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15 # assembler, which originally made it hard to motivate the effort.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 # Bulldozer 9.11 5.95/+53%
77 # VIA Nano 9.32 7.15/+30%
79 # Silvermont 13.1(*) 9.37/+40%
81 # (*) obviously suboptimal result, nothing was done about it,
82 # because SSSE3 code is compiled unconditionally;
86 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
88 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
90 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
91 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
92 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
93 die "can't locate x86_64-xlate.pl";
95 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
96 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
97 $avx = ($1>=2.19) + ($1>=2.22);
100 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
101 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
102 $avx = ($1>=2.09) + ($1>=2.10);
105 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
106 `ml64 2>&1` =~ /Version ([0-9]+)\./) {
107 $avx = ($1>=10) + ($1>=11);
110 open OUT,"| \"$^X\" $xlate $flavour $output";
113 $ctx="%rdi"; # 1st arg
114 $inp="%rsi"; # 2nd arg
115 $num="%rdx"; # 3rd arg
117 # reassign arguments in order to produce more compact code
125 @xi=("%edx","%ebp","%r14d");
135 my ($i,$a,$b,$c,$d,$e)=@_;
137 $code.=<<___ if ($i==0);
138 mov `4*$i`($inp),$xi[0]
141 $code.=<<___ if ($i<15);
142 mov `4*$j`($inp),$xi[1]
144 mov $xi[0],`4*$i`(%rsp)
150 lea 0x5a827999($xi[0],$e),$e
156 $code.=<<___ if ($i>=15);
157 xor `4*($j%16)`(%rsp),$xi[1]
159 mov $xi[0],`4*($i%16)`(%rsp)
161 xor `4*(($j+2)%16)`(%rsp),$xi[1]
164 xor `4*(($j+8)%16)`(%rsp),$xi[1]
166 lea 0x5a827999($xi[0],$e),$e
173 push(@xi,shift(@xi));
177 my ($i,$a,$b,$c,$d,$e)=@_;
179 my $K=($i<40)?0x6ed9eba1:0xca62c1d6;
180 $code.=<<___ if ($i<79);
181 xor `4*($j%16)`(%rsp),$xi[1]
183 `"mov $xi[0],".4*($i%16)."(%rsp)" if ($i<72)`
185 xor `4*(($j+2)%16)`(%rsp),$xi[1]
188 xor `4*(($j+8)%16)`(%rsp),$xi[1]
196 $code.=<<___ if ($i==79);
207 push(@xi,shift(@xi));
211 my ($i,$a,$b,$c,$d,$e)=@_;
214 xor `4*($j%16)`(%rsp),$xi[1]
216 mov $xi[0],`4*($i%16)`(%rsp)
218 xor `4*(($j+2)%16)`(%rsp),$xi[1]
221 xor `4*(($j+8)%16)`(%rsp),$xi[1]
222 lea 0x8f1bbcdc($xi[0],$e),$e
232 push(@xi,shift(@xi));
237 .extern OPENSSL_ia32cap_P
239 .globl sha1_block_data_order
240 .type sha1_block_data_order,\@function,3
242 sha1_block_data_order:
243 mov OPENSSL_ia32cap_P+0(%rip),%r9d
244 mov OPENSSL_ia32cap_P+4(%rip),%r8d
245 mov OPENSSL_ia32cap_P+8(%rip),%r10d
246 test \$`1<<9`,%r8d # check SSSE3 bit
249 test \$`1<<29`,%r10d # check SHA bit
252 $code.=<<___ if ($avx>1);
253 and \$`1<<3|1<<5|1<<8`,%r10d # check AVX2+BMI1+BMI2
254 cmp \$`1<<3|1<<5|1<<8`,%r10d
257 $code.=<<___ if ($avx);
258 and \$`1<<28`,%r8d # mask AVX bit
259 and \$`1<<30`,%r9d # mask "Intel CPU" bit
261 cmp \$`1<<28|1<<30`,%r8d
275 mov %rdi,$ctx # reassigned argument
277 mov %rsi,$inp # reassigned argument
279 mov %rdx,$num # reassigned argument
280 mov %rax,`16*4`(%rsp)
293 for($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); }
294 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
295 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
296 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
310 lea `16*4`($inp),$inp
313 mov `16*4`(%rsp),%rsi
322 .size sha1_block_data_order,.-sha1_block_data_order
325 ######################################################################
326 # Intel SHA Extensions implementation of SHA1 update function.
328 my ($ctx,$inp,$num)=("%rdi","%rsi","%rdx");
329 my ($ABCD,$E,$E_,$BSWAP,$ABCD_SAVE,$E_SAVE)=map("%xmm$_",(0..3,8,9));
330 my @MSG=map("%xmm$_",(4..7));
333 .type sha1_block_data_order_shaext,\@function,3
335 sha1_block_data_order_shaext:
338 $code.=<<___ if ($win64);
339 lea `-8-4*16`(%rsp),%rsp
340 movaps %xmm6,-8-4*16(%rax)
341 movaps %xmm7,-8-3*16(%rax)
342 movaps %xmm8,-8-2*16(%rax)
343 movaps %xmm9,-8-1*16(%rax)
349 movdqa K_XX_XX+0xa0(%rip),$BSWAP # byte-n-word swap
351 movdqu ($inp),@MSG[0]
352 pshufd \$0b00011011,$ABCD,$ABCD # flip word order
353 movdqu 0x10($inp),@MSG[1]
354 pshufd \$0b00011011,$E,$E # flip word order
355 movdqu 0x20($inp),@MSG[2]
356 pshufb $BSWAP,@MSG[0]
357 movdqu 0x30($inp),@MSG[3]
358 pshufb $BSWAP,@MSG[1]
359 pshufb $BSWAP,@MSG[2]
360 movdqa $E,$E_SAVE # offload $E
361 pshufb $BSWAP,@MSG[3]
367 lea 0x40($inp),%rax # next input block
370 movdqa $ABCD,$ABCD_SAVE # offload $ABCD
372 for($i=0;$i<20-4;$i+=2) {
374 sha1msg1 @MSG[1],@MSG[0]
376 sha1rnds4 \$`int($i/5)`,$E,$ABCD # 0-3...
