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 # ====================================================================
12 # This is RC4+MD5 "stitch" implementation. The idea, as spelled in
13 # http://download.intel.com/design/intarch/papers/323686.pdf, is that
14 # since both algorithms exhibit instruction-level parallelism, ILP,
15 # below theoretical maximum, interleaving them would allow to utilize
16 # processor resources better and achieve better performance. RC4
17 # instruction sequence is virtually identical to rc4-x86_64.pl, which
18 # is heavily based on submission by Maxim Perminov, Maxim Locktyukhin
19 # and Jim Guilford of Intel. MD5 is fresh implementation aiming to
20 # minimize register usage, which was used as "main thread" with RC4
21 # weaved into it, one RC4 round per one MD5 round. In addition to the
22 # stiched subroutine the script can generate standalone replacement
23 # md5_block_asm_data_order and RC4. Below are performance numbers in
24 # cycles per processed byte, less is better, for these the standalone
25 # subroutines, sum of them, and stitched one:
27 # RC4 MD5 RC4+MD5 stitch gain
28 # Opteron 6.5(*) 5.4 11.9 7.0 +70%(*)
29 # Core2 6.5 5.8 12.3 7.7 +60%
30 # Westmere 4.3 5.2 9.5 7.0 +36%
31 # Sandy Bridge 4.2 5.5 9.7 6.8 +43%
32 # Ivy Bridge 4.1 5.2 9.3 6.0 +54%
33 # Haswell 4.0 5.0 9.0 5.7 +60%
34 # Skylake 6.3(**) 5.0 11.3 5.3 +110%
35 # Atom 9.3 6.5 15.8 11.1 +42%
36 # VIA Nano 6.3 5.4 11.7 8.6 +37%
37 # Bulldozer 4.5 5.4 9.9 7.7 +29%
39 # (*) rc4-x86_64.pl delivers 5.3 on Opteron, so real improvement
41 # (**) unidentified anomaly;
43 my ($rc4,$md5)=(1,1); # what to generate?
44 my $D="#" if (!$md5); # if set to "#", MD5 is stitched into RC4(),
45 # but its result is discarded. Idea here is
46 # to be able to use 'openssl speed rc4' for
47 # benchmarking the stitched subroutine...
51 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
53 my $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
55 $0 =~ m/(.*[\/\\])[^\/\\]+$/; my $dir=$1; my $xlate;
56 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
57 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
58 die "can't locate x86_64-xlate.pl";
60 open OUT,"| \"$^X\" $xlate $flavour $output";
63 my ($dat,$in0,$out,$ctx,$inp,$len, $func,$nargs);
66 ($dat,$len,$in0,$out) = ("%rdi","%rsi","%rdx","%rcx");
67 $func="RC4"; $nargs=4;
68 } elsif ($md5 && !$rc4) {
69 ($ctx,$inp,$len) = ("%rdi","%rsi","%rdx");
70 $func="md5_block_asm_data_order"; $nargs=3;
