
270 cmp \$`1<<28|1<<30`,%r8d
284 mov %rdi,$ctx # reassigned argument
286 mov %rsi,$inp # reassigned argument
288 mov %rdx,$num # reassigned argument
289 mov %rax,`16*4`(%rsp)
302 for($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); }
303 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
304 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
305 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
319 lea `16*4`($inp),$inp
322 mov `16*4`(%rsp),%rsi
331 .size sha1_block_data_order,.-sha1_block_data_order
334 ######################################################################
335 # Intel SHA Extensions implementation of SHA1 update function.
337 my ($ctx,$inp,$num)=("%rdi","%rsi","%rdx");
338 my ($ABCD,$E,$E_,$BSWAP,$ABCD_SAVE,$E_SAVE)=map("%xmm$_",(0..3,8,9));
339 my @MSG=map("%xmm$_",(4..7));
342 .type sha1_block_data_order_shaext,\@function,3
344 sha1_block_data_order_shaext:
347 $code.=<<___ if ($win64);
348 lea `-8-4*16`(%rsp),%rsp
349 movaps %xmm6,-8-4*16(%rax)
350 movaps %xmm7,-8-3*16(%rax)
351 movaps %xmm8,-8-2*16(%rax)
352 movaps %xmm9,-8-1*16(%rax)
358 movdqa K_XX_XX+0xa0(%rip),$BSWAP # byte-n-word swap
360 movdqu ($inp),@MSG[0]
361 pshufd \$0b00011011,$ABCD,$ABCD # flip word order
362 movdqu 0x10($inp),@MSG[1]
363 pshufd \$0b00011011,$E,$E # flip word order
364 movdqu 0x20($inp),@MSG[2]
365 pshufb $BSWAP,@MSG[0]
366 movdqu 0x30($inp),@MSG[3]
367 pshufb $BSWAP,@MSG[1]
368 pshufb $BSWAP,@MSG[2]
369 movdqa $E,$E_SAVE # offload $E
370 pshufb $BSWAP,@MSG[3]
376 lea 0x40($inp),%rax # next input block
379 movdqa $ABCD,$ABCD_SAVE # offload $ABCD
381 for($i=0;$i<20-4;$i+=2) {
383 sha1msg1 @MSG[1],@MSG[0]
385 sha1rnds4 \$`int($i/5)`,$E,$ABCD # 0-3...
386 sha1nexte @MSG[1],$E_
388 sha1msg1 @MSG[2],@MSG[1]
389 sha1msg2 @MSG[3],@MSG[0]
392 sha1rnds4 \$`int(($i+1)/5)`,$E_,$ABCD
395 sha1msg2 @MSG[0],@MSG[1]
397 push(@MSG,shift(@MSG)); push(@MSG,shift(@MSG));
400 movdqu ($inp),@MSG[0]
402 sha1rnds4 \$3,$E,$ABCD # 64-67
403 sha1nexte @MSG[1],$E_
404 movdqu 0x10($inp),@MSG[1]
405 pshufb $BSWAP,@MSG[0]
408 sha1rnds4 \$3,$E_,$ABCD # 68-71
410 movdqu 0x20($inp),@MSG[2]
411 pshufb $BSWAP,@MSG[1]
414 sha1rnds4 \$3,$E,$ABCD # 72-75
415 sha1nexte @MSG[3],$E_
416 movdqu 0x30($inp),@MSG[3]
417 pshufb $BSWAP,@MSG[2]
420 sha1rnds4 \$3,$E_,$ABCD # 76-79
422 pshufb $BSWAP,@MSG[3]
424 paddd $ABCD_SAVE,$ABCD
425 movdqa $E,$E_SAVE # offload $E
429 pshufd \$0b00011011,$ABCD,$ABCD
430 pshufd \$0b00011011,$E,$E
434 $code.=<<___ if ($win64);
435 movaps -8-4*16(%rax),%xmm6
436 movaps -8-3*16(%rax),%xmm7
437 movaps -8-2*16(%rax),%xmm8
438 movaps -8-1*16(%rax),%xmm9
444 .size sha1_block_data_order_shaext,.-sha1_block_data_order_shaext
449 my @X=map("%xmm$_",(4..7,0..3));
450 my @Tx=map("%xmm$_",(8..10));
452 my @V=($A,$B,$C,$D,$E)=("%eax","%ebx","%ecx","%edx","%ebp"); # size optimization
453 my @T=("%esi","%edi");
458 my $_rol=sub { &rol(@_) };
459 my $_ror=sub { &ror(@_) };
465 jmp .Lalign32_$sn # see "Decoded ICache" in manual
473 .type sha1_block_data_order_ssse3,\@function,3
475 sha1_block_data_order_ssse3:
481 push %r13 # redundant, done to share Win64 SE handler
483 lea `-64-($win64?6*16:0)`(%rsp),%rsp
485 $code.=<<___ if ($win64);
486 movaps %xmm6,-40-6*16(%rax)
487 movaps %xmm7,-40-5*16(%rax)
488 movaps %xmm8,-40-4*16(%rax)
489 movaps %xmm9,-40-3*16(%rax)
490 movaps %xmm10,-40-2*16(%rax)
491 movaps %xmm11,-40-1*16(%rax)
495 mov %rax,%r14 # original %rsp
497 mov %rdi,$ctx # reassigned argument
498 mov %rsi,$inp # reassigned argument
499 mov %rdx,$num # reassigned argument
503 lea K_XX_XX+64(%rip),$K_XX_XX
505 mov 0($ctx),$A # load context
509 mov $B,@T[0] # magic seed
515 movdqa 64($K_XX_XX),@X[2] # pbswap mask
516 movdqa -64($K_XX_XX),@Tx[1] # K_00_19
517 movdqu 0($inp),@X[-4&7] # load input to %xmm[0-3]
518 movdqu 16($inp),@X[-3&7]
519 movdqu 32($inp),@X[-2&7]
520 movdqu 48($inp),@X[-1&7]
521 pshufb @X[2],@X[-4&7] # byte swap
522 pshufb @X[2],@X[-3&7]
523 pshufb @X[2],@X[-2&7]
525 paddd @Tx[1],@X[-4&7] # add K_00_19
526 pshufb @X[2],@X[-1&7]
527 paddd @Tx[1],@X[-3&7]
528 paddd @Tx[1],@X[-2&7]
529 movdqa @X[-4&7],0(%rsp) # X[]+K xfer to IALU
530 psubd @Tx[1],@X[-4&7] # restore X[]
531 movdqa @X[-3&7],16(%rsp)
532 psubd @Tx[1],@X[-3&7]
533 movdqa @X[-2&7],32(%rsp)
534 psubd @Tx[1],@X[-2&7]
538 sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm
539 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
541 $arg = "\$$arg" if ($arg*1 eq $arg);
542 $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
545 sub Xupdate_ssse3_16_31() # recall that $Xi starts wtih 4
