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15 # assembler, which originally made it hard to motivate the effort.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 $shaext=1; ### set to zero if compiling for 1.0.1
111 $avx=1 if (!$shaext && $avx);
113 open OUT,"| \"$^X\" $xlate $flavour $output";
116 $ctx="%rdi"; # 1st arg
117 $inp="%rsi"; # 2nd arg
118 $num="%rdx"; # 3rd arg
120 # reassign arguments in order to produce more compact code
128 @xi=("%edx","%ebp","%r14d");
138 my ($i,$a,$b,$c,$d,$e)=@_;
140 $code.=<<___ if ($i==0);
141 mov `4*$i`($inp),$xi[0]
144 $code.=<<___ if ($i<15);
145 mov `4*$j`($inp),$xi[1]
147 mov $xi[0],`4*$i`(%rsp)
153 lea 0x5a827999($xi[0],$e),$e
159 $code.=<<___ if ($i>=15);
160 xor `4*($j%16)`(%rsp),$xi[1]
162 mov $xi[0],`4*($i%16)`(%rsp)
164 xor `4*(($j+2)%16)`(%rsp),$xi[1]
167 xor `4*(($j+8)%16)`(%rsp),$xi[1]
169 lea 0x5a827999($xi[0],$e),$e
176 push(@xi,shift(@xi));
180 my ($i,$a,$b,$c,$d,$e)=@_;
182 my $K=($i<40)?0x6ed9eba1:0xca62c1d6;
183 $code.=<<___ if ($i<79);
184 xor `4*($j%16)`(%rsp),$xi[1]
186 `"mov $xi[0],".4*($i%16)."(%rsp)" if ($i<72)`
188 xor `4*(($j+2)%16)`(%rsp),$xi[1]
191 xor `4*(($j+8)%16)`(%rsp),$xi[1]
199 $code.=<<___ if ($i==79);
210 push(@xi,shift(@xi));
214 my ($i,$a,$b,$c,$d,$e)=@_;
217 xor `4*($j%16)`(%rsp),$xi[1]
219 mov $xi[0],`4*($i%16)`(%rsp)
221 xor `4*(($j+2)%16)`(%rsp),$xi[1]
224 xor `4*(($j+8)%16)`(%rsp),$xi[1]
225 lea 0x8f1bbcdc($xi[0],$e),$e
235 push(@xi,shift(@xi));
240 .extern OPENSSL_ia32cap_P
242 .globl sha1_block_data_order
243 .type sha1_block_data_order,\@function,3
245 sha1_block_data_order:
246 mov OPENSSL_ia32cap_P+0(%rip),%r9d
247 mov OPENSSL_ia32cap_P+4(%rip),%r8d
248 mov OPENSSL_ia32cap_P+8(%rip),%r10d
249 test \$`1<<9`,%r8d # check SSSE3 bit
252 $code.=<<___ if ($shaext);
253 test \$`1<<29`,%r10d # check SHA bit
256 $code.=<<___ if ($avx>1);
257 and \$`1<<3|1<<5|1<<8`,%r10d # check AVX2+BMI1+BMI2
258 cmp \$`1<<3|1<<5|1<<8`,%r10d
261 $code.=<<___ if ($avx);
262 and \$`1<<28`,%r8d # mask AVX bit
263 and \$`1<<30`,%r9d # mask "Intel CPU" bit
265 cmp \$`1<<28|1<<30`,%r8d
279 mov %rdi,$ctx # reassigned argument
281 mov %rsi,$inp # reassigned argument
283 mov %rdx,$num # reassigned argument
284 mov %rax,`16*4`(%rsp)
297 for($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); }
298 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
299 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
300 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
314 lea `16*4`($inp),$inp
317 mov `16*4`(%rsp),%rsi
326 .size sha1_block_data_order,.-sha1_block_data_order
329 ######################################################################
330 # Intel SHA Extensions implementation of SHA1 update function.
332 my ($ctx,$inp,$num)=("%rdi","%rsi","%rdx");
333 my ($ABCD,$E,$E_,$BSWAP,$ABCD_SAVE,$E_SAVE)=map("%xmm$_",(0..3,8,9));
334 my @MSG=map("%xmm$_",(4..7));
337 .type sha1_block_data_order_shaext,\@function,3
339 sha1_block_data_order_shaext:
342 $code.=<<___ if ($win64);
343 lea `-8-4*16`(%rsp),%rsp
344 movaps %xmm6,-8-4*16(%rax)
345 movaps %xmm7,-8-3*16(%rax)
346 movaps %xmm8,-8-2*16(%rax)
347 movaps %xmm9,-8-1*16(%rax)
353 movdqa K_XX_XX+0xa0(%rip),$BSWAP # byte-n-word swap
355 movdqu ($inp),@MSG[0]
356 pshufd \$0b00011011,$ABCD,$ABCD # flip word order
357 movdqu 0x10($inp),@MSG[1]
358 pshufd \$0b00011011,$E,$E # flip word order
359 movdqu 0x20($inp),@MSG[2]
360 pshufb $BSWAP,@MSG[0]
361 movdqu 0x30($inp),@MSG[3]
362 pshufb $BSWAP,@MSG[1]
363 pshufb $BSWAP,@MSG[2]
364 movdqa $E,$E_SAVE # offload $E
365 pshufb $BSWAP,@MSG[3]
371 lea 0x40($inp),%rax # next input block
374 movdqa $ABCD,$ABCD_SAVE # offload $ABCD
376 for($i=0;$i<20-4;$i+=2) {
378 sha1msg1 @MSG[1],@MSG[0]
380 sha1rnds4 \$`int($i/5)`,$E,$ABCD # 0-3...