377 sha1nexte @MSG[1],$E_
379 sha1msg1 @MSG[2],@MSG[1]
380 sha1msg2 @MSG[3],@MSG[0]
383 sha1rnds4 \$`int(($i+1)/5)`,$E_,$ABCD
386 sha1msg2 @MSG[0],@MSG[1]
388 push(@MSG,shift(@MSG)); push(@MSG,shift(@MSG));
391 movdqu ($inp),@MSG[0]
393 sha1rnds4 \$3,$E,$ABCD # 64-67
394 sha1nexte @MSG[1],$E_
395 movdqu 0x10($inp),@MSG[1]
396 pshufb $BSWAP,@MSG[0]
399 sha1rnds4 \$3,$E_,$ABCD # 68-71
401 movdqu 0x20($inp),@MSG[2]
402 pshufb $BSWAP,@MSG[1]
405 sha1rnds4 \$3,$E,$ABCD # 72-75
406 sha1nexte @MSG[3],$E_
407 movdqu 0x30($inp),@MSG[3]
408 pshufb $BSWAP,@MSG[2]
411 sha1rnds4 \$3,$E_,$ABCD # 76-79
413 pshufb $BSWAP,@MSG[3]
415 paddd $ABCD_SAVE,$ABCD
416 movdqa $E,$E_SAVE # offload $E
420 pshufd \$0b00011011,$ABCD,$ABCD
421 pshufd \$0b00011011,$E,$E
425 $code.=<<___ if ($win64);
426 movaps -8-4*16(%rax),%xmm6
427 movaps -8-3*16(%rax),%xmm7
428 movaps -8-2*16(%rax),%xmm8
429 movaps -8-1*16(%rax),%xmm9
435 .size sha1_block_data_order_shaext,.-sha1_block_data_order_shaext
440 my @X=map("%xmm$_",(4..7,0..3));
441 my @Tx=map("%xmm$_",(8..10));
443 my @V=($A,$B,$C,$D,$E)=("%eax","%ebx","%ecx","%edx","%ebp"); # size optimization
444 my @T=("%esi","%edi");
449 my $_rol=sub { &rol(@_) };
450 my $_ror=sub { &ror(@_) };
456 jmp .Lalign32_$sn # see "Decoded ICache" in manual
464 .type sha1_block_data_order_ssse3,\@function,3
466 sha1_block_data_order_ssse3:
472 push %r13 # redundant, done to share Win64 SE handler
474 lea `-64-($win64?6*16:0)`(%rsp),%rsp
476 $code.=<<___ if ($win64);
477 movaps %xmm6,-40-6*16(%rax)
478 movaps %xmm7,-40-5*16(%rax)
479 movaps %xmm8,-40-4*16(%rax)
480 movaps %xmm9,-40-3*16(%rax)
481 movaps %xmm10,-40-2*16(%rax)
482 movaps %xmm11,-40-1*16(%rax)
486 mov %rax,%r14 # original %rsp
488 mov %rdi,$ctx # reassigned argument
489 mov %rsi,$inp # reassigned argument
490 mov %rdx,$num # reassigned argument
494 lea K_XX_XX+64(%rip),$K_XX_XX
496 mov 0($ctx),$A # load context
500 mov $B,@T[0] # magic seed
506 movdqa 64($K_XX_XX),@X[2] # pbswap mask
507 movdqa -64($K_XX_XX),@Tx[1] # K_00_19
508 movdqu 0($inp),@X[-4&7] # load input to %xmm[0-3]
509 movdqu 16($inp),@X[-3&7]
510 movdqu 32($inp),@X[-2&7]
511 movdqu 48($inp),@X[-1&7]
512 pshufb @X[2],@X[-4&7] # byte swap
513 pshufb @X[2],@X[-3&7]
514 pshufb @X[2],@X[-2&7]
516 paddd @Tx[1],@X[-4&7] # add K_00_19
517 pshufb @X[2],@X[-1&7]
518 paddd @Tx[1],@X[-3&7]
519 paddd @Tx[1],@X[-2&7]
520 movdqa @X[-4&7],0(%rsp) # X[]+K xfer to IALU
521 psubd @Tx[1],@X[-4&7] # restore X[]
522 movdqa @X[-3&7],16(%rsp)
523 psubd @Tx[1],@X[-3&7]
524 movdqa @X[-2&7],32(%rsp)
525 psubd @Tx[1],@X[-2&7]
529 sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm
530 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
532 $arg = "\$$arg" if ($arg*1 eq $arg);
533 $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
536 sub Xupdate_ssse3_16_31() # recall that $Xi starts wtih 4
539 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
542 eval(shift(@insns)); # ror
543 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
545 &movdqa (@Tx[0],@X[-1&7]);
546 &paddd (@Tx[1],@X[-1&7]);
550 &punpcklqdq(@X[0],@X[-3&7]); # compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8);
552 eval(shift(@insns)); # rol
554 &psrldq (@Tx[0],4); # "X[-3]", 3 dwords
558 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
560 eval(shift(@insns)); # ror
561 &pxor (@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
566 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
568 eval(shift(@insns)); # rol
569 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
573 &movdqa (@Tx[2],@X[0]);
576 eval(shift(@insns)); # ror
577 &movdqa (@Tx[0],@X[0]);
580 &pslldq (@Tx[2],12); # "X[0]"<<96, extract one dword
581 &paddd (@X[0],@X[0]);
587 eval(shift(@insns)); # rol
589 &movdqa (@Tx[1],@Tx[2]);
595 eval(shift(@insns)); # ror
596 &por (@X[0],@Tx[0]); # "X[0]"<<<=1
602 &pxor (@X[0],@Tx[2]);
604 &movdqa (@Tx[2],eval(2*16*(($Xi)/5)-64)."($K_XX_XX)"); # K_XX_XX
605 eval(shift(@insns)); # rol
609 &pxor (@X[0],@Tx[1]); # "X[0]"^=("X[0]">>96)<<<2
610 &pshufd (@Tx[1],@X[-1&7],0xee) if ($Xi==7); # was &movdqa (@Tx[0],@X[-1&7]) in Xupdate_ssse3_32_79
612 foreach (@insns) { eval; } # remaining instructions [if any]
614 $Xi++; push(@X,shift(@X)); # "rotate" X[]
615 push(@Tx,shift(@Tx));
618 sub Xupdate_ssse3_32_79()
621 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
624 eval(shift(@insns)) if ($Xi==8);
625 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
626 eval(shift(@insns)) if ($Xi==8);
627 eval(shift(@insns)); # body_20_39
629 eval(shift(@insns)) if (@insns[1] =~ /_ror/);
630 eval(shift(@insns)) if (@insns[0] =~ /_ror/);
631 &punpcklqdq(@Tx[0],@X[-1&7]); # compose "X[-6]", was &palignr(@Tx[0],@X[-2&7],8);
633 eval(shift(@insns)); # rol
635 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
639 &movdqa (@Tx[2],@Tx[1]);# "perpetuate" K_XX_XX...