548 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
551 eval(shift(@insns)); # ror
552 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
554 &movdqa (@Tx[0],@X[-1&7]);
555 &paddd (@Tx[1],@X[-1&7]);
559 &punpcklqdq(@X[0],@X[-3&7]); # compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8);
561 eval(shift(@insns)); # rol
563 &psrldq (@Tx[0],4); # "X[-3]", 3 dwords
567 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
569 eval(shift(@insns)); # ror
570 &pxor (@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
575 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
577 eval(shift(@insns)); # rol
578 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
582 &movdqa (@Tx[2],@X[0]);
585 eval(shift(@insns)); # ror
586 &movdqa (@Tx[0],@X[0]);
589 &pslldq (@Tx[2],12); # "X[0]"<<96, extract one dword
590 &paddd (@X[0],@X[0]);
596 eval(shift(@insns)); # rol
598 &movdqa (@Tx[1],@Tx[2]);
604 eval(shift(@insns)); # ror
605 &por (@X[0],@Tx[0]); # "X[0]"<<<=1
611 &pxor (@X[0],@Tx[2]);
613 &movdqa (@Tx[2],eval(2*16*(($Xi)/5)-64)."($K_XX_XX)"); # K_XX_XX
614 eval(shift(@insns)); # rol
618 &pxor (@X[0],@Tx[1]); # "X[0]"^=("X[0]">>96)<<<2
619 &pshufd (@Tx[1],@X[-1&7],0xee) if ($Xi==7); # was &movdqa (@Tx[0],@X[-1&7]) in Xupdate_ssse3_32_79
621 foreach (@insns) { eval; } # remaining instructions [if any]
623 $Xi++; push(@X,shift(@X)); # "rotate" X[]
624 push(@Tx,shift(@Tx));
627 sub Xupdate_ssse3_32_79()
630 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
633 eval(shift(@insns)) if ($Xi==8);
634 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
635 eval(shift(@insns)) if ($Xi==8);
636 eval(shift(@insns)); # body_20_39
638 eval(shift(@insns)) if (@insns[1] =~ /_ror/);
639 eval(shift(@insns)) if (@insns[0] =~ /_ror/);
640 &punpcklqdq(@Tx[0],@X[-1&7]); # compose "X[-6]", was &palignr(@Tx[0],@X[-2&7],8);
642 eval(shift(@insns)); # rol
644 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
648 &movdqa (@Tx[2],@Tx[1]);# "perpetuate" K_XX_XX...
649 } else { # ... or load next one
650 &movdqa (@Tx[2],eval(2*16*($Xi/5)-64)."($K_XX_XX)");
652 eval(shift(@insns)); # ror
653 &paddd (@Tx[1],@X[-1&7]);
656 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-6]"
657 eval(shift(@insns)); # body_20_39
660 eval(shift(@insns)); # rol
661 eval(shift(@insns)) if (@insns[0] =~ /_ror/);
663 &movdqa (@Tx[0],@X[0]);
666 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
667 eval(shift(@insns)); # ror
669 eval(shift(@insns)); # body_20_39
675 eval(shift(@insns)) if (@insns[0] =~ /_rol/);# rol
678 eval(shift(@insns)); # ror
680 &por (@X[0],@Tx[0]); # "X[0]"<<<=2
682 eval(shift(@insns)); # body_20_39
683 eval(shift(@insns)) if (@insns[1] =~ /_rol/);
684 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
685 &pshufd(@Tx[1],@X[-1&7],0xee) if ($Xi<19); # was &movdqa (@Tx[1],@X[0])
687 eval(shift(@insns)); # rol
690 eval(shift(@insns)); # rol
693 foreach (@insns) { eval; } # remaining instructions
695 $Xi++; push(@X,shift(@X)); # "rotate" X[]
696 push(@Tx,shift(@Tx));
699 sub Xuplast_ssse3_80()
702 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
709 &paddd (@Tx[1],@X[-1&7]);
713 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU
715 foreach (@insns) { eval; } # remaining instructions
718 &je (".Ldone_ssse3");
720 unshift(@Tx,pop(@Tx));
722 &movdqa (@X[2],"64($K_XX_XX)"); # pbswap mask
723 &movdqa (@Tx[1],"-64($K_XX_XX)"); # K_00_19
724 &movdqu (@X[-4&7],"0($inp)"); # load input
725 &movdqu (@X[-3&7],"16($inp)");
726 &movdqu (@X[-2&7],"32($inp)");
727 &movdqu (@X[-1&7],"48($inp)");
728 &pshufb (@X[-4&7],@X[2]); # byte swap
737 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
743 &pshufb (@X[($Xi-3)&7],@X[2]);
748 &paddd (@X[($Xi-4)&7],@Tx[1]);
753 &movdqa (eval(16*$Xi)."(%rsp)",@X[($Xi-4)&7]); # X[]+K xfer to IALU
758 &psubd (@X[($Xi-4)&7],@Tx[1]);
760 foreach (@insns) { eval; }
767 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
770 foreach (@insns) { eval; }
773 sub body_00_19 () { # ((c^d)&b)^d
774 # on start @T[0]=(c^d)&b
775 return &body_20_39() if ($rx==19); $rx++;
777 '($a,$b,$c,$d,$e)=@V;'.
778 '&$_ror ($b,$j?7:2)', # $b>>>2
780 '&mov (@T[1],$a)', # $b for next round
782 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
783 '&xor ($b,$c)', # $c^$d for next round
787 '&and (@T[1],$b)', # ($b&($c^$d)) for next round
789 '&xor ($b,$c)', # restore $b
790 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
794 sub body_20_39 () { # b^d^c
796 return &body_40_59() if ($rx==39); $rx++;
798 '($a,$b,$c,$d,$e)=@V;'.