381 sha1nexte @MSG[1],$E_
383 sha1msg1 @MSG[2],@MSG[1]
384 sha1msg2 @MSG[3],@MSG[0]
387 sha1rnds4 \$`int(($i+1)/5)`,$E_,$ABCD
390 sha1msg2 @MSG[0],@MSG[1]
392 push(@MSG,shift(@MSG)); push(@MSG,shift(@MSG));
395 movdqu ($inp),@MSG[0]
397 sha1rnds4 \$3,$E,$ABCD # 64-67
398 sha1nexte @MSG[1],$E_
399 movdqu 0x10($inp),@MSG[1]
400 pshufb $BSWAP,@MSG[0]
403 sha1rnds4 \$3,$E_,$ABCD # 68-71
405 movdqu 0x20($inp),@MSG[2]
406 pshufb $BSWAP,@MSG[1]
409 sha1rnds4 \$3,$E,$ABCD # 72-75
410 sha1nexte @MSG[3],$E_
411 movdqu 0x30($inp),@MSG[3]
412 pshufb $BSWAP,@MSG[2]
415 sha1rnds4 \$3,$E_,$ABCD # 76-79
417 pshufb $BSWAP,@MSG[3]
419 paddd $ABCD_SAVE,$ABCD
420 movdqa $E,$E_SAVE # offload $E
424 pshufd \$0b00011011,$ABCD,$ABCD
425 pshufd \$0b00011011,$E,$E
429 $code.=<<___ if ($win64);
430 movaps -8-4*16(%rax),%xmm6
431 movaps -8-3*16(%rax),%xmm7
432 movaps -8-2*16(%rax),%xmm8
433 movaps -8-1*16(%rax),%xmm9
439 .size sha1_block_data_order_shaext,.-sha1_block_data_order_shaext
444 my @X=map("%xmm$_",(4..7,0..3));
445 my @Tx=map("%xmm$_",(8..10));
447 my @V=($A,$B,$C,$D,$E)=("%eax","%ebx","%ecx","%edx","%ebp"); # size optimization
448 my @T=("%esi","%edi");
453 my $_rol=sub { &rol(@_) };
454 my $_ror=sub { &ror(@_) };
460 jmp .Lalign32_$sn # see "Decoded ICache" in manual
468 .type sha1_block_data_order_ssse3,\@function,3
470 sha1_block_data_order_ssse3:
476 push %r13 # redundant, done to share Win64 SE handler
478 lea `-64-($win64?6*16:0)`(%rsp),%rsp
480 $code.=<<___ if ($win64);
481 movaps %xmm6,-40-6*16(%rax)
482 movaps %xmm7,-40-5*16(%rax)
483 movaps %xmm8,-40-4*16(%rax)
484 movaps %xmm9,-40-3*16(%rax)
485 movaps %xmm10,-40-2*16(%rax)
486 movaps %xmm11,-40-1*16(%rax)
490 mov %rax,%r14 # original %rsp
492 mov %rdi,$ctx # reassigned argument
493 mov %rsi,$inp # reassigned argument
494 mov %rdx,$num # reassigned argument
498 lea K_XX_XX+64(%rip),$K_XX_XX
500 mov 0($ctx),$A # load context
504 mov $B,@T[0] # magic seed
510 movdqa 64($K_XX_XX),@X[2] # pbswap mask
511 movdqa -64($K_XX_XX),@Tx[1] # K_00_19
512 movdqu 0($inp),@X[-4&7] # load input to %xmm[0-3]
513 movdqu 16($inp),@X[-3&7]
514 movdqu 32($inp),@X[-2&7]
515 movdqu 48($inp),@X[-1&7]
516 pshufb @X[2],@X[-4&7] # byte swap
517 pshufb @X[2],@X[-3&7]
518 pshufb @X[2],@X[-2&7]
520 paddd @Tx[1],@X[-4&7] # add K_00_19
521 pshufb @X[2],@X[-1&7]
522 paddd @Tx[1],@X[-3&7]
523 paddd @Tx[1],@X[-2&7]
524 movdqa @X[-4&7],0(%rsp) # X[]+K xfer to IALU
525 psubd @Tx[1],@X[-4&7] # restore X[]
526 movdqa @X[-3&7],16(%rsp)
527 psubd @Tx[1],@X[-3&7]
528 movdqa @X[-2&7],32(%rsp)
529 psubd @Tx[1],@X[-2&7]
533 sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm
534 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
536 $arg = "\$$arg" if ($arg*1 eq $arg);
537 $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
540 sub Xupdate_ssse3_16_31() # recall that $Xi starts wtih 4
543 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
546 eval(shift(@insns)); # ror
547 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]);
549 &movdqa (@Tx[0],@X[-1&7]);
550 &paddd (@Tx[1],@X[-1&7]);
554 &punpcklqdq(@X[0],@X[-3&7]); # compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8);
556 eval(shift(@insns)); # rol
558 &psrldq (@Tx[0],4); # "X[-3]", 3 dwords
562 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
564 eval(shift(@insns)); # ror
565 &pxor (@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
570 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
572 eval(shift(@insns)); # rol
573 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
577 &movdqa (@Tx[2],@X[0]);
580 eval(shift(@insns)); # ror
581 &movdqa (@Tx[0],@X[0]);
584 &pslldq (@Tx[2],12); # "X[0]"<<96, extract one dword
585 &paddd (@X[0],@X[0]);
591 eval(shift(@insns)); # rol
593 &movdqa (@Tx[1],@Tx[2]);
599 eval(shift(@insns)); # ror
600 &por (@X[0],@Tx[0]); # "X[0]"<<<=1
606 &pxor (@X[0],@Tx[2]);
608 &movdqa (@Tx[2],eval(2*16*(($Xi)/5)-64)."($K_XX_XX)"); # K_XX_XX
609 eval(shift(@insns)); # rol
613 &pxor (@X[0],@Tx[1]); # "X[0]"^=("X[0]">>96)<<<2
614 &pshufd (@Tx[1],@X[-1&7],0xee) if ($Xi==7); # was &movdqa (@Tx[0],@X[-1&7]) in Xupdate_ssse3_32_79
616 foreach (@insns) { eval; } # remaining instructions [if any]
618 $Xi++; push(@X,shift(@X)); # "rotate" X[]
619 push(@Tx,shift(@Tx));
622 sub Xupdate_ssse3_32_79()
625 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
628 eval(shift(@insns)) if ($Xi==8);
629 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
630 eval(shift(@insns)) if ($Xi==8);
631 eval(shift(@insns)); # body_20_39
633 eval(shift(@insns)) if (@insns[1] =~ /_ror/);
634 eval(shift(@insns)) if (@insns[0] =~ /_ror/);
635 &punpcklqdq(@Tx[0],@X[-1&7]); # compose "X[-6]", was &palignr(@Tx[0],@X[-2&7],8);
637 eval(shift(@insns)); # rol
639 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
643 &movdqa (@Tx[2],@Tx[1]);# "perpetuate" K_XX_XX...