72 ($dat,$in0,$out,$ctx,$inp,$len) = ("%rdi","%rsi","%rdx","%rcx","%r8","%r9");
73 $func="rc4_md5_enc"; $nargs=6;
76 # const void *in0, # RC4 input
77 # void *out, # RC4 output
79 # const void *inp, # MD5 input
80 # size_t len); # number of 64-byte blocks
83 my @K=( 0xd76aa478,0xe8c7b756,0x242070db,0xc1bdceee,
84 0xf57c0faf,0x4787c62a,0xa8304613,0xfd469501,
85 0x698098d8,0x8b44f7af,0xffff5bb1,0x895cd7be,
86 0x6b901122,0xfd987193,0xa679438e,0x49b40821,
88 0xf61e2562,0xc040b340,0x265e5a51,0xe9b6c7aa,
89 0xd62f105d,0x02441453,0xd8a1e681,0xe7d3fbc8,
90 0x21e1cde6,0xc33707d6,0xf4d50d87,0x455a14ed,
91 0xa9e3e905,0xfcefa3f8,0x676f02d9,0x8d2a4c8a,
93 0xfffa3942,0x8771f681,0x6d9d6122,0xfde5380c,
94 0xa4beea44,0x4bdecfa9,0xf6bb4b60,0xbebfbc70,
95 0x289b7ec6,0xeaa127fa,0xd4ef3085,0x04881d05,
96 0xd9d4d039,0xe6db99e5,0x1fa27cf8,0xc4ac5665,
98 0xf4292244,0x432aff97,0xab9423a7,0xfc93a039,
99 0x655b59c3,0x8f0ccc92,0xffeff47d,0x85845dd1,
100 0x6fa87e4f,0xfe2ce6e0,0xa3014314,0x4e0811a1,
101 0xf7537e82,0xbd3af235,0x2ad7d2bb,0xeb86d391 );
103 my @V=("%r8d","%r9d","%r10d","%r11d"); # MD5 registers
106 my @XX=("%rbp","%rsi"); # RC4 registers
107 my @TX=("%rax","%rbx");
111 my $MOD=32; # 16, 32 or 64
118 .type $func,\@function,$nargs
133 $D#md5# mov $ctx,%r11 # reassign arguments
137 $D#md5# mov $inp,%r15
139 $ctx="%r11" if ($md5); # reassign arguments
143 $inp="%r15" if ($md5);
144 $inp=$in0 if (!$md5);
150 mov -8($dat),$XX[0]#b
155 movl ($dat,$XX[0],4),$TX[0]#d
157 $code.=<<___ if (!$md5);
162 and \$`$MOD-1`,$TX[1]
163 jz .Loop${MOD}_is_hot
167 movl ($dat,$YY,4),$TY#d
168 movl $TX[0]#d,($dat,$YY,4)
169 movl $TY#d,($dat,$XX[0],4)
172 movl ($dat,$TX[0],4),$TY#d
173 movl ($dat,$XX[0],4),$TX[0]#d
175 movb $TY#b,($out,$in0)
178 jnz .Loop${MOD}_warmup
185 mov $len,32(%rsp) # save original $len
186 shr \$6,$len # number of 64-byte blocks
188 if ($D && !$md5) { # stitch in dummy MD5
199 #rc4# add $TX[0]#b,$YY#b
200 #rc4# lea ($dat,$XX[0],4),$XX[1]
202 add $inp,$len # pointer to the end of input
205 #md5# mov $ctx,24(%rsp) # save pointer to MD5_CTX
206 #md5# mov 0*4($ctx),$V[0] # load current hash value from MD5_CTX
207 #md5# mov 1*4($ctx),$V[1]
208 #md5# mov 2*4($ctx),$V[2]
209 #md5# mov 3*4($ctx),$V[3]
214 #md5# mov $V[0],0*4(%rsp) # put aside current hash value
215 #md5# mov $V[1],1*4(%rsp)
216 #md5# mov $V[2],2*4(%rsp)
217 #md5# mov $V[3],$tmp # forward reference
218 #md5# mov $V[3],3*4(%rsp)
222 my ($i,$a,$b,$c,$d)=@_;
223 my @rot0=(7,12,17,22);
226 my $xmm="%xmm".($j&1);
227 $code.=" movdqu ($in0),%xmm2\n" if ($rc4 && $j==15);