799 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
800 '&xor (@T[0],$d) if($j==19);'.
801 '&xor (@T[0],$c) if($j> 19)', # ($b^$d^$c)
802 '&mov (@T[1],$a)', # $b for next round
806 '&xor (@T[1],$c) if ($j< 79)', # $b^$d for next round
808 '&$_ror ($b,7)', # $b>>>2
809 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
813 sub body_40_59 () { # ((b^c)&(c^d))^c
814 # on entry @T[0]=(b^c), (c^=d)
817 '($a,$b,$c,$d,$e)=@V;'.
818 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
819 '&and (@T[0],$c) if ($j>=40)', # (b^c)&(c^d)
820 '&xor ($c,$d) if ($j>=40)', # restore $c
822 '&$_ror ($b,7)', # $b>>>2
823 '&mov (@T[1],$a)', # $b for next round
828 '&xor (@T[1],$c) if ($j==59);'.
829 '&xor (@T[1],$b) if ($j< 59)', # b^c for next round
831 '&xor ($b,$c) if ($j< 59)', # c^d for next round
832 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
839 &Xupdate_ssse3_16_31(\&body_00_19);
840 &Xupdate_ssse3_16_31(\&body_00_19);
841 &Xupdate_ssse3_16_31(\&body_00_19);
842 &Xupdate_ssse3_16_31(\&body_00_19);
843 &Xupdate_ssse3_32_79(\&body_00_19);
844 &Xupdate_ssse3_32_79(\&body_20_39);
845 &Xupdate_ssse3_32_79(\&body_20_39);
846 &Xupdate_ssse3_32_79(\&body_20_39);
847 &Xupdate_ssse3_32_79(\&body_20_39);
848 &Xupdate_ssse3_32_79(\&body_20_39);
849 &Xupdate_ssse3_32_79(\&body_40_59);
850 &Xupdate_ssse3_32_79(\&body_40_59);
851 &Xupdate_ssse3_32_79(\&body_40_59);
852 &Xupdate_ssse3_32_79(\&body_40_59);
853 &Xupdate_ssse3_32_79(\&body_40_59);
854 &Xupdate_ssse3_32_79(\&body_20_39);
855 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done"
857 $saved_j=$j; @saved_V=@V;
859 &Xloop_ssse3(\&body_20_39);
860 &Xloop_ssse3(\&body_20_39);
861 &Xloop_ssse3(\&body_20_39);
864 add 0($ctx),$A # update context
871 mov @T[0],$B # magic seed
883 $j=$saved_j; @V=@saved_V;
885 &Xtail_ssse3(\&body_20_39);
886 &Xtail_ssse3(\&body_20_39);
887 &Xtail_ssse3(\&body_20_39);
890 add 0($ctx),$A # update context
901 $code.=<<___ if ($win64);
902 movaps -40-6*16(%r14),%xmm6
903 movaps -40-5*16(%r14),%xmm7
904 movaps -40-4*16(%r14),%xmm8
905 movaps -40-3*16(%r14),%xmm9
906 movaps -40-2*16(%r14),%xmm10
907 movaps -40-1*16(%r14),%xmm11
919 .size sha1_block_data_order_ssse3,.-sha1_block_data_order_ssse3
923 $Xi=4; # reset variables
924 @X=map("%xmm$_",(4..7,0..3));
925 @Tx=map("%xmm$_",(8..10));
929 my $done_avx_label=".Ldone_avx";
931 my $_rol=sub { &shld(@_[0],@_) };
932 my $_ror=sub { &shrd(@_[0],@_) };
935 .type sha1_block_data_order_avx,\@function,3
937 sha1_block_data_order_avx:
943 push %r13 # redundant, done to share Win64 SE handler
945 lea `-64-($win64?6*16:0)`(%rsp),%rsp
948 $code.=<<___ if ($win64);
949 vmovaps %xmm6,-40-6*16(%rax)
950 vmovaps %xmm7,-40-5*16(%rax)
951 vmovaps %xmm8,-40-4*16(%rax)
952 vmovaps %xmm9,-40-3*16(%rax)
953 vmovaps %xmm10,-40-2*16(%rax)
954 vmovaps %xmm11,-40-1*16(%rax)
958 mov %rax,%r14 # original %rsp
960 mov %rdi,$ctx # reassigned argument
961 mov %rsi,$inp # reassigned argument
962 mov %rdx,$num # reassigned argument
966 lea K_XX_XX+64(%rip),$K_XX_XX
968 mov 0($ctx),$A # load context
972 mov $B,@T[0] # magic seed
978 vmovdqa 64($K_XX_XX),@X[2] # pbswap mask
979 vmovdqa -64($K_XX_XX),$Kx # K_00_19
980 vmovdqu 0($inp),@X[-4&7] # load input to %xmm[0-3]
981 vmovdqu 16($inp),@X[-3&7]
982 vmovdqu 32($inp),@X[-2&7]
983 vmovdqu 48($inp),@X[-1&7]
984 vpshufb @X[2],@X[-4&7],@X[-4&7] # byte swap
986 vpshufb @X[2],@X[-3&7],@X[-3&7]
987 vpshufb @X[2],@X[-2&7],@X[-2&7]
988 vpshufb @X[2],@X[-1&7],@X[-1&7]
989 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19
990 vpaddd $Kx,@X[-3&7],@X[1]
991 vpaddd $Kx,@X[-2&7],@X[2]
992 vmovdqa @X[0],0(%rsp) # X[]+K xfer to IALU
993 vmovdqa @X[1],16(%rsp)
994 vmovdqa @X[2],32(%rsp)
998 sub Xupdate_avx_16_31() # recall that $Xi starts wtih 4
1001 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
1002 my ($a,$b,$c,$d,$e);
1004 eval(shift(@insns));
1005 eval(shift(@insns));
1006 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1007 eval(shift(@insns));
1008 eval(shift(@insns));
1010 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1011 eval(shift(@insns));
1012 eval(shift(@insns));
1013 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1014 eval(shift(@insns));
1015 eval(shift(@insns));
1016 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1017 eval(shift(@insns));
1018 eval(shift(@insns));
1020 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1021 eval(shift(@insns));
1022 eval(shift(@insns));
1023 eval(shift(@insns));
1024 eval(shift(@insns));
1026 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1027 eval(shift(@insns));
1028 eval(shift(@insns));
1029 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1030 eval(shift(@insns));
1031 eval(shift(@insns));
1033 &vpsrld (@Tx[0],@X[0],31);