640 } else { # ... or load next one
641 &movdqa (@Tx[2],eval(2*16*($Xi/5)-64)."($K_XX_XX)");
643 eval(shift(@insns)); # ror
644 &paddd (@Tx[1],@X[-1&7]);
647 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-6]"
648 eval(shift(@insns)); # body_20_39
651 eval(shift(@insns)); # rol
652 eval(shift(@insns)) if (@insns[0] =~ /_ror/);
654 &movdqa (@Tx[0],@X[0]);
657 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
658 eval(shift(@insns)); # ror
660 eval(shift(@insns)); # body_20_39
666 eval(shift(@insns)) if (@insns[0] =~ /_rol/);# rol
669 eval(shift(@insns)); # ror
671 &por (@X[0],@Tx[0]); # "X[0]"<<<=2
673 eval(shift(@insns)); # body_20_39
674 eval(shift(@insns)) if (@insns[1] =~ /_rol/);
675 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
676 &pshufd(@Tx[1],@X[-1&7],0xee) if ($Xi<19); # was &movdqa (@Tx[1],@X[0])
678 eval(shift(@insns)); # rol
681 eval(shift(@insns)); # rol
684 foreach (@insns) { eval; } # remaining instructions
686 $Xi++; push(@X,shift(@X)); # "rotate" X[]
687 push(@Tx,shift(@Tx));
690 sub Xuplast_ssse3_80()
693 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
700 &paddd (@Tx[1],@X[-1&7]);
704 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU
706 foreach (@insns) { eval; } # remaining instructions
709 &je (".Ldone_ssse3");
711 unshift(@Tx,pop(@Tx));
713 &movdqa (@X[2],"64($K_XX_XX)"); # pbswap mask
714 &movdqa (@Tx[1],"-64($K_XX_XX)"); # K_00_19
715 &movdqu (@X[-4&7],"0($inp)"); # load input
716 &movdqu (@X[-3&7],"16($inp)");
717 &movdqu (@X[-2&7],"32($inp)");
718 &movdqu (@X[-1&7],"48($inp)");
719 &pshufb (@X[-4&7],@X[2]); # byte swap
728 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
734 &pshufb (@X[($Xi-3)&7],@X[2]);
739 &paddd (@X[($Xi-4)&7],@Tx[1]);
744 &movdqa (eval(16*$Xi)."(%rsp)",@X[($Xi-4)&7]); # X[]+K xfer to IALU
749 &psubd (@X[($Xi-4)&7],@Tx[1]);
751 foreach (@insns) { eval; }
758 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
761 foreach (@insns) { eval; }
764 sub body_00_19 () { # ((c^d)&b)^d
765 # on start @T[0]=(c^d)&b
766 return &body_20_39() if ($rx==19); $rx++;
768 '($a,$b,$c,$d,$e)=@V;'.
769 '&$_ror ($b,$j?7:2)', # $b>>>2
771 '&mov (@T[1],$a)', # $b for next round
773 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
774 '&xor ($b,$c)', # $c^$d for next round
778 '&and (@T[1],$b)', # ($b&($c^$d)) for next round
780 '&xor ($b,$c)', # restore $b
781 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
785 sub body_20_39 () { # b^d^c
787 return &body_40_59() if ($rx==39); $rx++;
789 '($a,$b,$c,$d,$e)=@V;'.
790 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
791 '&xor (@T[0],$d) if($j==19);'.
792 '&xor (@T[0],$c) if($j> 19)', # ($b^$d^$c)
793 '&mov (@T[1],$a)', # $b for next round
797 '&xor (@T[1],$c) if ($j< 79)', # $b^$d for next round
799 '&$_ror ($b,7)', # $b>>>2
800 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
804 sub body_40_59 () { # ((b^c)&(c^d))^c
805 # on entry @T[0]=(b^c), (c^=d)
808 '($a,$b,$c,$d,$e)=@V;'.
809 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
810 '&and (@T[0],$c) if ($j>=40)', # (b^c)&(c^d)
811 '&xor ($c,$d) if ($j>=40)', # restore $c
813 '&$_ror ($b,7)', # $b>>>2
814 '&mov (@T[1],$a)', # $b for next round
819 '&xor (@T[1],$c) if ($j==59);'.
820 '&xor (@T[1],$b) if ($j< 59)', # b^c for next round
822 '&xor ($b,$c) if ($j< 59)', # c^d for next round
823 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
830 &Xupdate_ssse3_16_31(\&body_00_19);
831 &Xupdate_ssse3_16_31(\&body_00_19);
832 &Xupdate_ssse3_16_31(\&body_00_19);
833 &Xupdate_ssse3_16_31(\&body_00_19);
834 &Xupdate_ssse3_32_79(\&body_00_19);
835 &Xupdate_ssse3_32_79(\&body_20_39);
836 &Xupdate_ssse3_32_79(\&body_20_39);
837 &Xupdate_ssse3_32_79(\&body_20_39);
838 &Xupdate_ssse3_32_79(\&body_20_39);
839 &Xupdate_ssse3_32_79(\&body_20_39);
840 &Xupdate_ssse3_32_79(\&body_40_59);
841 &Xupdate_ssse3_32_79(\&body_40_59);
842 &Xupdate_ssse3_32_79(\&body_40_59);
843 &Xupdate_ssse3_32_79(\&body_40_59);
844 &Xupdate_ssse3_32_79(\&body_40_59);
845 &Xupdate_ssse3_32_79(\&body_20_39);
846 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done"
848 $saved_j=$j; @saved_V=@V;
850 &Xloop_ssse3(\&body_20_39);
851 &Xloop_ssse3(\&body_20_39);
852 &Xloop_ssse3(\&body_20_39);
855 add 0($ctx),$A # update context
862 mov @T[0],$B # magic seed
874 $j=$saved_j; @V=@saved_V;
876 &Xtail_ssse3(\&body_20_39);
877 &Xtail_ssse3(\&body_20_39);
878 &Xtail_ssse3(\&body_20_39);
881 add 0($ctx),$A # update context
892 $code.=<<___ if ($win64);
893 movaps -40-6*16(%r14),%xmm6
894 movaps -40-5*16(%r14),%xmm7
895 movaps -40-4*16(%r14),%xmm8
896 movaps -40-3*16(%r14),%xmm9
897 movaps -40-2*16(%r14),%xmm10
898 movaps -40-1*16(%r14),%xmm11
910 .size sha1_block_data_order_ssse3,.-sha1_block_data_order_ssse3
914 $Xi=4; # reset variables
915 @X=map("%xmm$_",(4..7,0..3));
916 @Tx=map("%xmm$_",(8..10));
920 my $done_avx_label=".Ldone_avx";
922 my $_rol=sub { &shld(@_[0],@_) };
923 my $_ror=sub { &shrd(@_[0],@_) };
926 .type sha1_block_data_order_avx,\@function,3
928 sha1_block_data_order_avx:
934 push %r13 # redundant, done to share Win64 SE handler