228 $code.=" add \$$MOD,$XX[0]#b\n" if ($rc4 && $j==15 && $k==$MOD-1);
229 $code.=" pxor $xmm,$xmm\n" if ($rc4 && $j<=1);
231 #rc4# movl ($dat,$YY,4),$TY#d
233 #rc4# movl $TX[0]#d,($dat,$YY,4)
235 #md5# add 4*`$j`($inp),$a
236 #rc4# add $TY#b,$TX[0]#b
237 #rc4# movl `4*(($k+1)%$MOD)`(`$k==$MOD-1?"$dat,$XX[0],4":"$XX[1]"`),$TX[1]#d
238 #md5# add \$$K[$i],$a
240 #rc4# movz $TX[0]#b,$TX[0]#d
241 #rc4# movl $TY#d,4*$k($XX[1])
243 #rc4# add $TX[1]#b,$YY#b
244 #md5# rol \$$rot0[$j%4],$a
245 #md5# mov `$j==15?"$b":"$c"`,$tmp # forward reference
246 #rc4# pinsrw \$`($j>>1)&7`,($dat,$TX[0],4),$xmm\n
249 $code.=<<___ if ($rc4 && $j==15 && $k==$MOD-1);
251 xor $YY,$YY # keyword to partial register
253 lea ($dat,$XX[0],4),$XX[1]
255 $code.=<<___ if ($rc4 && $j==15);
262 my ($i,$a,$b,$c,$d)=@_;
263 my @rot1=(5,9,14,20);
266 my $xmm="%xmm".($j&1);
267 $code.=" movdqu 16($in0),%xmm3\n" if ($rc4 && $j==15);
268 $code.=" add \$$MOD,$XX[0]#b\n" if ($rc4 && $j==15 && $k==$MOD-1);
269 $code.=" pxor $xmm,$xmm\n" if ($rc4 && $j<=1);
271 #rc4# movl ($dat,$YY,4),$TY#d
273 #rc4# movl $TX[0]#d,($dat,$YY,4)
275 #md5# add 4*`((1+5*$j)%16)`($inp),$a
276 #rc4# add $TY#b,$TX[0]#b
277 #rc4# movl `4*(($k+1)%$MOD)`(`$k==$MOD-1?"$dat,$XX[0],4":"$XX[1]"`),$TX[1]#d
278 #md5# add \$$K[$i],$a
280 #rc4# movz $TX[0]#b,$TX[0]#d
281 #rc4# movl $TY#d,4*$k($XX[1])
283 #rc4# add $TX[1]#b,$YY#b
284 #md5# rol \$$rot1[$j%4],$a
285 #md5# mov `$j==15?"$c":"$b"`,$tmp # forward reference
286 #rc4# pinsrw \$`($j>>1)&7`,($dat,$TX[0],4),$xmm\n
289 $code.=<<___ if ($rc4 && $j==15 && $k==$MOD-1);
291 xor $YY,$YY # keyword to partial register
293 lea ($dat,$XX[0],4),$XX[1]
295 $code.=<<___ if ($rc4 && $j==15);
302 my ($i,$a,$b,$c,$d)=@_;
303 my @rot2=(4,11,16,23);
306 my $xmm="%xmm".($j&1);
307 $code.=" movdqu 32($in0),%xmm4\n" if ($rc4 && $j==15);
308 $code.=" add \$$MOD,$XX[0]#b\n" if ($rc4 && $j==15 && $k==$MOD-1);
309 $code.=" pxor $xmm,$xmm\n" if ($rc4 && $j<=1);
311 #rc4# movl ($dat,$YY,4),$TY#d
313 #rc4# movl $TX[0]#d,($dat,$YY,4)
315 #md5# add 4*`((5+3*$j)%16)`($inp),$a
316 #rc4# add $TY#b,$TX[0]#b
317 #rc4# movl `4*(($k+1)%$MOD)`(`$k==$MOD-1?"$dat,$XX[0],4":"$XX[1]"`),$TX[1]#d
318 #md5# add \$$K[$i],$a
319 #rc4# movz $TX[0]#b,$TX[0]#d
321 #rc4# movl $TY#d,4*$k($XX[1])
322 #rc4# add $TX[1]#b,$YY#b
323 #md5# rol \$$rot2[$j%4],$a
324 #md5# mov `$j==15?"\\\$-1":"$c"`,$tmp # forward reference
325 #rc4# pinsrw \$`($j>>1)&7`,($dat,$TX[0],4),$xmm\n
328 $code.=<<___ if ($rc4 && $j==15 && $k==$MOD-1);
330 xor $YY,$YY # keyword to partial register
332 lea ($dat,$XX[0],4),$XX[1]