1034 eval(shift(@insns));
1035 eval(shift(@insns));
1036 eval(shift(@insns));
1037 eval(shift(@insns));
1039 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1040 &vpaddd (@X[0],@X[0],@X[0]);
1041 eval(shift(@insns));
1042 eval(shift(@insns));
1043 eval(shift(@insns));
1044 eval(shift(@insns));
1046 &vpsrld (@Tx[1],@Tx[2],30);
1047 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1048 eval(shift(@insns));
1049 eval(shift(@insns));
1050 eval(shift(@insns));
1051 eval(shift(@insns));
1053 &vpslld (@Tx[2],@Tx[2],2);
1054 &vpxor (@X[0],@X[0],@Tx[1]);
1055 eval(shift(@insns));
1056 eval(shift(@insns));
1057 eval(shift(@insns));
1058 eval(shift(@insns));
1060 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1061 eval(shift(@insns));
1062 eval(shift(@insns));
1063 &vmovdqa ($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1064 eval(shift(@insns));
1065 eval(shift(@insns));
1068 foreach (@insns) { eval; } # remaining instructions [if any]
1070 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1073 sub Xupdate_avx_32_79()
1076 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
1077 my ($a,$b,$c,$d,$e);
1079 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1080 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1081 eval(shift(@insns)); # body_20_39
1082 eval(shift(@insns));
1083 eval(shift(@insns));
1084 eval(shift(@insns)); # rol
1086 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1087 eval(shift(@insns));
1088 eval(shift(@insns)) if (@insns[0] !~ /&ro[rl]/);
1089 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1090 &vmovdqa ($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0);
1091 eval(shift(@insns)); # ror
1092 eval(shift(@insns));
1094 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]"
1095 eval(shift(@insns)); # body_20_39
1096 eval(shift(@insns));
1097 eval(shift(@insns));
1098 eval(shift(@insns)); # rol
1100 &vpsrld (@Tx[0],@X[0],30);
1101 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1102 eval(shift(@insns));
1103 eval(shift(@insns));
1104 eval(shift(@insns)); # ror
1105 eval(shift(@insns));
1107 &vpslld (@X[0],@X[0],2);
1108 eval(shift(@insns)); # body_20_39
1109 eval(shift(@insns));
1110 eval(shift(@insns));
1111 eval(shift(@insns)); # rol
1112 eval(shift(@insns));
1113 eval(shift(@insns));
1114 eval(shift(@insns)); # ror
1115 eval(shift(@insns));
1117 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2
1118 eval(shift(@insns)); # body_20_39
1119 eval(shift(@insns));
1120 eval(shift(@insns));
1121 eval(shift(@insns)); # rol
1122 eval(shift(@insns));
1123 eval(shift(@insns));
1124 eval(shift(@insns)); # rol
1125 eval(shift(@insns));
1127 foreach (@insns) { eval; } # remaining instructions
1129 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1132 sub Xuplast_avx_80()
1135 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1136 my ($a,$b,$c,$d,$e);
1138 eval(shift(@insns));
1139 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1140 eval(shift(@insns));
1141 eval(shift(@insns));
1142 eval(shift(@insns));
1143 eval(shift(@insns));
1145 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU
1147 foreach (@insns) { eval; } # remaining instructions
1150 &je ($done_avx_label);
1152 &vmovdqa(@X[2],"64($K_XX_XX)"); # pbswap mask
1153 &vmovdqa($Kx,"-64($K_XX_XX)"); # K_00_19
1154 &vmovdqu(@X[-4&7],"0($inp)"); # load input
1155 &vmovdqu(@X[-3&7],"16($inp)");
1156 &vmovdqu(@X[-2&7],"32($inp)");
1157 &vmovdqu(@X[-1&7],"48($inp)");
1158 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1167 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1168 my ($a,$b,$c,$d,$e);
1170 eval(shift(@insns));
1171 eval(shift(@insns));
1172 &vpshufb(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1173 eval(shift(@insns));
1174 eval(shift(@insns));
1175 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],$Kx);
1176 eval(shift(@insns));
1177 eval(shift(@insns));
1178 eval(shift(@insns));
1179 eval(shift(@insns));
1180 &vmovdqa(eval(16*$Xi)."(%rsp)",@X[$Xi&7]); # X[]+K xfer to IALU
1181 eval(shift(@insns));
1182 eval(shift(@insns));
1184 foreach (@insns) { eval; }
1191 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1192 my ($a,$b,$c,$d,$e);
1194 foreach (@insns) { eval; }
1201 &Xupdate_avx_16_31(\&body_00_19);
1202 &Xupdate_avx_16_31(\&body_00_19);
1203 &Xupdate_avx_16_31(\&body_00_19);
1204 &Xupdate_avx_16_31(\&body_00_19);
1205 &Xupdate_avx_32_79(\&body_00_19);
1206 &Xupdate_avx_32_79(\&body_20_39);
1207 &Xupdate_avx_32_79(\&body_20_39);
1208 &Xupdate_avx_32_79(\&body_20_39);
1209 &Xupdate_avx_32_79(\&body_20_39);
1210 &Xupdate_avx_32_79(\&body_20_39);
1211 &Xupdate_avx_32_79(\&body_40_59);
1212 &Xupdate_avx_32_79(\&body_40_59);
1213 &Xupdate_avx_32_79(\&body_40_59);
1214 &Xupdate_avx_32_79(\&body_40_59);