936 lea `-64-($win64?6*16:0)`(%rsp),%rsp
939 $code.=<<___ if ($win64);
940 vmovaps %xmm6,-40-6*16(%rax)
941 vmovaps %xmm7,-40-5*16(%rax)
942 vmovaps %xmm8,-40-4*16(%rax)
943 vmovaps %xmm9,-40-3*16(%rax)
944 vmovaps %xmm10,-40-2*16(%rax)
945 vmovaps %xmm11,-40-1*16(%rax)
949 mov %rax,%r14 # original %rsp
951 mov %rdi,$ctx # reassigned argument
952 mov %rsi,$inp # reassigned argument
953 mov %rdx,$num # reassigned argument
957 lea K_XX_XX+64(%rip),$K_XX_XX
959 mov 0($ctx),$A # load context
963 mov $B,@T[0] # magic seed
969 vmovdqa 64($K_XX_XX),@X[2] # pbswap mask
970 vmovdqa -64($K_XX_XX),$Kx # K_00_19
971 vmovdqu 0($inp),@X[-4&7] # load input to %xmm[0-3]
972 vmovdqu 16($inp),@X[-3&7]
973 vmovdqu 32($inp),@X[-2&7]
974 vmovdqu 48($inp),@X[-1&7]
975 vpshufb @X[2],@X[-4&7],@X[-4&7] # byte swap
977 vpshufb @X[2],@X[-3&7],@X[-3&7]
978 vpshufb @X[2],@X[-2&7],@X[-2&7]
979 vpshufb @X[2],@X[-1&7],@X[-1&7]
980 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19
981 vpaddd $Kx,@X[-3&7],@X[1]
982 vpaddd $Kx,@X[-2&7],@X[2]
983 vmovdqa @X[0],0(%rsp) # X[]+K xfer to IALU
984 vmovdqa @X[1],16(%rsp)
985 vmovdqa @X[2],32(%rsp)
989 sub Xupdate_avx_16_31() # recall that $Xi starts wtih 4
992 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
997 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1001 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1002 eval(shift(@insns));
1003 eval(shift(@insns));
1004 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1005 eval(shift(@insns));
1006 eval(shift(@insns));
1007 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1008 eval(shift(@insns));
1009 eval(shift(@insns));
1011 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1012 eval(shift(@insns));
1013 eval(shift(@insns));
1014 eval(shift(@insns));
1015 eval(shift(@insns));
1017 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1018 eval(shift(@insns));
1019 eval(shift(@insns));
1020 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1021 eval(shift(@insns));
1022 eval(shift(@insns));
1024 &vpsrld (@Tx[0],@X[0],31);
1025 eval(shift(@insns));
1026 eval(shift(@insns));
1027 eval(shift(@insns));
1028 eval(shift(@insns));
1030 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1031 &vpaddd (@X[0],@X[0],@X[0]);
1032 eval(shift(@insns));
1033 eval(shift(@insns));
1034 eval(shift(@insns));
1035 eval(shift(@insns));
1037 &vpsrld (@Tx[1],@Tx[2],30);
1038 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1039 eval(shift(@insns));
1040 eval(shift(@insns));
1041 eval(shift(@insns));
1042 eval(shift(@insns));
1044 &vpslld (@Tx[2],@Tx[2],2);
1045 &vpxor (@X[0],@X[0],@Tx[1]);
1046 eval(shift(@insns));
1047 eval(shift(@insns));
1048 eval(shift(@insns));
1049 eval(shift(@insns));
1051 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1052 eval(shift(@insns));
1053 eval(shift(@insns));
1054 &vmovdqa ($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1055 eval(shift(@insns));
1056 eval(shift(@insns));
1059 foreach (@insns) { eval; } # remaining instructions [if any]
1061 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1064 sub Xupdate_avx_32_79()
1067 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
1068 my ($a,$b,$c,$d,$e);
1070 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1071 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1072 eval(shift(@insns)); # body_20_39
1073 eval(shift(@insns));
1074 eval(shift(@insns));
1075 eval(shift(@insns)); # rol
1077 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1078 eval(shift(@insns));
1079 eval(shift(@insns)) if (@insns[0] !~ /&ro[rl]/);
1080 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1081 &vmovdqa ($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0);
1082 eval(shift(@insns)); # ror
1083 eval(shift(@insns));
1085 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]"
1086 eval(shift(@insns)); # body_20_39
1087 eval(shift(@insns));
1088 eval(shift(@insns));
1089 eval(shift(@insns)); # rol
1091 &vpsrld (@Tx[0],@X[0],30);
1092 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1093 eval(shift(@insns));
1094 eval(shift(@insns));
1095 eval(shift(@insns)); # ror
1096 eval(shift(@insns));
1098 &vpslld (@X[0],@X[0],2);
1099 eval(shift(@insns)); # body_20_39
1100 eval(shift(@insns));
1101 eval(shift(@insns));
1102 eval(shift(@insns)); # rol
1103 eval(shift(@insns));
1104 eval(shift(@insns));
1105 eval(shift(@insns)); # ror
1106 eval(shift(@insns));
1108 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2
1109 eval(shift(@insns)); # body_20_39
1110 eval(shift(@insns));
1111 eval(shift(@insns));
1112 eval(shift(@insns)); # rol
1113 eval(shift(@insns));
1114 eval(shift(@insns));
1115 eval(shift(@insns)); # rol
1116 eval(shift(@insns));
1118 foreach (@insns) { eval; } # remaining instructions
1120 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1123 sub Xuplast_avx_80()
1126 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1127 my ($a,$b,$c,$d,$e);
1129 eval(shift(@insns));
1130 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1131 eval(shift(@insns));
1132 eval(shift(@insns));
1133 eval(shift(@insns));
1134 eval(shift(@insns));