644 } else { # ... or load next one
645 &movdqa (@Tx[2],eval(2*16*($Xi/5)-64)."($K_XX_XX)");
647 eval(shift(@insns)); # ror
648 &paddd (@Tx[1],@X[-1&7]);
651 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-6]"
652 eval(shift(@insns)); # body_20_39
655 eval(shift(@insns)); # rol
656 eval(shift(@insns)) if (@insns[0] =~ /_ror/);
658 &movdqa (@Tx[0],@X[0]);
661 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
662 eval(shift(@insns)); # ror
664 eval(shift(@insns)); # body_20_39
670 eval(shift(@insns)) if (@insns[0] =~ /_rol/);# rol
673 eval(shift(@insns)); # ror
675 &por (@X[0],@Tx[0]); # "X[0]"<<<=2
677 eval(shift(@insns)); # body_20_39
678 eval(shift(@insns)) if (@insns[1] =~ /_rol/);
679 eval(shift(@insns)) if (@insns[0] =~ /_rol/);
680 &pshufd(@Tx[1],@X[-1&7],0xee) if ($Xi<19); # was &movdqa (@Tx[1],@X[0])
682 eval(shift(@insns)); # rol
685 eval(shift(@insns)); # rol
688 foreach (@insns) { eval; } # remaining instructions
690 $Xi++; push(@X,shift(@X)); # "rotate" X[]
691 push(@Tx,shift(@Tx));
694 sub Xuplast_ssse3_80()
697 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
704 &paddd (@Tx[1],@X[-1&7]);
708 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU
710 foreach (@insns) { eval; } # remaining instructions
713 &je (".Ldone_ssse3");
715 unshift(@Tx,pop(@Tx));
717 &movdqa (@X[2],"64($K_XX_XX)"); # pbswap mask
718 &movdqa (@Tx[1],"-64($K_XX_XX)"); # K_00_19
719 &movdqu (@X[-4&7],"0($inp)"); # load input
720 &movdqu (@X[-3&7],"16($inp)");
721 &movdqu (@X[-2&7],"32($inp)");
722 &movdqu (@X[-1&7],"48($inp)");
723 &pshufb (@X[-4&7],@X[2]); # byte swap
732 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
738 &pshufb (@X[($Xi-3)&7],@X[2]);
743 &paddd (@X[($Xi-4)&7],@Tx[1]);
748 &movdqa (eval(16*$Xi)."(%rsp)",@X[($Xi-4)&7]); # X[]+K xfer to IALU
753 &psubd (@X[($Xi-4)&7],@Tx[1]);
755 foreach (@insns) { eval; }
762 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
765 foreach (@insns) { eval; }
768 sub body_00_19 () { # ((c^d)&b)^d
769 # on start @T[0]=(c^d)&b
770 return &body_20_39() if ($rx==19); $rx++;
772 '($a,$b,$c,$d,$e)=@V;'.
773 '&$_ror ($b,$j?7:2)', # $b>>>2
775 '&mov (@T[1],$a)', # $b for next round
777 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
778 '&xor ($b,$c)', # $c^$d for next round
782 '&and (@T[1],$b)', # ($b&($c^$d)) for next round
784 '&xor ($b,$c)', # restore $b
785 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
789 sub body_20_39 () { # b^d^c
791 return &body_40_59() if ($rx==39); $rx++;
793 '($a,$b,$c,$d,$e)=@V;'.
794 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
795 '&xor (@T[0],$d) if($j==19);'.
796 '&xor (@T[0],$c) if($j> 19)', # ($b^$d^$c)
797 '&mov (@T[1],$a)', # $b for next round
801 '&xor (@T[1],$c) if ($j< 79)', # $b^$d for next round
803 '&$_ror ($b,7)', # $b>>>2
804 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
808 sub body_40_59 () { # ((b^c)&(c^d))^c
809 # on entry @T[0]=(b^c), (c^=d)
812 '($a,$b,$c,$d,$e)=@V;'.
813 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer
814 '&and (@T[0],$c) if ($j>=40)', # (b^c)&(c^d)
815 '&xor ($c,$d) if ($j>=40)', # restore $c
817 '&$_ror ($b,7)', # $b>>>2
818 '&mov (@T[1],$a)', # $b for next round
823 '&xor (@T[1],$c) if ($j==59);'.
824 '&xor (@T[1],$b) if ($j< 59)', # b^c for next round
826 '&xor ($b,$c) if ($j< 59)', # c^d for next round
827 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
834 &Xupdate_ssse3_16_31(\&body_00_19);
835 &Xupdate_ssse3_16_31(\&body_00_19);
836 &Xupdate_ssse3_16_31(\&body_00_19);
837 &Xupdate_ssse3_16_31(\&body_00_19);
838 &Xupdate_ssse3_32_79(\&body_00_19);
839 &Xupdate_ssse3_32_79(\&body_20_39);
840 &Xupdate_ssse3_32_79(\&body_20_39);
841 &Xupdate_ssse3_32_79(\&body_20_39);
842 &Xupdate_ssse3_32_79(\&body_20_39);
843 &Xupdate_ssse3_32_79(\&body_20_39);
844 &Xupdate_ssse3_32_79(\&body_40_59);
845 &Xupdate_ssse3_32_79(\&body_40_59);
846 &Xupdate_ssse3_32_79(\&body_40_59);
847 &Xupdate_ssse3_32_79(\&body_40_59);
848 &Xupdate_ssse3_32_79(\&body_40_59);
849 &Xupdate_ssse3_32_79(\&body_20_39);
850 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done"
852 $saved_j=$j; @saved_V=@V;
854 &Xloop_ssse3(\&body_20_39);
855 &Xloop_ssse3(\&body_20_39);
856 &Xloop_ssse3(\&body_20_39);
859 add 0($ctx),$A # update context
866 mov @T[0],$B # magic seed
878 $j=$saved_j; @V=@saved_V;
880 &Xtail_ssse3(\&body_20_39);
881 &Xtail_ssse3(\&body_20_39);
882 &Xtail_ssse3(\&body_20_39);
885 add 0($ctx),$A # update context
896 $code.=<<___ if ($win64);
897 movaps -40-6*16(%r14),%xmm6
898 movaps -40-5*16(%r14),%xmm7
899 movaps -40-4*16(%r14),%xmm8
900 movaps -40-3*16(%r14),%xmm9
901 movaps -40-2*16(%r14),%xmm10
902 movaps -40-1*16(%r14),%xmm11
914 .size sha1_block_data_order_ssse3,.-sha1_block_data_order_ssse3
918 $Xi=4; # reset variables
919 @X=map("%xmm$_",(4..7,0..3));
920 @Tx=map("%xmm$_",(8..10));
924 my $done_avx_label=".Ldone_avx";
926 my $_rol=sub { &shld(@_[0],@_) };
927 my $_ror=sub { &shrd(@_[0],@_) };
930 .type sha1_block_data_order_avx,\@function,3
932 sha1_block_data_order_avx:
938 push %r13 # redundant, done to share Win64 SE handler
940 lea `-64-($win64?6*16:0)`(%rsp),%rsp
943 $code.=<<___ if ($win64);
944 vmovaps %xmm6,-40-6*16(%rax)
945 vmovaps %xmm7,-40-5*16(%rax)
946 vmovaps %xmm8,-40-4*16(%rax)
947 vmovaps %xmm9,-40-3*16(%rax)
948 vmovaps %xmm10,-40-2*16(%rax)
949 vmovaps %xmm11,-40-1*16(%rax)
953 mov %rax,%r14 # original %rsp
955 mov %rdi,$ctx # reassigned argument
956 mov %rsi,$inp # reassigned argument
957 mov %rdx,$num # reassigned argument
961 lea K_XX_XX+64(%rip),$K_XX_XX
963 mov 0($ctx),$A # load context
967 mov $B,@T[0] # magic seed