1215 &Xupdate_avx_32_79(\&body_40_59);
1216 &Xupdate_avx_32_79(\&body_20_39);
1217 &Xuplast_avx_80(\&body_20_39); # can jump to "done"
1219 $saved_j=$j; @saved_V=@V;
1221 &Xloop_avx(\&body_20_39);
1222 &Xloop_avx(\&body_20_39);
1223 &Xloop_avx(\&body_20_39);
1226 add 0($ctx),$A # update context
1233 mov @T[0],$B # magic seed
1245 $j=$saved_j; @V=@saved_V;
1247 &Xtail_avx(\&body_20_39);
1248 &Xtail_avx(\&body_20_39);
1249 &Xtail_avx(\&body_20_39);
1254 add 0($ctx),$A # update context
1265 $code.=<<___ if ($win64);
1266 movaps -40-6*16(%r14),%xmm6
1267 movaps -40-5*16(%r14),%xmm7
1268 movaps -40-4*16(%r14),%xmm8
1269 movaps -40-3*16(%r14),%xmm9
1270 movaps -40-2*16(%r14),%xmm10
1271 movaps -40-1*16(%r14),%xmm11
1283 .size sha1_block_data_order_avx,.-sha1_block_data_order_avx
1288 $Xi=4; # reset variables
1289 @X=map("%ymm$_",(4..7,0..3));
1290 @Tx=map("%ymm$_",(8..10));
1294 my @ROTX=("%eax","%ebp","%ebx","%ecx","%edx","%esi");
1295 my ($a5,$t0)=("%r12d","%edi");
1297 my ($A,$F,$B,$C,$D,$E)=@ROTX;
1302 .type sha1_block_data_order_avx2,\@function,3
1304 sha1_block_data_order_avx2:
1314 $code.=<<___ if ($win64);
1315 lea -6*16(%rsp),%rsp
1316 vmovaps %xmm6,-40-6*16(%rax)
1317 vmovaps %xmm7,-40-5*16(%rax)
1318 vmovaps %xmm8,-40-4*16(%rax)
1319 vmovaps %xmm9,-40-3*16(%rax)
1320 vmovaps %xmm10,-40-2*16(%rax)
1321 vmovaps %xmm11,-40-1*16(%rax)
1325 mov %rax,%r14 # original %rsp
1326 mov %rdi,$ctx # reassigned argument
1327 mov %rsi,$inp # reassigned argument
1328 mov %rdx,$num # reassigned argument
1335 lea K_XX_XX+64(%rip),$K_XX_XX
1337 mov 0($ctx),$A # load context
1339 cmovae $inp,$frame # next or same block
1344 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask
1346 vmovdqu ($inp),%xmm0
1347 vmovdqu 16($inp),%xmm1
1348 vmovdqu 32($inp),%xmm2
1349 vmovdqu 48($inp),%xmm3
1351 vinserti128 \$1,($frame),@X[-4&7],@X[-4&7]
1352 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7]
1353 vpshufb @X[2],@X[-4&7],@X[-4&7]
1354 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7]
1355 vpshufb @X[2],@X[-3&7],@X[-3&7]
1356 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7]
1357 vpshufb @X[2],@X[-2&7],@X[-2&7]
1358 vmovdqu -64($K_XX_XX),$Kx # K_00_19
1359 vpshufb @X[2],@X[-1&7],@X[-1&7]
1361 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19
1362 vpaddd $Kx,@X[-3&7],@X[1]
1363 vmovdqu @X[0],0(%rsp) # X[]+K xfer to IALU
1364 vpaddd $Kx,@X[-2&7],@X[2]
1365 vmovdqu @X[1],32(%rsp)
1366 vpaddd $Kx,@X[-1&7],@X[3]
1367 vmovdqu @X[2],64(%rsp)
1368 vmovdqu @X[3],96(%rsp)
1370 for (;$Xi<8;$Xi++) { # Xupdate_avx2_16_31
1373 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1374 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1375 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1376 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1377 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1378 &vpsrld (@Tx[0],@X[0],31);
1379 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1380 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1381 &vpaddd (@X[0],@X[0],@X[0]);
1382 &vpsrld (@Tx[1],@Tx[2],30);
1383 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1384 &vpslld (@Tx[2],@Tx[2],2);
1385 &vpxor (@X[0],@X[0],@Tx[1]);
1386 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1387 &vpaddd (@Tx[1],@X[0],$Kx);
1388 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1390 push(@X,shift(@X)); # "rotate" X[]
1393 lea 128(%rsp),$frame
1402 sub bodyx_00_19 () { # 8 instructions, 3 cycles critical path
1403 # at start $f=(b&c)^(~b&d), $b>>>=2
1404 return &bodyx_20_39() if ($rx==19); $rx++;
1406 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1408 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1409 '&lea ($frame,"256($frame)") if ($j%32==31);',
1410 '&andn ($t0,$a,$c)', # ~b&d for next round
1412 '&add ($e,$f)', # e+=(b&c)^(~b&d)
1413 '&rorx ($a5,$a,27)', # a<<<5
1414 '&rorx ($f,$a,2)', # b>>>2 for next round
1415 '&and ($a,$b)', # b&c for next round
1417 '&add ($e,$a5)', # e+=a<<<5
1418 '&xor ($a,$t0);'. # f=(b&c)^(~b&d) for next round
1420 'unshift(@ROTX,pop(@ROTX)); $j++;'
1424 sub bodyx_20_39 () { # 7 instructions, 2 cycles critical path
1425 # on entry $f=b^c^d, $b>>>=2
1426 return &bodyx_40_59() if ($rx==39); $rx++;
1428 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1430 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1431 '&lea ($frame,"256($frame)") if ($j%32==31);',
1433 '&lea ($e,"($e,$f)")', # e+=b^c^d
1434 '&rorx ($a5,$a,27)', # a<<<5
1435 '&rorx ($f,$a,2) if ($j<79)', # b>>>2 in next round
1436 '&xor ($a,$b) if ($j<79)', # b^c for next round
1438 '&add ($e,$a5)', # e+=a<<<5
1439 '&xor ($a,$c) if ($j<79);'. # f=b^c^d for next round
1441 'unshift(@ROTX,pop(@ROTX)); $j++;'