1136 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU
1138 foreach (@insns) { eval; } # remaining instructions
1141 &je ($done_avx_label);
1143 &vmovdqa(@X[2],"64($K_XX_XX)"); # pbswap mask
1144 &vmovdqa($Kx,"-64($K_XX_XX)"); # K_00_19
1145 &vmovdqu(@X[-4&7],"0($inp)"); # load input
1146 &vmovdqu(@X[-3&7],"16($inp)");
1147 &vmovdqu(@X[-2&7],"32($inp)");
1148 &vmovdqu(@X[-1&7],"48($inp)");
1149 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1158 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1159 my ($a,$b,$c,$d,$e);
1161 eval(shift(@insns));
1162 eval(shift(@insns));
1163 &vpshufb(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1164 eval(shift(@insns));
1165 eval(shift(@insns));
1166 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],$Kx);
1167 eval(shift(@insns));
1168 eval(shift(@insns));
1169 eval(shift(@insns));
1170 eval(shift(@insns));
1171 &vmovdqa(eval(16*$Xi)."(%rsp)",@X[$Xi&7]); # X[]+K xfer to IALU
1172 eval(shift(@insns));
1173 eval(shift(@insns));
1175 foreach (@insns) { eval; }
1182 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1183 my ($a,$b,$c,$d,$e);
1185 foreach (@insns) { eval; }
1192 &Xupdate_avx_16_31(\&body_00_19);
1193 &Xupdate_avx_16_31(\&body_00_19);
1194 &Xupdate_avx_16_31(\&body_00_19);
1195 &Xupdate_avx_16_31(\&body_00_19);
1196 &Xupdate_avx_32_79(\&body_00_19);
1197 &Xupdate_avx_32_79(\&body_20_39);
1198 &Xupdate_avx_32_79(\&body_20_39);
1199 &Xupdate_avx_32_79(\&body_20_39);
1200 &Xupdate_avx_32_79(\&body_20_39);
1201 &Xupdate_avx_32_79(\&body_20_39);
1202 &Xupdate_avx_32_79(\&body_40_59);
1203 &Xupdate_avx_32_79(\&body_40_59);
1204 &Xupdate_avx_32_79(\&body_40_59);
1205 &Xupdate_avx_32_79(\&body_40_59);
1206 &Xupdate_avx_32_79(\&body_40_59);
1207 &Xupdate_avx_32_79(\&body_20_39);
1208 &Xuplast_avx_80(\&body_20_39); # can jump to "done"
1210 $saved_j=$j; @saved_V=@V;
1212 &Xloop_avx(\&body_20_39);
1213 &Xloop_avx(\&body_20_39);
1214 &Xloop_avx(\&body_20_39);
1217 add 0($ctx),$A # update context
1224 mov @T[0],$B # magic seed
1236 $j=$saved_j; @V=@saved_V;
1238 &Xtail_avx(\&body_20_39);
1239 &Xtail_avx(\&body_20_39);
1240 &Xtail_avx(\&body_20_39);
1245 add 0($ctx),$A # update context
1256 $code.=<<___ if ($win64);
1257 movaps -40-6*16(%r14),%xmm6
1258 movaps -40-5*16(%r14),%xmm7
1259 movaps -40-4*16(%r14),%xmm8
1260 movaps -40-3*16(%r14),%xmm9
1261 movaps -40-2*16(%r14),%xmm10
1262 movaps -40-1*16(%r14),%xmm11
1274 .size sha1_block_data_order_avx,.-sha1_block_data_order_avx
1279 $Xi=4; # reset variables
1280 @X=map("%ymm$_",(4..7,0..3));
1281 @Tx=map("%ymm$_",(8..10));
1285 my @ROTX=("%eax","%ebp","%ebx","%ecx","%edx","%esi");
1286 my ($a5,$t0)=("%r12d","%edi");
1288 my ($A,$F,$B,$C,$D,$E)=@ROTX;
1293 .type sha1_block_data_order_avx2,\@function,3
1295 sha1_block_data_order_avx2:
1305 $code.=<<___ if ($win64);
1306 lea -6*16(%rsp),%rsp
1307 vmovaps %xmm6,-40-6*16(%rax)
1308 vmovaps %xmm7,-40-5*16(%rax)
1309 vmovaps %xmm8,-40-4*16(%rax)
1310 vmovaps %xmm9,-40-3*16(%rax)
1311 vmovaps %xmm10,-40-2*16(%rax)
1312 vmovaps %xmm11,-40-1*16(%rax)
1316 mov %rax,%r14 # original %rsp
1317 mov %rdi,$ctx # reassigned argument
1318 mov %rsi,$inp # reassigned argument
1319 mov %rdx,$num # reassigned argument
1326 lea K_XX_XX+64(%rip),$K_XX_XX
1328 mov 0($ctx),$A # load context
1330 cmovae $inp,$frame # next or same block
1335 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask
1337 vmovdqu ($inp),%xmm0
1338 vmovdqu 16($inp),%xmm1
1339 vmovdqu 32($inp),%xmm2
1340 vmovdqu 48($inp),%xmm3
1342 vinserti128 \$1,($frame),@X[-4&7],@X[-4&7]
1343 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7]
1344 vpshufb @X[2],@X[-4&7],@X[-4&7]
1345 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7]
1346 vpshufb @X[2],@X[-3&7],@X[-3&7]
1347 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7]
1348 vpshufb @X[2],@X[-2&7],@X[-2&7]
1349 vmovdqu -64($K_XX_XX),$Kx # K_00_19
1350 vpshufb @X[2],@X[-1&7],@X[-1&7]
1352 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19
1353 vpaddd $Kx,@X[-3&7],@X[1]
1354 vmovdqu @X[0],0(%rsp) # X[]+K xfer to IALU
1355 vpaddd $Kx,@X[-2&7],@X[2]
1356 vmovdqu @X[1],32(%rsp)
1357 vpaddd $Kx,@X[-1&7],@X[3]
1358 vmovdqu @X[2],64(%rsp)
1359 vmovdqu @X[3],96(%rsp)
1361 for (;$Xi<8;$Xi++) { # Xupdate_avx2_16_31
1364 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1365 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1366 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1367 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1368 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1369 &vpsrld (@Tx[0],@X[0],31);
1370 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1371 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1372 &vpaddd (@X[0],@X[0],@X[0]);
1373 &vpsrld (@Tx[1],@Tx[2],30);
1374 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1375 &vpslld (@Tx[2],@Tx[2],2);
1376 &vpxor (@X[0],@X[0],@Tx[1]);
1377 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1378 &vpaddd (@Tx[1],@X[0],$Kx);
1379 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1381 push(@X,shift(@X)); # "rotate" X[]
1384 lea 128(%rsp),$frame