973 vmovdqa 64($K_XX_XX),@X[2] # pbswap mask
974 vmovdqa -64($K_XX_XX),$Kx # K_00_19
975 vmovdqu 0($inp),@X[-4&7] # load input to %xmm[0-3]
976 vmovdqu 16($inp),@X[-3&7]
977 vmovdqu 32($inp),@X[-2&7]
978 vmovdqu 48($inp),@X[-1&7]
979 vpshufb @X[2],@X[-4&7],@X[-4&7] # byte swap
981 vpshufb @X[2],@X[-3&7],@X[-3&7]
982 vpshufb @X[2],@X[-2&7],@X[-2&7]
983 vpshufb @X[2],@X[-1&7],@X[-1&7]
984 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19
985 vpaddd $Kx,@X[-3&7],@X[1]
986 vpaddd $Kx,@X[-2&7],@X[2]
987 vmovdqa @X[0],0(%rsp) # X[]+K xfer to IALU
988 vmovdqa @X[1],16(%rsp)
989 vmovdqa @X[2],32(%rsp)
993 sub Xupdate_avx_16_31() # recall that $Xi starts wtih 4
996 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions
1000 eval(shift(@insns));
1001 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1002 eval(shift(@insns));
1003 eval(shift(@insns));
1005 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1006 eval(shift(@insns));
1007 eval(shift(@insns));
1008 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1009 eval(shift(@insns));
1010 eval(shift(@insns));
1011 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1012 eval(shift(@insns));
1013 eval(shift(@insns));
1015 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1016 eval(shift(@insns));
1017 eval(shift(@insns));
1018 eval(shift(@insns));
1019 eval(shift(@insns));
1021 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1022 eval(shift(@insns));
1023 eval(shift(@insns));
1024 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1025 eval(shift(@insns));
1026 eval(shift(@insns));
1028 &vpsrld (@Tx[0],@X[0],31);
1029 eval(shift(@insns));
1030 eval(shift(@insns));
1031 eval(shift(@insns));
1032 eval(shift(@insns));
1034 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1035 &vpaddd (@X[0],@X[0],@X[0]);
1036 eval(shift(@insns));
1037 eval(shift(@insns));
1038 eval(shift(@insns));
1039 eval(shift(@insns));
1041 &vpsrld (@Tx[1],@Tx[2],30);
1042 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1043 eval(shift(@insns));
1044 eval(shift(@insns));
1045 eval(shift(@insns));
1046 eval(shift(@insns));
1048 &vpslld (@Tx[2],@Tx[2],2);
1049 &vpxor (@X[0],@X[0],@Tx[1]);
1050 eval(shift(@insns));
1051 eval(shift(@insns));
1052 eval(shift(@insns));
1053 eval(shift(@insns));
1055 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1056 eval(shift(@insns));
1057 eval(shift(@insns));
1058 &vmovdqa ($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1059 eval(shift(@insns));
1060 eval(shift(@insns));
1063 foreach (@insns) { eval; } # remaining instructions [if any]
1065 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1068 sub Xupdate_avx_32_79()
1071 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions
1072 my ($a,$b,$c,$d,$e);
1074 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1075 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1076 eval(shift(@insns)); # body_20_39
1077 eval(shift(@insns));
1078 eval(shift(@insns));
1079 eval(shift(@insns)); # rol
1081 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1082 eval(shift(@insns));
1083 eval(shift(@insns)) if (@insns[0] !~ /&ro[rl]/);
1084 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1085 &vmovdqa ($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0);
1086 eval(shift(@insns)); # ror
1087 eval(shift(@insns));
1089 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]"
1090 eval(shift(@insns)); # body_20_39
1091 eval(shift(@insns));
1092 eval(shift(@insns));
1093 eval(shift(@insns)); # rol
1095 &vpsrld (@Tx[0],@X[0],30);
1096 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1097 eval(shift(@insns));
1098 eval(shift(@insns));
1099 eval(shift(@insns)); # ror
1100 eval(shift(@insns));
1102 &vpslld (@X[0],@X[0],2);
1103 eval(shift(@insns)); # body_20_39
1104 eval(shift(@insns));
1105 eval(shift(@insns));
1106 eval(shift(@insns)); # rol
1107 eval(shift(@insns));
1108 eval(shift(@insns));
1109 eval(shift(@insns)); # ror
1110 eval(shift(@insns));
1112 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2
1113 eval(shift(@insns)); # body_20_39
1114 eval(shift(@insns));
1115 eval(shift(@insns));
1116 eval(shift(@insns)); # rol
1117 eval(shift(@insns));
1118 eval(shift(@insns));
1119 eval(shift(@insns)); # rol
1120 eval(shift(@insns));
1122 foreach (@insns) { eval; } # remaining instructions
1124 $Xi++; push(@X,shift(@X)); # "rotate" X[]
1127 sub Xuplast_avx_80()
1130 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1131 my ($a,$b,$c,$d,$e);
1133 eval(shift(@insns));
1134 &vpaddd (@Tx[1],$Kx,@X[-1&7]);
1135 eval(shift(@insns));
1136 eval(shift(@insns));
1137 eval(shift(@insns));
1138 eval(shift(@insns));
1140 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU
1142 foreach (@insns) { eval; } # remaining instructions
1145 &je ($done_avx_label);
1147 &vmovdqa(@X[2],"64($K_XX_XX)"); # pbswap mask
1148 &vmovdqa($Kx,"-64($K_XX_XX)"); # K_00_19
1149 &vmovdqu(@X[-4&7],"0($inp)"); # load input
1150 &vmovdqu(@X[-3&7],"16($inp)");
1151 &vmovdqu(@X[-2&7],"32($inp)");
1152 &vmovdqu(@X[-1&7],"48($inp)");
1153 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
1162 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1163 my ($a,$b,$c,$d,$e);
1165 eval(shift(@insns));
1166 eval(shift(@insns));
1167 &vpshufb(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1168 eval(shift(@insns));
1169 eval(shift(@insns));
1170 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],$Kx);
1171 eval(shift(@insns));
1172 eval(shift(@insns));
1173 eval(shift(@insns));
1174 eval(shift(@insns));
1175 &vmovdqa(eval(16*$Xi)."(%rsp)",@X[$Xi&7]); # X[]+K xfer to IALU