1445 sub bodyx_40_59 () { # 10 instructions, 3 cycles critical path
1446 # on entry $f=((b^c)&(c^d)), $b>>>=2
1449 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1451 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1452 '&lea ($frame,"256($frame)") if ($j%32==31);',
1453 '&xor ($f,$c) if ($j>39)', # (b^c)&(c^d)^c
1454 '&mov ($t0,$b) if ($j<59)', # count on zero latency
1455 '&xor ($t0,$c) if ($j<59)', # c^d for next round
1457 '&lea ($e,"($e,$f)")', # e+=(b^c)&(c^d)^c
1458 '&rorx ($a5,$a,27)', # a<<<5
1459 '&rorx ($f,$a,2)', # b>>>2 in next round
1460 '&xor ($a,$b)', # b^c for next round
1462 '&add ($e,$a5)', # e+=a<<<5
1463 '&and ($a,$t0) if ($j< 59);'. # f=(b^c)&(c^d) for next round
1464 '&xor ($a,$c) if ($j==59);'. # f=b^c^d for next round
1466 'unshift(@ROTX,pop(@ROTX)); $j++;'
1470 sub Xupdate_avx2_16_31() # recall that $Xi starts wtih 4
1473 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 instructions
1474 my ($a,$b,$c,$d,$e);
1476 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1477 eval(shift(@insns));
1478 eval(shift(@insns));
1479 eval(shift(@insns));
1480 eval(shift(@insns));
1482 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1483 eval(shift(@insns));
1484 eval(shift(@insns));
1485 eval(shift(@insns));
1487 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1488 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1489 eval(shift(@insns));
1490 eval(shift(@insns));
1492 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1493 eval(shift(@insns));
1494 eval(shift(@insns));
1495 eval(shift(@insns));
1496 eval(shift(@insns));
1498 &vpsrld (@Tx[0],@X[0],31);
1499 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1500 eval(shift(@insns));
1501 eval(shift(@insns));
1502 eval(shift(@insns));
1504 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1505 &vpaddd (@X[0],@X[0],@X[0]);
1506 eval(shift(@insns));
1507 eval(shift(@insns));
1509 &vpsrld (@Tx[1],@Tx[2],30);
1510 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1511 eval(shift(@insns));
1512 eval(shift(@insns));
1514 &vpslld (@Tx[2],@Tx[2],2);
1515 &vpxor (@X[0],@X[0],@Tx[1]);
1516 eval(shift(@insns));
1517 eval(shift(@insns));
1519 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1520 eval(shift(@insns));
1521 eval(shift(@insns));
1522 eval(shift(@insns));
1524 &vpaddd (@Tx[1],@X[0],$Kx);
1525 eval(shift(@insns));
1526 eval(shift(@insns));
1527 eval(shift(@insns));
1528 &vmovdqu(eval(32*($Xi))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1530 foreach (@insns) { eval; } # remaining instructions [if any]
1533 push(@X,shift(@X)); # "rotate" X[]
1536 sub Xupdate_avx2_32_79()
1539 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 to 50 instructions
1540 my ($a,$b,$c,$d,$e);
1542 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1543 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1544 eval(shift(@insns));
1545 eval(shift(@insns));
1547 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1548 &vmovdqu($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0);
1549 eval(shift(@insns));
1550 eval(shift(@insns));
1551 eval(shift(@insns));
1553 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]"
1554 eval(shift(@insns));
1555 eval(shift(@insns));
1556 eval(shift(@insns));
1558 &vpsrld (@Tx[0],@X[0],30);
1559 &vpslld (@X[0],@X[0],2);
1560 eval(shift(@insns));
1561 eval(shift(@insns));
1562 eval(shift(@insns));
1564 #&vpslld (@X[0],@X[0],2);
1565 eval(shift(@insns));
1566 eval(shift(@insns));
1567 eval(shift(@insns));
1569 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2
1570 eval(shift(@insns));
1571 eval(shift(@insns));
1572 eval(shift(@insns));
1573 eval(shift(@insns));
1575 &vpaddd (@Tx[1],@X[0],$Kx);
1576 eval(shift(@insns));
1577 eval(shift(@insns));
1578 eval(shift(@insns));
1579 eval(shift(@insns));
1581 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1583 foreach (@insns) { eval; } # remaining instructions
1586 push(@X,shift(@X)); # "rotate" X[]
1592 my @insns = (&$body,&$body,&$body,&$body,&$body); # 32 instructions
1593 my ($a,$b,$c,$d,$e);
1595 foreach (@insns) { eval; }
1599 &Xupdate_avx2_32_79(\&bodyx_00_19);
1600 &Xupdate_avx2_32_79(\&bodyx_00_19);
1601 &Xupdate_avx2_32_79(\&bodyx_00_19);
1602 &Xupdate_avx2_32_79(\&bodyx_00_19);
1604 &Xupdate_avx2_32_79(\&bodyx_20_39);
1605 &Xupdate_avx2_32_79(\&bodyx_20_39);
1606 &Xupdate_avx2_32_79(\&bodyx_20_39);
1607 &Xupdate_avx2_32_79(\&bodyx_20_39);
1610 &Xupdate_avx2_32_79(\&bodyx_40_59);
1611 &Xupdate_avx2_32_79(\&bodyx_40_59);
1612 &Xupdate_avx2_32_79(\&bodyx_40_59);
1613 &Xupdate_avx2_32_79(\&bodyx_40_59);
1615 &Xloop_avx2(\&bodyx_20_39);
1616 &Xloop_avx2(\&bodyx_20_39);
1617 &Xloop_avx2(\&bodyx_20_39);
1618 &Xloop_avx2(\&bodyx_20_39);
1621 lea 128($inp),$frame
1622 lea 128($inp),%rdi # borrow $t0