1393 sub bodyx_00_19 () { # 8 instructions, 3 cycles critical path
1394 # at start $f=(b&c)^(~b&d), $b>>>=2
1395 return &bodyx_20_39() if ($rx==19); $rx++;
1397 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1399 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1400 '&lea ($frame,"256($frame)") if ($j%32==31);',
1401 '&andn ($t0,$a,$c)', # ~b&d for next round
1403 '&add ($e,$f)', # e+=(b&c)^(~b&d)
1404 '&rorx ($a5,$a,27)', # a<<<5
1405 '&rorx ($f,$a,2)', # b>>>2 for next round
1406 '&and ($a,$b)', # b&c for next round
1408 '&add ($e,$a5)', # e+=a<<<5
1409 '&xor ($a,$t0);'. # f=(b&c)^(~b&d) for next round
1411 'unshift(@ROTX,pop(@ROTX)); $j++;'
1415 sub bodyx_20_39 () { # 7 instructions, 2 cycles critical path
1416 # on entry $f=b^c^d, $b>>>=2
1417 return &bodyx_40_59() if ($rx==39); $rx++;
1419 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1421 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1422 '&lea ($frame,"256($frame)") if ($j%32==31);',
1424 '&lea ($e,"($e,$f)")', # e+=b^c^d
1425 '&rorx ($a5,$a,27)', # a<<<5
1426 '&rorx ($f,$a,2) if ($j<79)', # b>>>2 in next round
1427 '&xor ($a,$b) if ($j<79)', # b^c for next round
1429 '&add ($e,$a5)', # e+=a<<<5
1430 '&xor ($a,$c) if ($j<79);'. # f=b^c^d for next round
1432 'unshift(@ROTX,pop(@ROTX)); $j++;'
1436 sub bodyx_40_59 () { # 10 instructions, 3 cycles critical path
1437 # on entry $f=((b^c)&(c^d)), $b>>>=2
1440 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1442 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1443 '&lea ($frame,"256($frame)") if ($j%32==31);',
1444 '&xor ($f,$c) if ($j>39)', # (b^c)&(c^d)^c
1445 '&mov ($t0,$b) if ($j<59)', # count on zero latency
1446 '&xor ($t0,$c) if ($j<59)', # c^d for next round
1448 '&lea ($e,"($e,$f)")', # e+=(b^c)&(c^d)^c
1449 '&rorx ($a5,$a,27)', # a<<<5
1450 '&rorx ($f,$a,2)', # b>>>2 in next round
1451 '&xor ($a,$b)', # b^c for next round
1453 '&add ($e,$a5)', # e+=a<<<5
1454 '&and ($a,$t0) if ($j< 59);'. # f=(b^c)&(c^d) for next round
1455 '&xor ($a,$c) if ($j==59);'. # f=b^c^d for next round
1457 'unshift(@ROTX,pop(@ROTX)); $j++;'
1461 sub Xupdate_avx2_16_31() # recall that $Xi starts wtih 4
1464 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 instructions
1465 my ($a,$b,$c,$d,$e);
1467 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1468 eval(shift(@insns));
1469 eval(shift(@insns));
1470 eval(shift(@insns));
1471 eval(shift(@insns));
1473 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1474 eval(shift(@insns));
1475 eval(shift(@insns));
1476 eval(shift(@insns));
1478 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1479 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1480 eval(shift(@insns));
1481 eval(shift(@insns));
1483 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1484 eval(shift(@insns));
1485 eval(shift(@insns));
1486 eval(shift(@insns));
1487 eval(shift(@insns));
1489 &vpsrld (@Tx[0],@X[0],31);
1490 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1491 eval(shift(@insns));
1492 eval(shift(@insns));
1493 eval(shift(@insns));
1495 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1496 &vpaddd (@X[0],@X[0],@X[0]);
1497 eval(shift(@insns));
1498 eval(shift(@insns));
1500 &vpsrld (@Tx[1],@Tx[2],30);
1501 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1502 eval(shift(@insns));
1503 eval(shift(@insns));
1505 &vpslld (@Tx[2],@Tx[2],2);
1506 &vpxor (@X[0],@X[0],@Tx[1]);
1507 eval(shift(@insns));
1508 eval(shift(@insns));
1510 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1511 eval(shift(@insns));
1512 eval(shift(@insns));
1513 eval(shift(@insns));
1515 &vpaddd (@Tx[1],@X[0],$Kx);
1516 eval(shift(@insns));
1517 eval(shift(@insns));
1518 eval(shift(@insns));
1519 &vmovdqu(eval(32*($Xi))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1521 foreach (@insns) { eval; } # remaining instructions [if any]
1524 push(@X,shift(@X)); # "rotate" X[]
1527 sub Xupdate_avx2_32_79()
1530 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 to 50 instructions
1531 my ($a,$b,$c,$d,$e);
1533 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1534 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1535 eval(shift(@insns));
1536 eval(shift(@insns));
1538 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1539 &vmovdqu($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0);
1540 eval(shift(@insns));
1541 eval(shift(@insns));
1542 eval(shift(@insns));
1544 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]"
1545 eval(shift(@insns));
1546 eval(shift(@insns));
1547 eval(shift(@insns));
1549 &vpsrld (@Tx[0],@X[0],30);
1550 &vpslld (@X[0],@X[0],2);
1551 eval(shift(@insns));
1552 eval(shift(@insns));
1553 eval(shift(@insns));
1555 #&vpslld (@X[0],@X[0],2);
1556 eval(shift(@insns));
1557 eval(shift(@insns));
1558 eval(shift(@insns));
1560 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2
1561 eval(shift(@insns));
1562 eval(shift(@insns));
1563 eval(shift(@insns));
1564 eval(shift(@insns));
1566 &vpaddd (@Tx[1],@X[0],$Kx);
1567 eval(shift(@insns));
1568 eval(shift(@insns));
1569 eval(shift(@insns));
1570 eval(shift(@insns));