1176 eval(shift(@insns));
1177 eval(shift(@insns));
1179 foreach (@insns) { eval; }
1186 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions
1187 my ($a,$b,$c,$d,$e);
1189 foreach (@insns) { eval; }
1196 &Xupdate_avx_16_31(\&body_00_19);
1197 &Xupdate_avx_16_31(\&body_00_19);
1198 &Xupdate_avx_16_31(\&body_00_19);
1199 &Xupdate_avx_16_31(\&body_00_19);
1200 &Xupdate_avx_32_79(\&body_00_19);
1201 &Xupdate_avx_32_79(\&body_20_39);
1202 &Xupdate_avx_32_79(\&body_20_39);
1203 &Xupdate_avx_32_79(\&body_20_39);
1204 &Xupdate_avx_32_79(\&body_20_39);
1205 &Xupdate_avx_32_79(\&body_20_39);
1206 &Xupdate_avx_32_79(\&body_40_59);
1207 &Xupdate_avx_32_79(\&body_40_59);
1208 &Xupdate_avx_32_79(\&body_40_59);
1209 &Xupdate_avx_32_79(\&body_40_59);
1210 &Xupdate_avx_32_79(\&body_40_59);
1211 &Xupdate_avx_32_79(\&body_20_39);
1212 &Xuplast_avx_80(\&body_20_39); # can jump to "done"
1214 $saved_j=$j; @saved_V=@V;
1216 &Xloop_avx(\&body_20_39);
1217 &Xloop_avx(\&body_20_39);
1218 &Xloop_avx(\&body_20_39);
1221 add 0($ctx),$A # update context
1228 mov @T[0],$B # magic seed
1240 $j=$saved_j; @V=@saved_V;
1242 &Xtail_avx(\&body_20_39);
1243 &Xtail_avx(\&body_20_39);
1244 &Xtail_avx(\&body_20_39);
1249 add 0($ctx),$A # update context
1260 $code.=<<___ if ($win64);
1261 movaps -40-6*16(%r14),%xmm6
1262 movaps -40-5*16(%r14),%xmm7
1263 movaps -40-4*16(%r14),%xmm8
1264 movaps -40-3*16(%r14),%xmm9
1265 movaps -40-2*16(%r14),%xmm10
1266 movaps -40-1*16(%r14),%xmm11
1278 .size sha1_block_data_order_avx,.-sha1_block_data_order_avx
1283 $Xi=4; # reset variables
1284 @X=map("%ymm$_",(4..7,0..3));
1285 @Tx=map("%ymm$_",(8..10));
1289 my @ROTX=("%eax","%ebp","%ebx","%ecx","%edx","%esi");
1290 my ($a5,$t0)=("%r12d","%edi");
1292 my ($A,$F,$B,$C,$D,$E)=@ROTX;
1297 .type sha1_block_data_order_avx2,\@function,3
1299 sha1_block_data_order_avx2:
1309 $code.=<<___ if ($win64);
1310 lea -6*16(%rsp),%rsp
1311 vmovaps %xmm6,-40-6*16(%rax)
1312 vmovaps %xmm7,-40-5*16(%rax)
1313 vmovaps %xmm8,-40-4*16(%rax)
1314 vmovaps %xmm9,-40-3*16(%rax)
1315 vmovaps %xmm10,-40-2*16(%rax)
1316 vmovaps %xmm11,-40-1*16(%rax)
1320 mov %rax,%r14 # original %rsp
1321 mov %rdi,$ctx # reassigned argument
1322 mov %rsi,$inp # reassigned argument
1323 mov %rdx,$num # reassigned argument
1330 lea K_XX_XX+64(%rip),$K_XX_XX
1332 mov 0($ctx),$A # load context
1334 cmovae $inp,$frame # next or same block
1339 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask
1341 vmovdqu ($inp),%xmm0
1342 vmovdqu 16($inp),%xmm1
1343 vmovdqu 32($inp),%xmm2
1344 vmovdqu 48($inp),%xmm3
1346 vinserti128 \$1,($frame),@X[-4&7],@X[-4&7]
1347 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7]
1348 vpshufb @X[2],@X[-4&7],@X[-4&7]
1349 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7]
1350 vpshufb @X[2],@X[-3&7],@X[-3&7]
1351 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7]
1352 vpshufb @X[2],@X[-2&7],@X[-2&7]
1353 vmovdqu -64($K_XX_XX),$Kx # K_00_19
1354 vpshufb @X[2],@X[-1&7],@X[-1&7]
1356 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19
1357 vpaddd $Kx,@X[-3&7],@X[1]
1358 vmovdqu @X[0],0(%rsp) # X[]+K xfer to IALU
1359 vpaddd $Kx,@X[-2&7],@X[2]
1360 vmovdqu @X[1],32(%rsp)
1361 vpaddd $Kx,@X[-1&7],@X[3]
1362 vmovdqu @X[2],64(%rsp)
1363 vmovdqu @X[3],96(%rsp)
1365 for (;$Xi<8;$Xi++) { # Xupdate_avx2_16_31
1368 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1369 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1370 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1371 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1372 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1373 &vpsrld (@Tx[0],@X[0],31);
1374 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1375 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1376 &vpaddd (@X[0],@X[0],@X[0]);
1377 &vpsrld (@Tx[1],@Tx[2],30);
1378 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1379 &vpslld (@Tx[2],@Tx[2],2);
1380 &vpxor (@X[0],@X[0],@Tx[1]);
1381 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1382 &vpaddd (@Tx[1],@X[0],$Kx);
1383 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1385 push(@X,shift(@X)); # "rotate" X[]
1388 lea 128(%rsp),$frame
1397 sub bodyx_00_19 () { # 8 instructions, 3 cycles critical path
1398 # at start $f=(b&c)^(~b&d), $b>>>=2
1399 return &bodyx_20_39() if ($rx==19); $rx++;
1401 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1403 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1404 '&lea ($frame,"256($frame)") if ($j%32==31);',
1405 '&andn ($t0,$a,$c)', # ~b&d for next round
1407 '&add ($e,$f)', # e+=(b&c)^(~b&d)
1408 '&rorx ($a5,$a,27)', # a<<<5
1409 '&rorx ($f,$a,2)', # b>>>2 for next round
1410 '&and ($a,$b)', # b&c for next round
1412 '&add ($e,$a5)', # e+=a<<<5
1413 '&xor ($a,$t0);'. # f=(b&c)^(~b&d) for next round
1415 'unshift(@ROTX,pop(@ROTX)); $j++;'
1419 sub bodyx_20_39 () { # 7 instructions, 2 cycles critical path
1420 # on entry $f=b^c^d, $b>>>=2
1421 return &bodyx_40_59() if ($rx==39); $rx++;
1423 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1425 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1426 '&lea ($frame,"256($frame)") if ($j%32==31);',
1428 '&lea ($e,"($e,$f)")', # e+=b^c^d
1429 '&rorx ($a5,$a,27)', # a<<<5
1430 '&rorx ($f,$a,2) if ($j<79)', # b>>>2 in next round
1431 '&xor ($a,$b) if ($j<79)', # b^c for next round
1433 '&add ($e,$a5)', # e+=a<<<5
1434 '&xor ($a,$c) if ($j<79);'. # f=b^c^d for next round
1436 'unshift(@ROTX,pop(@ROTX)); $j++;'
1440 sub bodyx_40_59 () { # 10 instructions, 3 cycles critical path
1441 # on entry $f=((b^c)&(c^d)), $b>>>=2
1444 '($a,$f,$b,$c,$d,$e)=@ROTX;'.