334 $code.=<<___ if ($rc4 && $j==15);
341 my ($i,$a,$b,$c,$d)=@_;
342 my @rot3=(6,10,15,21);
345 my $xmm="%xmm".($j&1);
346 $code.=" movdqu 48($in0),%xmm5\n" if ($rc4 && $j==15);
347 $code.=" add \$$MOD,$XX[0]#b\n" if ($rc4 && $j==15 && $k==$MOD-1);
348 $code.=" pxor $xmm,$xmm\n" if ($rc4 && $j<=1);
350 #rc4# movl ($dat,$YY,4),$TY#d
352 #rc4# movl $TX[0]#d,($dat,$YY,4)
354 #md5# add 4*`((7*$j)%16)`($inp),$a
355 #rc4# add $TY#b,$TX[0]#b
356 #rc4# movl `4*(($k+1)%$MOD)`(`$k==$MOD-1?"$dat,$XX[0],4":"$XX[1]"`),$TX[1]#d
357 #md5# add \$$K[$i],$a
358 #rc4# movz $TX[0]#b,$TX[0]#d
360 #rc4# movl $TY#d,4*$k($XX[1])
362 #rc4# add $TX[1]#b,$YY#b
363 #md5# rol \$$rot3[$j%4],$a
364 #md5# mov \$-1,$tmp # forward reference
365 #rc4# pinsrw \$`($j>>1)&7`,($dat,$TX[0],4),$xmm\n
368 $code.=<<___ if ($rc4 && $j==15);
370 xor $XX[0],$XX[0] # keyword to partial register
371 mov $XX[1]#b,$XX[0]#b
373 xor $YY,$YY # keyword to partial register
375 lea ($dat,$XX[0],4),$XX[1]
383 for(;$i<16;$i++) { R0($i,@V); unshift(@V,pop(@V)); push(@TX,shift(@TX)); }
384 for(;$i<32;$i++) { R1($i,@V); unshift(@V,pop(@V)); push(@TX,shift(@TX)); }
385 for(;$i<48;$i++) { R2($i,@V); unshift(@V,pop(@V)); push(@TX,shift(@TX)); }
386 for(;$i<64;$i++) { R3($i,@V); unshift(@V,pop(@V)); push(@TX,shift(@TX)); }
389 #md5# add 0*4(%rsp),$V[0] # accumulate hash value
390 #md5# add 1*4(%rsp),$V[1]
391 #md5# add 2*4(%rsp),$V[2]
392 #md5# add 3*4(%rsp),$V[3]
394 #rc4# movdqu %xmm2,($out,$in0) # write RC4 output
395 #rc4# movdqu %xmm3,16($out,$in0)
396 #rc4# movdqu %xmm4,32($out,$in0)
397 #rc4# movdqu %xmm5,48($out,$in0)
398 #md5# lea 64($inp),$inp
399 #rc4# lea 64($in0),$in0
400 cmp 16(%rsp),$inp # are we done?
403 #md5# mov 24(%rsp),$len # restore pointer to MD5_CTX
404 #rc4# sub $TX[0]#b,$YY#b # correct $YY
405 #md5# mov $V[0],0*4($len) # write MD5_CTX
406 #md5# mov $V[1],1*4($len)
407 #md5# mov $V[2],2*4($len)
408 #md5# mov $V[3],3*4($len)
410 $code.=<<___ if ($rc4 && (!$md5 || $D));
411 mov 32(%rsp),$len # restore original $len
412 and \$63,$len # remaining bytes
419 movl ($dat,$YY,4),$TY#d
420 movl $TX[0]#d,($dat,$YY,4)
421 movl $TY#d,($dat,$XX[0],4)
424 movl ($dat,$TX[0],4),$TY#d
425 movl ($dat,$XX[0],4),$TX[0]#d
427 movb $TY#b,($out,$in0)
435 #rc4# sub \$1,$XX[0]#b
436 #rc4# movl $XX[0]#d,-8($dat)
437 #rc4# movl $YY#d,-4($dat)
452 if ($rc4 && $D) { # sole purpose of this section is to provide
453 # option to use the generated module as drop-in
454 # replacement for rc4-x86_64.pl for debugging
455 # and testing purposes...
