3 # ====================================================================
4 # Written by Andy Polyakov <appro@fy.chalmers.se> 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 # 2.22x RC4 tune-up:-) It should be noted though that my hand [as in
11 # "hand-coded assembler"] doesn't stand for the whole improvement
12 # coefficient. It turned out that eliminating RC4_CHAR from config
13 # line results in ~40% improvement (yes, even for C implementation).
14 # Presumably it has everything to do with AMD cache architecture and
15 # RAW or whatever penalties. Once again! The module *requires* config
16 # line *without* RC4_CHAR! As for coding "secret," I bet on partial
17 # register arithmetics. For example instead of 'inc %r8; and $255,%r8'
18 # I simply 'inc %r8b'. Even though optimization manual discourages
19 # to operate on partial registers, it turned out to be the best bet.
20 # At least for AMD... How IA32E would perform remains to be seen...
22 # As was shown by Marc Bevand reordering of couple of load operations
23 # results in even higher performance gain of 3.3x:-) At least on
24 # Opteron... For reference, 1x in this case is RC4_CHAR C-code
25 # compiled with gcc 3.3.2, which performs at ~54MBps per 1GHz clock.
26 # Latter means that if you want to *estimate* what to expect from
27 # *your* Opteron, then multiply 54 by 3.3 and clock frequency in GHz.
29 # Intel P4 EM64T core was found to run the AMD64 code really slow...
30 # The only way to achieve comparable performance on P4 was to keep
31 # RC4_CHAR. Kind of ironic, huh? As it's apparently impossible to
32 # compose blended code, which would perform even within 30% marginal
33 # on either AMD and Intel platforms, I implement both cases. See
34 # rc4_skey.c for further details...
36 # P4 EM64T core appears to be "allergic" to 64-bit inc/dec. Replacing
37 # those with add/sub results in 50% performance improvement of folded
40 # As was shown by Zou Nanhai loop unrolling can improve Intel EM64T
41 # performance by >30% [unlike P4 32-bit case that is]. But this is
42 # provided that loads are reordered even more aggressively! Both code
43 # pathes, AMD64 and EM64T, reorder loads in essentially same manner
44 # as my IA-64 implementation. On Opteron this resulted in modest 5%
45 # improvement [I had to test it], while final Intel P4 performance
46 # achieves respectful 432MBps on 2.8GHz processor now. For reference.
47 # If executed on Xeon, current RC4_CHAR code-path is 2.7x faster than
48 # RC4_INT code-path. While if executed on Opteron, it's only 25%
49 # slower than the RC4_INT one [meaning that if CPU ยต-arch detection
50 # is not implemented, then this final RC4_CHAR code-path should be
51 # preferred, as it provides better *all-round* performance].
53 # Intel Core2 was observed to perform poorly on both code paths:-( It
54 # apparently suffers from some kind of partial register stall, which
55 # occurs in 64-bit mode only [as virtually identical 32-bit loop was
56 # observed to outperform 64-bit one by almost 50%]. Adding two movzb to
57 # cloop1 boosts its performance by 80%! This loop appears to be optimal
58 # fit for Core2 and therefore the code was modified to skip cloop8 on
63 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
64 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
65 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
66 die "can't locate x86_64-xlate.pl";
68 open STDOUT,"| $^X $xlate $output";
84 .type RC4,\@function,4
94 movl -8($dat),$XX[0]#d
99 movl ($dat,$XX[0],4),$TX[0]#d
106 for ($i=0;$i<8;$i++) {
110 movl ($dat,$YY,4),$TY#d
111 ror \$8,%rax # ror is redundant when $i=0
113 movl ($dat,$XX[1],4),$TX[1]#d
115 movl $TX[0]#d,($dat,$YY,4)
117 movl $TY#d,($dat,$XX[0],4)
119 movb ($dat,$TY,4),%al
121 push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers
140 movl $XX[0]#d,-8($dat)
149 movl ($dat,$YY,4),$TY#d
150 movl $TX[0]#d,($dat,$YY,4)
151 movl $TY#d,($dat,$XX[0],4)
154 movl ($dat,$TX[0],4),$TY#d
155 movl ($dat,$XX[0],4),$TX[0]#d
167 movzb ($dat,$XX[0]),$TX[0]#d
179 # unroll 2x4-wise, because 64-bit rotates kill Intel P4...
180 for ($i=0;$i<4;$i++) {
184 movzb ($dat,$YY),$TY#d
185 movzb $XX[1]#b,$XX[1]#d
186 movzb ($dat,$XX[1]),$TX[1]#d
187 movb $TX[0]#b,($dat,$YY)
189 movb $TY#b,($dat,$XX[0])
190 jne .Lcmov$i # Intel cmov is sloooow...
197 push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers
199 for ($i=4;$i<8;$i++) {
203 movzb ($dat,$YY),$TY#d
204 movzb $XX[1]#b,$XX[1]#d
205 movzb ($dat,$XX[1]),$TX[1]#d
206 movb $TX[0]#b,($dat,$YY)
208 movb $TY#b,($dat,$XX[0])
209 jne .Lcmov$i # Intel cmov is sloooow...
216 push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers
236 movzb ($dat,$YY),$TY#d
237 movb $TX[0]#b,($dat,$YY)
238 movb $TY#b,($dat,$XX[0])
242 movzb $XX[0]#b,$XX[0]#d
243 movzb ($dat,$TY),$TY#d
244 movzb ($dat,$XX[0]),$TX[0]#d
259 .extern OPENSSL_ia32cap_P
261 .type RC4_set_key,\@function,3
273 mov OPENSSL_ia32cap_P(%rip),$idx#d
283 mov %eax,($dat,%rax,4)
291 mov ($dat,$ido,4),%r10d
292 add ($inp,$len,1),$idx#b
295 mov ($dat,$idx,4),%r11d
297 mov %r10d,($dat,$idx,4)
298 mov %r11d,($dat,$ido,4)
313 mov ($dat,$ido),%r10b
314 add ($inp,$len),$idx#b
317 mov ($dat,$idx),%r11b
321 mov %r10b,($dat,$idx)
322 mov %r11b,($dat,$ido)
333 .size RC4_set_key,.-RC4_set_key
336 .type RC4_options,\@function,0
340 lea .Lopts-.(%rax),%rax
341 mov OPENSSL_ia32cap_P(%rip),%edx
353 .asciz "rc4(8x,char)"
354 .asciz "rc4(1x,char)"
355 .asciz "RC4 for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
357 .size RC4_options,.-RC4_options
360 $code =~ s/#([bwd])/$1/gm;