1446 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K
1447 '&lea ($frame,"256($frame)") if ($j%32==31);',
1448 '&xor ($f,$c) if ($j>39)', # (b^c)&(c^d)^c
1449 '&mov ($t0,$b) if ($j<59)', # count on zero latency
1450 '&xor ($t0,$c) if ($j<59)', # c^d for next round
1452 '&lea ($e,"($e,$f)")', # e+=(b^c)&(c^d)^c
1453 '&rorx ($a5,$a,27)', # a<<<5
1454 '&rorx ($f,$a,2)', # b>>>2 in next round
1455 '&xor ($a,$b)', # b^c for next round
1457 '&add ($e,$a5)', # e+=a<<<5
1458 '&and ($a,$t0) if ($j< 59);'. # f=(b^c)&(c^d) for next round
1459 '&xor ($a,$c) if ($j==59);'. # f=b^c^d for next round
1461 'unshift(@ROTX,pop(@ROTX)); $j++;'
1465 sub Xupdate_avx2_16_31() # recall that $Xi starts wtih 4
1468 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 instructions
1469 my ($a,$b,$c,$d,$e);
1471 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]"
1472 eval(shift(@insns));
1473 eval(shift(@insns));
1474 eval(shift(@insns));
1475 eval(shift(@insns));
1477 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords
1478 eval(shift(@insns));
1479 eval(shift(@insns));
1480 eval(shift(@insns));
1482 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
1483 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
1484 eval(shift(@insns));
1485 eval(shift(@insns));
1487 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]"
1488 eval(shift(@insns));
1489 eval(shift(@insns));
1490 eval(shift(@insns));
1491 eval(shift(@insns));
1493 &vpsrld (@Tx[0],@X[0],31);
1494 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX
1495 eval(shift(@insns));
1496 eval(shift(@insns));
1497 eval(shift(@insns));
1499 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword
1500 &vpaddd (@X[0],@X[0],@X[0]);
1501 eval(shift(@insns));
1502 eval(shift(@insns));
1504 &vpsrld (@Tx[1],@Tx[2],30);
1505 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1
1506 eval(shift(@insns));
1507 eval(shift(@insns));
1509 &vpslld (@Tx[2],@Tx[2],2);
1510 &vpxor (@X[0],@X[0],@Tx[1]);
1511 eval(shift(@insns));
1512 eval(shift(@insns));
1514 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2
1515 eval(shift(@insns));
1516 eval(shift(@insns));
1517 eval(shift(@insns));
1519 &vpaddd (@Tx[1],@X[0],$Kx);
1520 eval(shift(@insns));
1521 eval(shift(@insns));
1522 eval(shift(@insns));
1523 &vmovdqu(eval(32*($Xi))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1525 foreach (@insns) { eval; } # remaining instructions [if any]
1528 push(@X,shift(@X)); # "rotate" X[]
1531 sub Xupdate_avx2_32_79()
1534 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 to 50 instructions
1535 my ($a,$b,$c,$d,$e);
1537 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]"
1538 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
1539 eval(shift(@insns));
1540 eval(shift(@insns));
1542 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
1543 &vmovdqu($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0);
1544 eval(shift(@insns));
1545 eval(shift(@insns));
1546 eval(shift(@insns));
1548 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]"
1549 eval(shift(@insns));
1550 eval(shift(@insns));
1551 eval(shift(@insns));
1553 &vpsrld (@Tx[0],@X[0],30);
1554 &vpslld (@X[0],@X[0],2);
1555 eval(shift(@insns));
1556 eval(shift(@insns));
1557 eval(shift(@insns));
1559 #&vpslld (@X[0],@X[0],2);
1560 eval(shift(@insns));
1561 eval(shift(@insns));
1562 eval(shift(@insns));
1564 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2
1565 eval(shift(@insns));
1566 eval(shift(@insns));
1567 eval(shift(@insns));
1568 eval(shift(@insns));
1570 &vpaddd (@Tx[1],@X[0],$Kx);
1571 eval(shift(@insns));
1572 eval(shift(@insns));
1573 eval(shift(@insns));
1574 eval(shift(@insns));
1576 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU
1578 foreach (@insns) { eval; } # remaining instructions
1581 push(@X,shift(@X)); # "rotate" X[]
1587 my @insns = (&$body,&$body,&$body,&$body,&$body); # 32 instructions
1588 my ($a,$b,$c,$d,$e);
1590 foreach (@insns) { eval; }
1594 &Xupdate_avx2_32_79(\&bodyx_00_19);
1595 &Xupdate_avx2_32_79(\&bodyx_00_19);
1596 &Xupdate_avx2_32_79(\&bodyx_00_19);
1597 &Xupdate_avx2_32_79(\&bodyx_00_19);
1599 &Xupdate_avx2_32_79(\&bodyx_20_39);
1600 &Xupdate_avx2_32_79(\&bodyx_20_39);
1601 &Xupdate_avx2_32_79(\&bodyx_20_39);
1602 &Xupdate_avx2_32_79(\&bodyx_20_39);
1605 &Xupdate_avx2_32_79(\&bodyx_40_59);
1606 &Xupdate_avx2_32_79(\&bodyx_40_59);
1607 &Xupdate_avx2_32_79(\&bodyx_40_59);
1608 &Xupdate_avx2_32_79(\&bodyx_40_59);
1610 &Xloop_avx2(\&bodyx_20_39);
1611 &Xloop_avx2(\&bodyx_20_39);
1612 &Xloop_avx2(\&bodyx_20_39);
1613 &Xloop_avx2(\&bodyx_20_39);
1616 lea 128($inp),$frame
1617 lea 128($inp),%rdi # borrow $t0