456 my ($idx,$ido)=("%r8","%r9");
457 my ($dat,$len,$inp)=("%rdi","%rsi","%rdx");
461 .type RC4_set_key,\@function,3
476 mov %eax,($dat,%rax,4)
484 mov ($dat,$ido,4),%r10d
485 add ($inp,$len,1),$idx#b
488 mov ($dat,$idx,4),%r11d
490 mov %r10d,($dat,$idx,4)
491 mov %r11d,($dat,$ido,4)
499 .size RC4_set_key,.-RC4_set_key
502 .type RC4_options,\@abi-omnipotent
505 lea .Lopts(%rip),%rax
509 .asciz "rc4(64x,int)"
511 .size RC4_options,.-RC4_options
514 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
515 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
523 .extern __imp_RtlVirtualUnwind
524 .type se_handler,\@abi-omnipotent
538 mov 120($context),%rax # pull context->Rax
539 mov 248($context),%rbx # pull context->Rip
541 lea .Lbody(%rip),%r10
542 cmp %r10,%rbx # context->Rip<.Lbody
545 mov 152($context),%rax # pull context->Rsp
547 lea .Lepilogue(%rip),%r10
548 cmp %r10,%rbx # context->Rip>=.Lepilogue
559 mov %rbx,144($context) # restore context->Rbx
560 mov %rbp,160($context) # restore context->Rbp
561 mov %r12,216($context) # restore context->R12
562 mov %r13,224($context) # restore context->R12
563 mov %r14,232($context) # restore context->R14
564 mov %r15,240($context) # restore context->R15
569 mov %rax,152($context) # restore context->Rsp
570 mov %rsi,168($context) # restore context->Rsi
571 mov %rdi,176($context) # restore context->Rdi
573 mov 40($disp),%rdi # disp->ContextRecord
574 mov $context,%rsi # context
575 mov \$154,%ecx # sizeof(CONTEXT)
576 .long 0xa548f3fc # cld; rep movsq
579 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
580 mov 8(%rsi),%rdx # arg2, disp->ImageBase
581 mov 0(%rsi),%r8 # arg3, disp->ControlPc
582 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
583 mov 40(%rsi),%r10 # disp->ContextRecord
584 lea 56(%rsi),%r11 # &disp->HandlerData
585 lea 24(%rsi),%r12 # &disp->EstablisherFrame
586 mov %r10,32(%rsp) # arg5
587 mov %r11,40(%rsp) # arg6
588 mov %r12,48(%rsp) # arg7
589 mov %rcx,56(%rsp) # arg8, (NULL)
590 call *__imp_RtlVirtualUnwind(%rip)
592 mov \$1,%eax # ExceptionContinueSearch
604 .size se_handler,.-se_handler
608 .rva .LSEH_begin_$func
610 .rva .LSEH_info_$func
622 if ($reg =~ /%r[0-9]+/) { $reg .= $conv; }
623 elsif ($conv eq "b") { $reg =~ s/%[er]([^x]+)x?/%$1l/; }
624 elsif ($conv eq "w") { $reg =~ s/%[er](.+)/%$1/; }
625 elsif ($conv eq "d") { $reg =~ s/%[er](.+)/%e$1/; }
629 $code =~ s/(%[a-z0-9]+)#([bwd])/reg_part($1,$2)/gem;
630 $code =~ s/\`([^\`]*)\`/eval $1/gem;
631 $code =~ s/pinsrw\s+\$0,/movd /gm;
633 $code =~ s/#md5#//gm if ($md5);
634 $code =~ s/#rc4#//gm if ($rc4);