1572 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1574 foreach (@insns) { eval; } # remaining instructions
1577 push(@X,shift(@X)); # "rotate" X[]
1583 my @insns = (&$body,&$body,&$body,&$body,&$body); # 32 instructions
1584 my ($a,$b,$c,$d,$e);
1586 foreach (@insns) { eval; }
1590 &Xupdate_avx2_32_79(\&bodyx_00_19);
1591 &Xupdate_avx2_32_79(\&bodyx_00_19);
1592 &Xupdate_avx2_32_79(\&bodyx_00_19);
1593 &Xupdate_avx2_32_79(\&bodyx_00_19);
1595 &Xupdate_avx2_32_79(\&bodyx_20_39);
1596 &Xupdate_avx2_32_79(\&bodyx_20_39);
1597 &Xupdate_avx2_32_79(\&bodyx_20_39);
1598 &Xupdate_avx2_32_79(\&bodyx_20_39);
1601 &Xupdate_avx2_32_79(\&bodyx_40_59);
1602 &Xupdate_avx2_32_79(\&bodyx_40_59);
1603 &Xupdate_avx2_32_79(\&bodyx_40_59);
1604 &Xupdate_avx2_32_79(\&bodyx_40_59);
1606 &Xloop_avx2(\&bodyx_20_39);
1607 &Xloop_avx2(\&bodyx_20_39);
1608 &Xloop_avx2(\&bodyx_20_39);
1609 &Xloop_avx2(\&bodyx_20_39);
1612 lea 128($inp),$frame
1613 lea 128($inp),%rdi # borrow $t0
1615 cmovae $inp,$frame # next or previous block
1617 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c
1618 add 0($ctx),@ROTX[0] # update context
1619 add 4($ctx),@ROTX[1]
1620 add 8($ctx),@ROTX[3]
1621 mov @ROTX[0],0($ctx)
1622 add 12($ctx),@ROTX[4]
1623 mov @ROTX[1],4($ctx)
1624 mov @ROTX[0],$A # A=d
1625 add 16($ctx),@ROTX[5]
1627 mov @ROTX[3],8($ctx)
1628 mov @ROTX[4],$D # D=b
1629 #xchg @ROTX[5],$F # F=c, C=f
1630 mov @ROTX[4],12($ctx)
1631 mov @ROTX[1],$F # F=e
1632 mov @ROTX[5],16($ctx)
1634 mov @ROTX[5],$E # E=c
1636 #xchg $F,$E # E=c, F=e
1642 $Xi=4; # reset variables
1643 @X=map("%ymm$_",(4..7,0..3));
1646 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask
1647 cmp $num,%rdi # borrowed $t0
1650 vmovdqu -64(%rdi),%xmm0 # low part of @X[-4&7]
1651 vmovdqu -48(%rdi),%xmm1
1652 vmovdqu -32(%rdi),%xmm2
1653 vmovdqu -16(%rdi),%xmm3
1654 vinserti128 \$1,0($frame),@X[-4&7],@X[-4&7]
1655 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7]
1656 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7]
1657 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7]
1662 lea 128+16(%rsp),$frame
1669 $rx=$j=0; @ROTX=($A,$F,$B,$C,$D,$E);
1671 &Xloop_avx2 (\&bodyx_00_19);
1672 &Xloop_avx2 (\&bodyx_00_19);
1673 &Xloop_avx2 (\&bodyx_00_19);
1674 &Xloop_avx2 (\&bodyx_00_19);
1676 &Xloop_avx2 (\&bodyx_20_39);
1677 &vmovdqu ($Kx,"-64($K_XX_XX)"); # K_00_19
1678 &vpshufb (@X[-4&7],@X[-4&7],@X[2]); # byte swap
1679 &Xloop_avx2 (\&bodyx_20_39);
1680 &vpshufb (@X[-3&7],@X[-3&7],@X[2]);
1681 &vpaddd (@Tx[0],@X[-4&7],$Kx); # add K_00_19
1682 &Xloop_avx2 (\&bodyx_20_39);
1683 &vmovdqu ("0(%rsp)",@Tx[0]);
1684 &vpshufb (@X[-2&7],@X[-2&7],@X[2]);
1685 &vpaddd (@Tx[1],@X[-3&7],$Kx);
1686 &Xloop_avx2 (\&bodyx_20_39);
1687 &vmovdqu ("32(%rsp)",@Tx[1]);
1688 &vpshufb (@X[-1&7],@X[-1&7],@X[2]);
1689 &vpaddd (@X[2],@X[-2&7],$Kx);
1691 &Xloop_avx2 (\&bodyx_40_59);
1693 &vmovdqu ("64(%rsp)",@X[2]);
1694 &vpaddd (@X[3],@X[-1&7],$Kx);
1695 &Xloop_avx2 (\&bodyx_40_59);
1696 &vmovdqu ("96(%rsp)",@X[3]);
1697 &Xloop_avx2 (\&bodyx_40_59);
1698 &Xupdate_avx2_16_31(\&bodyx_40_59);
1700 &Xupdate_avx2_16_31(\&bodyx_20_39);
1701 &Xupdate_avx2_16_31(\&bodyx_20_39);
1702 &Xupdate_avx2_16_31(\&bodyx_20_39);
1703 &Xloop_avx2 (\&bodyx_20_39);
1706 lea 128(%rsp),$frame
1708 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c
1709 add 0($ctx),@ROTX[0] # update context
1710 add 4($ctx),@ROTX[1]
1711 add 8($ctx),@ROTX[3]
1712 mov @ROTX[0],0($ctx)
1713 add 12($ctx),@ROTX[4]
1714 mov @ROTX[1],4($ctx)
1715 mov @ROTX[0],$A # A=d
1716 add 16($ctx),@ROTX[5]
1718 mov @ROTX[3],8($ctx)
1719 mov @ROTX[4],$D # D=b
1720 #xchg @ROTX[5],$F # F=c, C=f
1721 mov @ROTX[4],12($ctx)
1722 mov @ROTX[1],$F # F=e
1723 mov @ROTX[5],16($ctx)
1725 mov @ROTX[5],$E # E=c
1727 #xchg $F,$E # E=c, F=e
1735 $code.=<<___ if ($win64);
1736 movaps -40-6*16(%r14),%xmm6
1737 movaps -40-5*16(%r14),%xmm7
1738 movaps -40-4*16(%r14),%xmm8
1739 movaps -40-3*16(%r14),%xmm9
1740 movaps -40-2*16(%r14),%xmm10
1741 movaps -40-1*16(%r14),%xmm11
1753 .size sha1_block_data_order_avx2,.-sha1_block_data_order_avx2
1760 .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19
1761 .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19
1762 .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39
1763 .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39
1764 .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59
1765 .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59
1766 .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79
1767 .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79
1768 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask
1769 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask
1770 .byte 0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0
1774 .asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1778 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1779 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
1787 .extern __imp_RtlVirtualUnwind
1788 .type se_handler,\@abi-omnipotent
1802 mov 120($context),%rax # pull context->Rax
1803 mov 248($context),%rbx # pull context->Rip
1805 lea .Lprologue(%rip),%r10