1619 cmovae $inp,$frame # next or previous block
1621 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c
1622 add 0($ctx),@ROTX[0] # update context
1623 add 4($ctx),@ROTX[1]
1624 add 8($ctx),@ROTX[3]
1625 mov @ROTX[0],0($ctx)
1626 add 12($ctx),@ROTX[4]
1627 mov @ROTX[1],4($ctx)
1628 mov @ROTX[0],$A # A=d
1629 add 16($ctx),@ROTX[5]
1631 mov @ROTX[3],8($ctx)
1632 mov @ROTX[4],$D # D=b
1633 #xchg @ROTX[5],$F # F=c, C=f
1634 mov @ROTX[4],12($ctx)
1635 mov @ROTX[1],$F # F=e
1636 mov @ROTX[5],16($ctx)
1638 mov @ROTX[5],$E # E=c
1640 #xchg $F,$E # E=c, F=e
1646 $Xi=4; # reset variables
1647 @X=map("%ymm$_",(4..7,0..3));
1650 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask
1651 cmp $num,%rdi # borrowed $t0
1654 vmovdqu -64(%rdi),%xmm0 # low part of @X[-4&7]
1655 vmovdqu -48(%rdi),%xmm1
1656 vmovdqu -32(%rdi),%xmm2
1657 vmovdqu -16(%rdi),%xmm3
1658 vinserti128 \$1,0($frame),@X[-4&7],@X[-4&7]
1659 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7]
1660 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7]
1661 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7]
1666 lea 128+16(%rsp),$frame
1673 $rx=$j=0; @ROTX=($A,$F,$B,$C,$D,$E);
1675 &Xloop_avx2 (\&bodyx_00_19);
1676 &Xloop_avx2 (\&bodyx_00_19);
1677 &Xloop_avx2 (\&bodyx_00_19);
1678 &Xloop_avx2 (\&bodyx_00_19);
1680 &Xloop_avx2 (\&bodyx_20_39);
1681 &vmovdqu ($Kx,"-64($K_XX_XX)"); # K_00_19
1682 &vpshufb (@X[-4&7],@X[-4&7],@X[2]); # byte swap
1683 &Xloop_avx2 (\&bodyx_20_39);
1684 &vpshufb (@X[-3&7],@X[-3&7],@X[2]);
1685 &vpaddd (@Tx[0],@X[-4&7],$Kx); # add K_00_19
1686 &Xloop_avx2 (\&bodyx_20_39);
1687 &vmovdqu ("0(%rsp)",@Tx[0]);
1688 &vpshufb (@X[-2&7],@X[-2&7],@X[2]);
1689 &vpaddd (@Tx[1],@X[-3&7],$Kx);
1690 &Xloop_avx2 (\&bodyx_20_39);
1691 &vmovdqu ("32(%rsp)",@Tx[1]);
1692 &vpshufb (@X[-1&7],@X[-1&7],@X[2]);
1693 &vpaddd (@X[2],@X[-2&7],$Kx);
1695 &Xloop_avx2 (\&bodyx_40_59);
1697 &vmovdqu ("64(%rsp)",@X[2]);
1698 &vpaddd (@X[3],@X[-1&7],$Kx);
1699 &Xloop_avx2 (\&bodyx_40_59);
1700 &vmovdqu ("96(%rsp)",@X[3]);
1701 &Xloop_avx2 (\&bodyx_40_59);
1702 &Xupdate_avx2_16_31(\&bodyx_40_59);
1704 &Xupdate_avx2_16_31(\&bodyx_20_39);
1705 &Xupdate_avx2_16_31(\&bodyx_20_39);
1706 &Xupdate_avx2_16_31(\&bodyx_20_39);
1707 &Xloop_avx2 (\&bodyx_20_39);
1710 lea 128(%rsp),$frame
1712 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c
1713 add 0($ctx),@ROTX[0] # update context
1714 add 4($ctx),@ROTX[1]
1715 add 8($ctx),@ROTX[3]
1716 mov @ROTX[0],0($ctx)
1717 add 12($ctx),@ROTX[4]
1718 mov @ROTX[1],4($ctx)
1719 mov @ROTX[0],$A # A=d
1720 add 16($ctx),@ROTX[5]
1722 mov @ROTX[3],8($ctx)
1723 mov @ROTX[4],$D # D=b
1724 #xchg @ROTX[5],$F # F=c, C=f
1725 mov @ROTX[4],12($ctx)
1726 mov @ROTX[1],$F # F=e
1727 mov @ROTX[5],16($ctx)
1729 mov @ROTX[5],$E # E=c
1731 #xchg $F,$E # E=c, F=e
1739 $code.=<<___ if ($win64);
1740 movaps -40-6*16(%r14),%xmm6
1741 movaps -40-5*16(%r14),%xmm7
1742 movaps -40-4*16(%r14),%xmm8
1743 movaps -40-3*16(%r14),%xmm9
1744 movaps -40-2*16(%r14),%xmm10
1745 movaps -40-1*16(%r14),%xmm11
1757 .size sha1_block_data_order_avx2,.-sha1_block_data_order_avx2
1764 .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19
1765 .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19
1766 .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39
1767 .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39
1768 .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59
1769 .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59
1770 .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79
1771 .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79
1772 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask
1773 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask
1774 .byte 0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0
1778 .asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1782 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1783 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
1791 .extern __imp_RtlVirtualUnwind
1792 .type se_handler,\@abi-omnipotent
1806 mov 120($context),%rax # pull context->Rax
1807 mov 248($context),%rbx # pull context->Rip
1809 lea .Lprologue(%rip),%r10
1810 cmp %r10,%rbx # context->Rip<.Lprologue
1811 jb .Lcommon_seh_tail
1813 mov 152($context),%rax # pull context->Rsp
1815 lea .Lepilogue(%rip),%r10
1816 cmp %r10,%rbx # context->Rip>=.Lepilogue
1817 jae .Lcommon_seh_tail
1819 mov `16*4`(%rax),%rax # pull saved stack pointer
1826 mov %rbx,144($context) # restore context->Rbx
1827 mov %rbp,160($context) # restore context->Rbp
1828 mov %r12,216($context) # restore context->R12
1829 mov %r13,224($context) # restore context->R13