1624 cmovae $inp,$frame # next or previous block
1626 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c
1627 add 0($ctx),@ROTX[0] # update context
1628 add 4($ctx),@ROTX[1]
1629 add 8($ctx),@ROTX[3]
1630 mov @ROTX[0],0($ctx)
1631 add 12($ctx),@ROTX[4]
1632 mov @ROTX[1],4($ctx)
1633 mov @ROTX[0],$A # A=d
1634 add 16($ctx),@ROTX[5]
1636 mov @ROTX[3],8($ctx)
1637 mov @ROTX[4],$D # D=b
1638 #xchg @ROTX[5],$F # F=c, C=f
1639 mov @ROTX[4],12($ctx)
1640 mov @ROTX[1],$F # F=e
1641 mov @ROTX[5],16($ctx)
1643 mov @ROTX[5],$E # E=c
1645 #xchg $F,$E # E=c, F=e
1651 $Xi=4; # reset variables
1652 @X=map("%ymm$_",(4..7,0..3));
1655 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask
1656 cmp $num,%rdi # borrowed $t0
1659 vmovdqu -64(%rdi),%xmm0 # low part of @X[-4&7]
1660 vmovdqu -48(%rdi),%xmm1
1661 vmovdqu -32(%rdi),%xmm2
1662 vmovdqu -16(%rdi),%xmm3
1663 vinserti128 \$1,0($frame),@X[-4&7],@X[-4&7]
1664 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7]
1665 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7]
1666 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7]
1671 lea 128+16(%rsp),$frame
1678 $rx=$j=0; @ROTX=($A,$F,$B,$C,$D,$E);
1680 &Xloop_avx2 (\&bodyx_00_19);
1681 &Xloop_avx2 (\&bodyx_00_19);
1682 &Xloop_avx2 (\&bodyx_00_19);
1683 &Xloop_avx2 (\&bodyx_00_19);
1685 &Xloop_avx2 (\&bodyx_20_39);
1686 &vmovdqu ($Kx,"-64($K_XX_XX)"); # K_00_19
1687 &vpshufb (@X[-4&7],@X[-4&7],@X[2]); # byte swap
1688 &Xloop_avx2 (\&bodyx_20_39);
1689 &vpshufb (@X[-3&7],@X[-3&7],@X[2]);
1690 &vpaddd (@Tx[0],@X[-4&7],$Kx); # add K_00_19
1691 &Xloop_avx2 (\&bodyx_20_39);
1692 &vmovdqu ("0(%rsp)",@Tx[0]);
1693 &vpshufb (@X[-2&7],@X[-2&7],@X[2]);
1694 &vpaddd (@Tx[1],@X[-3&7],$Kx);
1695 &Xloop_avx2 (\&bodyx_20_39);
1696 &vmovdqu ("32(%rsp)",@Tx[1]);
1697 &vpshufb (@X[-1&7],@X[-1&7],@X[2]);
1698 &vpaddd (@X[2],@X[-2&7],$Kx);
1700 &Xloop_avx2 (\&bodyx_40_59);
1702 &vmovdqu ("64(%rsp)",@X[2]);
1703 &vpaddd (@X[3],@X[-1&7],$Kx);
1704 &Xloop_avx2 (\&bodyx_40_59);
1705 &vmovdqu ("96(%rsp)",@X[3]);
1706 &Xloop_avx2 (\&bodyx_40_59);
1707 &Xupdate_avx2_16_31(\&bodyx_40_59);
1709 &Xupdate_avx2_16_31(\&bodyx_20_39);
1710 &Xupdate_avx2_16_31(\&bodyx_20_39);
1711 &Xupdate_avx2_16_31(\&bodyx_20_39);
1712 &Xloop_avx2 (\&bodyx_20_39);
1715 lea 128(%rsp),$frame
1717 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c
1718 add 0($ctx),@ROTX[0] # update context
1719 add 4($ctx),@ROTX[1]
1720 add 8($ctx),@ROTX[3]
1721 mov @ROTX[0],0($ctx)
1722 add 12($ctx),@ROTX[4]
1723 mov @ROTX[1],4($ctx)
1724 mov @ROTX[0],$A # A=d
1725 add 16($ctx),@ROTX[5]
1727 mov @ROTX[3],8($ctx)
1728 mov @ROTX[4],$D # D=b
1729 #xchg @ROTX[5],$F # F=c, C=f
1730 mov @ROTX[4],12($ctx)
1731 mov @ROTX[1],$F # F=e
1732 mov @ROTX[5],16($ctx)
1734 mov @ROTX[5],$E # E=c
1736 #xchg $F,$E # E=c, F=e
1744 $code.=<<___ if ($win64);
1745 movaps -40-6*16(%r14),%xmm6
1746 movaps -40-5*16(%r14),%xmm7
1747 movaps -40-4*16(%r14),%xmm8
1748 movaps -40-3*16(%r14),%xmm9
1749 movaps -40-2*16(%r14),%xmm10
1750 movaps -40-1*16(%r14),%xmm11
1762 .size sha1_block_data_order_avx2,.-sha1_block_data_order_avx2
1769 .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19
1770 .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19
1771 .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39
1772 .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39
1773 .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59
1774 .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59
1775 .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79
1776 .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79
1777 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask
1778 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask
1779 .byte 0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0
1783 .asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1787 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1788 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
1796 .extern __imp_RtlVirtualUnwind
1797 .type se_handler,\@abi-omnipotent
1811 mov 120($context),%rax # pull context->Rax
1812 mov 248($context),%rbx # pull context->Rip
1814 lea .Lprologue(%rip),%r10
1815 cmp %r10,%rbx # context->Rip<.Lprologue
1816 jb .Lcommon_seh_tail
1818 mov 152($context),%rax # pull context->Rsp
1820 lea .Lepilogue(%rip),%r10
1821 cmp %r10,%rbx # context->Rip>=.Lepilogue
1822 jae .Lcommon_seh_tail
1824 mov `16*4`(%rax),%rax # pull saved stack pointer
1831 mov %rbx,144($context) # restore context->Rbx
1832 mov %rbp,160($context) # restore context->Rbp
1833 mov %r12,216($context) # restore context->R12
1834 mov %r13,224($context) # restore context->R13
1835 mov %r14,232($context) # restore context->R14
1837 jmp .Lcommon_seh_tail