1806 cmp %r10,%rbx # context->Rip<.Lprologue
1807 jb .Lcommon_seh_tail
1809 mov 152($context),%rax # pull context->Rsp
1811 lea .Lepilogue(%rip),%r10
1812 cmp %r10,%rbx # context->Rip>=.Lepilogue
1813 jae .Lcommon_seh_tail
1815 mov `16*4`(%rax),%rax # pull saved stack pointer
1822 mov %rbx,144($context) # restore context->Rbx
1823 mov %rbp,160($context) # restore context->Rbp
1824 mov %r12,216($context) # restore context->R12
1825 mov %r13,224($context) # restore context->R13
1826 mov %r14,232($context) # restore context->R14
1828 jmp .Lcommon_seh_tail
1829 .size se_handler,.-se_handler
1831 .type shaext_handler,\@abi-omnipotent
1845 mov 120($context),%rax # pull context->Rax
1846 mov 248($context),%rbx # pull context->Rip
1848 lea .Lprologue_shaext(%rip),%r10
1849 cmp %r10,%rbx # context->Rip<.Lprologue
1850 jb .Lcommon_seh_tail
1852 lea .Lepilogue_shaext(%rip),%r10
1853 cmp %r10,%rbx # context->Rip>=.Lepilogue
1854 jae .Lcommon_seh_tail
1856 lea -8-4*16(%rax),%rsi
1857 lea 512($context),%rdi # &context.Xmm6
1859 .long 0xa548f3fc # cld; rep movsq
1861 jmp .Lcommon_seh_tail
1862 .size shaext_handler,.-shaext_handler
1864 .type ssse3_handler,\@abi-omnipotent
1878 mov 120($context),%rax # pull context->Rax
1879 mov 248($context),%rbx # pull context->Rip
1881 mov 8($disp),%rsi # disp->ImageBase
1882 mov 56($disp),%r11 # disp->HandlerData
1884 mov 0(%r11),%r10d # HandlerData[0]
1885 lea (%rsi,%r10),%r10 # prologue label
1886 cmp %r10,%rbx # context->Rip<prologue label
1887 jb .Lcommon_seh_tail
1889 mov 152($context),%rax # pull context->Rsp
1891 mov 4(%r11),%r10d # HandlerData[1]
1892 lea (%rsi,%r10),%r10 # epilogue label
1893 cmp %r10,%rbx # context->Rip>=epilogue label
1894 jae .Lcommon_seh_tail
1896 mov 232($context),%rax # pull context->R14
1898 lea -40-6*16(%rax),%rsi
1899 lea 512($context),%rdi # &context.Xmm6
1901 .long 0xa548f3fc # cld; rep movsq
1908 mov %rbx,144($context) # restore context->Rbx
1909 mov %rbp,160($context) # restore context->Rbp
1910 mov %r12,216($context) # restore cotnext->R12
1911 mov %r13,224($context) # restore cotnext->R13
1912 mov %r14,232($context) # restore cotnext->R14
1917 mov %rax,152($context) # restore context->Rsp
1918 mov %rsi,168($context) # restore context->Rsi
1919 mov %rdi,176($context) # restore context->Rdi
1921 mov 40($disp),%rdi # disp->ContextRecord
1922 mov $context,%rsi # context
1923 mov \$154,%ecx # sizeof(CONTEXT)
1924 .long 0xa548f3fc # cld; rep movsq
1927 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
1928 mov 8(%rsi),%rdx # arg2, disp->ImageBase
1929 mov 0(%rsi),%r8 # arg3, disp->ControlPc
1930 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
1931 mov 40(%rsi),%r10 # disp->ContextRecord
1932 lea 56(%rsi),%r11 # &disp->HandlerData
1933 lea 24(%rsi),%r12 # &disp->EstablisherFrame
1934 mov %r10,32(%rsp) # arg5
1935 mov %r11,40(%rsp) # arg6
1936 mov %r12,48(%rsp) # arg7
1937 mov %rcx,56(%rsp) # arg8, (NULL)
1938 call *__imp_RtlVirtualUnwind(%rip)
1940 mov \$1,%eax # ExceptionContinueSearch
1952 .size ssse3_handler,.-ssse3_handler
1956 .rva .LSEH_begin_sha1_block_data_order
1957 .rva .LSEH_end_sha1_block_data_order
1958 .rva .LSEH_info_sha1_block_data_order
1959 .rva .LSEH_begin_sha1_block_data_order_shaext
1960 .rva .LSEH_end_sha1_block_data_order_shaext
1961 .rva .LSEH_info_sha1_block_data_order_shaext
1962 .rva .LSEH_begin_sha1_block_data_order_ssse3
1963 .rva .LSEH_end_sha1_block_data_order_ssse3
1964 .rva .LSEH_info_sha1_block_data_order_ssse3
1966 $code.=<<___ if ($avx);
1967 .rva .LSEH_begin_sha1_block_data_order_avx
1968 .rva .LSEH_end_sha1_block_data_order_avx
1969 .rva .LSEH_info_sha1_block_data_order_avx
1971 $code.=<<___ if ($avx>1);
1972 .rva .LSEH_begin_sha1_block_data_order_avx2
1973 .rva .LSEH_end_sha1_block_data_order_avx2
1974 .rva .LSEH_info_sha1_block_data_order_avx2
1979 .LSEH_info_sha1_block_data_order:
1982 .LSEH_info_sha1_block_data_order_shaext:
1985 .LSEH_info_sha1_block_data_order_ssse3:
1988 .rva .Lprologue_ssse3,.Lepilogue_ssse3 # HandlerData[]
1990 $code.=<<___ if ($avx);
1991 .LSEH_info_sha1_block_data_order_avx:
1994 .rva .Lprologue_avx,.Lepilogue_avx # HandlerData[]
1996 $code.=<<___ if ($avx>1);
1997 .LSEH_info_sha1_block_data_order_avx2:
2000 .rva .Lprologue_avx2,.Lepilogue_avx2 # HandlerData[]
2004 ####################################################################
2007 if (@_[0] =~ /\$([x0-9a-f]+),\s*%xmm([0-7]),\s*%xmm([0-7])/) {
2008 my @opcode=(0x0f,0x3a,0xcc);
2009 push @opcode,0xc0|($2&7)|(($3&7)<<3); # ModR/M
2011 push @opcode,$c=~/^0/?oct($c):$c;
2012 return ".byte\t".join(',',@opcode);
2014 return "sha1rnds4\t".@_[0];
2021 "sha1nexte" => 0xc8,
2023 "sha1msg2" => 0xca );
2025 if (defined($opcodelet{$instr}) && @_[0] =~ /%xmm([0-7]),\s*%xmm([0-7])/) {
2026 my @opcode=(0x0f,0x38);
2027 push @opcode,$opcodelet{$instr};
2028 push @opcode,0xc0|($1&7)|(($2&7)<<3); # ModR/M
2029 return ".byte\t".join(',',@opcode);
2031 return $instr."\t".@_[0];
2035 foreach (split("\n",$code)) {
2036 s/\`([^\`]*)\`/eval $1/geo;
2038 s/\b(sha1rnds4)\s+(.*)/sha1rnds4($2)/geo or
2039 s/\b(sha1[^\s]*)\s+(.*)/sha1op38($1,$2)/geo;