1830 mov %r14,232($context) # restore context->R14
1832 jmp .Lcommon_seh_tail
1833 .size se_handler,.-se_handler
1835 .type shaext_handler,\@abi-omnipotent
1849 mov 120($context),%rax # pull context->Rax
1850 mov 248($context),%rbx # pull context->Rip
1852 lea .Lprologue_shaext(%rip),%r10
1853 cmp %r10,%rbx # context->Rip<.Lprologue
1854 jb .Lcommon_seh_tail
1856 lea .Lepilogue_shaext(%rip),%r10
1857 cmp %r10,%rbx # context->Rip>=.Lepilogue
1858 jae .Lcommon_seh_tail
1860 lea -8-4*16(%rax),%rsi
1861 lea 512($context),%rdi # &context.Xmm6
1863 .long 0xa548f3fc # cld; rep movsq
1865 jmp .Lcommon_seh_tail
1866 .size shaext_handler,.-shaext_handler
1868 .type ssse3_handler,\@abi-omnipotent
1882 mov 120($context),%rax # pull context->Rax
1883 mov 248($context),%rbx # pull context->Rip
1885 mov 8($disp),%rsi # disp->ImageBase
1886 mov 56($disp),%r11 # disp->HandlerData
1888 mov 0(%r11),%r10d # HandlerData[0]
1889 lea (%rsi,%r10),%r10 # prologue label
1890 cmp %r10,%rbx # context->Rip<prologue label
1891 jb .Lcommon_seh_tail
1893 mov 152($context),%rax # pull context->Rsp
1895 mov 4(%r11),%r10d # HandlerData[1]
1896 lea (%rsi,%r10),%r10 # epilogue label
1897 cmp %r10,%rbx # context->Rip>=epilogue label
1898 jae .Lcommon_seh_tail
1900 mov 232($context),%rax # pull context->R14
1902 lea -40-6*16(%rax),%rsi
1903 lea 512($context),%rdi # &context.Xmm6
1905 .long 0xa548f3fc # cld; rep movsq
1912 mov %rbx,144($context) # restore context->Rbx
1913 mov %rbp,160($context) # restore context->Rbp
1914 mov %r12,216($context) # restore cotnext->R12
1915 mov %r13,224($context) # restore cotnext->R13
1916 mov %r14,232($context) # restore cotnext->R14
1921 mov %rax,152($context) # restore context->Rsp
1922 mov %rsi,168($context) # restore context->Rsi
1923 mov %rdi,176($context) # restore context->Rdi
1925 mov 40($disp),%rdi # disp->ContextRecord
1926 mov $context,%rsi # context
1927 mov \$154,%ecx # sizeof(CONTEXT)
1928 .long 0xa548f3fc # cld; rep movsq
1931 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
1932 mov 8(%rsi),%rdx # arg2, disp->ImageBase
1933 mov 0(%rsi),%r8 # arg3, disp->ControlPc
1934 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
1935 mov 40(%rsi),%r10 # disp->ContextRecord
1936 lea 56(%rsi),%r11 # &disp->HandlerData
1937 lea 24(%rsi),%r12 # &disp->EstablisherFrame
1938 mov %r10,32(%rsp) # arg5
1939 mov %r11,40(%rsp) # arg6
1940 mov %r12,48(%rsp) # arg7
1941 mov %rcx,56(%rsp) # arg8, (NULL)
1942 call *__imp_RtlVirtualUnwind(%rip)
1944 mov \$1,%eax # ExceptionContinueSearch
1956 .size ssse3_handler,.-ssse3_handler
1960 .rva .LSEH_begin_sha1_block_data_order
1961 .rva .LSEH_end_sha1_block_data_order
1962 .rva .LSEH_info_sha1_block_data_order
1964 $code.=<<___ if ($shaext);
1965 .rva .LSEH_begin_sha1_block_data_order_shaext
1966 .rva .LSEH_end_sha1_block_data_order_shaext
1967 .rva .LSEH_info_sha1_block_data_order_shaext
1970 .rva .LSEH_begin_sha1_block_data_order_ssse3
1971 .rva .LSEH_end_sha1_block_data_order_ssse3
1972 .rva .LSEH_info_sha1_block_data_order_ssse3
1974 $code.=<<___ if ($avx);
1975 .rva .LSEH_begin_sha1_block_data_order_avx
1976 .rva .LSEH_end_sha1_block_data_order_avx
1977 .rva .LSEH_info_sha1_block_data_order_avx
1979 $code.=<<___ if ($avx>1);
1980 .rva .LSEH_begin_sha1_block_data_order_avx2
1981 .rva .LSEH_end_sha1_block_data_order_avx2
1982 .rva .LSEH_info_sha1_block_data_order_avx2
1987 .LSEH_info_sha1_block_data_order:
1990 .LSEH_info_sha1_block_data_order_shaext:
1993 .LSEH_info_sha1_block_data_order_ssse3:
1996 .rva .Lprologue_ssse3,.Lepilogue_ssse3 # HandlerData[]
1998 $code.=<<___ if ($avx);
1999 .LSEH_info_sha1_block_data_order_avx:
2002 .rva .Lprologue_avx,.Lepilogue_avx # HandlerData[]
2004 $code.=<<___ if ($avx>1);
2005 .LSEH_info_sha1_block_data_order_avx2:
2008 .rva .Lprologue_avx2,.Lepilogue_avx2 # HandlerData[]
2012 ####################################################################
2015 if (@_[0] =~ /\$([x0-9a-f]+),\s*%xmm([0-7]),\s*%xmm([0-7])/) {
2016 my @opcode=(0x0f,0x3a,0xcc);
2017 push @opcode,0xc0|($2&7)|(($3&7)<<3); # ModR/M
2019 push @opcode,$c=~/^0/?oct($c):$c;
2020 return ".byte\t".join(',',@opcode);
2022 return "sha1rnds4\t".@_[0];
2029 "sha1nexte" => 0xc8,
2031 "sha1msg2" => 0xca );
2033 if (defined($opcodelet{$instr}) && @_[0] =~ /%xmm([0-7]),\s*%xmm([0-7])/) {
2034 my @opcode=(0x0f,0x38);
2035 push @opcode,$opcodelet{$instr};
2036 push @opcode,0xc0|($1&7)|(($2&7)<<3); # ModR/M
2037 return ".byte\t".join(',',@opcode);
2039 return $instr."\t".@_[0];
2043 foreach (split("\n",$code)) {
2044 s/\`([^\`]*)\`/eval $1/geo;
2046 s/\b(sha1rnds4)\s+(.*)/sha1rnds4($2)/geo or
2047 s/\b(sha1[^\s]*)\s+(.*)/sha1op38($1,$2)/geo;