1838 .size se_handler,.-se_handler
1841 $code.=<<___ if ($shaext);
1842 .type shaext_handler,\@abi-omnipotent
1856 mov 120($context),%rax # pull context->Rax
1857 mov 248($context),%rbx # pull context->Rip
1859 lea .Lprologue_shaext(%rip),%r10
1860 cmp %r10,%rbx # context->Rip<.Lprologue
1861 jb .Lcommon_seh_tail
1863 lea .Lepilogue_shaext(%rip),%r10
1864 cmp %r10,%rbx # context->Rip>=.Lepilogue
1865 jae .Lcommon_seh_tail
1867 lea -8-4*16(%rax),%rsi
1868 lea 512($context),%rdi # &context.Xmm6
1870 .long 0xa548f3fc # cld; rep movsq
1872 jmp .Lcommon_seh_tail
1873 .size shaext_handler,.-shaext_handler
1877 .type ssse3_handler,\@abi-omnipotent
1891 mov 120($context),%rax # pull context->Rax
1892 mov 248($context),%rbx # pull context->Rip
1894 mov 8($disp),%rsi # disp->ImageBase
1895 mov 56($disp),%r11 # disp->HandlerData
1897 mov 0(%r11),%r10d # HandlerData[0]
1898 lea (%rsi,%r10),%r10 # prologue label
1899 cmp %r10,%rbx # context->Rip<prologue label
1900 jb .Lcommon_seh_tail
1902 mov 152($context),%rax # pull context->Rsp
1904 mov 4(%r11),%r10d # HandlerData[1]
1905 lea (%rsi,%r10),%r10 # epilogue label
1906 cmp %r10,%rbx # context->Rip>=epilogue label
1907 jae .Lcommon_seh_tail
1909 mov 232($context),%rax # pull context->R14
1911 lea -40-6*16(%rax),%rsi
1912 lea 512($context),%rdi # &context.Xmm6
1914 .long 0xa548f3fc # cld; rep movsq
1921 mov %rbx,144($context) # restore context->Rbx
1922 mov %rbp,160($context) # restore context->Rbp
1923 mov %r12,216($context) # restore cotnext->R12
1924 mov %r13,224($context) # restore cotnext->R13
1925 mov %r14,232($context) # restore cotnext->R14
1930 mov %rax,152($context) # restore context->Rsp
1931 mov %rsi,168($context) # restore context->Rsi
1932 mov %rdi,176($context) # restore context->Rdi
1934 mov 40($disp),%rdi # disp->ContextRecord
1935 mov $context,%rsi # context
1936 mov \$154,%ecx # sizeof(CONTEXT)
1937 .long 0xa548f3fc # cld; rep movsq
1940 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
1941 mov 8(%rsi),%rdx # arg2, disp->ImageBase
1942 mov 0(%rsi),%r8 # arg3, disp->ControlPc
1943 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
1944 mov 40(%rsi),%r10 # disp->ContextRecord
1945 lea 56(%rsi),%r11 # &disp->HandlerData
1946 lea 24(%rsi),%r12 # &disp->EstablisherFrame
1947 mov %r10,32(%rsp) # arg5
1948 mov %r11,40(%rsp) # arg6
1949 mov %r12,48(%rsp) # arg7
1950 mov %rcx,56(%rsp) # arg8, (NULL)
1951 call *__imp_RtlVirtualUnwind(%rip)
1953 mov \$1,%eax # ExceptionContinueSearch
1965 .size ssse3_handler,.-ssse3_handler
1969 .rva .LSEH_begin_sha1_block_data_order
1970 .rva .LSEH_end_sha1_block_data_order
1971 .rva .LSEH_info_sha1_block_data_order
1973 $code.=<<___ if ($shaext);
1974 .rva .LSEH_begin_sha1_block_data_order_shaext
1975 .rva .LSEH_end_sha1_block_data_order_shaext
1976 .rva .LSEH_info_sha1_block_data_order_shaext
1979 .rva .LSEH_begin_sha1_block_data_order_ssse3
1980 .rva .LSEH_end_sha1_block_data_order_ssse3
1981 .rva .LSEH_info_sha1_block_data_order_ssse3
1983 $code.=<<___ if ($avx);
1984 .rva .LSEH_begin_sha1_block_data_order_avx
1985 .rva .LSEH_end_sha1_block_data_order_avx
1986 .rva .LSEH_info_sha1_block_data_order_avx
1988 $code.=<<___ if ($avx>1);
1989 .rva .LSEH_begin_sha1_block_data_order_avx2
1990 .rva .LSEH_end_sha1_block_data_order_avx2
1991 .rva .LSEH_info_sha1_block_data_order_avx2
1996 .LSEH_info_sha1_block_data_order:
2000 $code.=<<___ if ($shaext);
2001 .LSEH_info_sha1_block_data_order_shaext:
2006 .LSEH_info_sha1_block_data_order_ssse3:
2009 .rva .Lprologue_ssse3,.Lepilogue_ssse3 # HandlerData[]
2011 $code.=<<___ if ($avx);
2012 .LSEH_info_sha1_block_data_order_avx:
2015 .rva .Lprologue_avx,.Lepilogue_avx # HandlerData[]
2017 $code.=<<___ if ($avx>1);
2018 .LSEH_info_sha1_block_data_order_avx2:
2021 .rva .Lprologue_avx2,.Lepilogue_avx2 # HandlerData[]
2025 ####################################################################
2028 if (@_[0] =~ /\$([x0-9a-f]+),\s*%xmm([0-7]),\s*%xmm([0-7])/) {
2029 my @opcode=(0x0f,0x3a,0xcc);
2030 push @opcode,0xc0|($2&7)|(($3&7)<<3); # ModR/M
2032 push @opcode,$c=~/^0/?oct($c):$c;
2033 return ".byte\t".join(',',@opcode);
2035 return "sha1rnds4\t".@_[0];
2042 "sha1nexte" => 0xc8,
2044 "sha1msg2" => 0xca );
2046 if (defined($opcodelet{$instr}) && @_[0] =~ /%xmm([0-9]+),\s*%xmm([0-9]+)/) {
2047 my @opcode=(0x0f,0x38);
2049 $rex|=0x04 if ($2>=8);
2050 $rex|=0x01 if ($1>=8);
2051 unshift @opcode,0x40|$rex if ($rex);
2052 push @opcode,$opcodelet{$instr};
2053 push @opcode,0xc0|($1&7)|(($2&7)<<3); # ModR/M
2054 return ".byte\t".join(',',@opcode);
2056 return $instr."\t".@_[0];
2060 foreach (split("\n",$code)) {
2061 s/\`([^\`]*)\`/eval $1/geo;
2063 s/\b(sha1rnds4)\s+(.*)/sha1rnds4($2)/geo or
2064 s/\b(sha1[^\s]*)\s+(.*)/sha1op38($1,$2)/geo;