3 # ====================================================================
4 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
5 # project. Rights for redistribution and usage in source and binary
6 # forms are granted according to the OpenSSL license.
7 # ====================================================================
9 # 2.22x RC4 tune-up:-) It should be noted though that my hand [as in
10 # "hand-coded assembler"] doesn't stand for the whole improvement
11 # coefficient. It turned out that eliminating RC4_CHAR from config
12 # line results in ~40% improvement (yes, even for C implementation).
13 # Presumably it has everything to do with AMD cache architecture and
14 # RAW or whatever penalties. Once again! The module *requires* config
15 # line *without* RC4_CHAR! As for coding "secret," I bet on partial
16 # register arithmetics. For example instead of 'inc %r8; and $255,%r8'
17 # I simply 'inc %r8b'. Even though optimization manual discourages
18 # to operate on partial registers, it turned out to be the best bet.
19 # At least for AMD... How IA32E would perform remains to be seen...
21 # As was shown by Marc Bevand reordering of couple of load operations
22 # results in even higher performance gain of 3.3x:-) At least on
23 # Opteron... For reference, 1x in this case is RC4_CHAR C-code
24 # compiled with gcc 3.3.2, which performs at ~54MBps per 1GHz clock.
25 # Latter means that if you want to *estimate* what to expect from
26 # *your* Opteron, then multiply 54 by 3.3 and clock frequency in GHz.
28 # Intel P4 EM64T core was found to run the AMD64 code really slow...
29 # The only way to achieve comparable performance on P4 was to keep
30 # RC4_CHAR. Kind of ironic, huh? As it's apparently impossible to
31 # compose blended code, which would perform even within 30% marginal
32 # on either AMD and Intel platforms, I implement both cases. See
33 # rc4_skey.c for further details...
35 # P4 EM64T core appears to be "allergic" to 64-bit inc/dec. Replacing
36 # those with add/sub results in 50% performance improvement of folded
39 # As was shown by Zou Nanhai loop unrolling can improve Intel EM64T
40 # performance by >30% [unlike P4 32-bit case that is]. But this is
41 # provided that loads are reordered even more aggressively! Both code
42 # pathes, AMD64 and EM64T, reorder loads in essentially same manner
43 # as my IA-64 implementation. On Opteron this resulted in modest 5%
44 # improvement [I had to test it], while final Intel P4 performance
45 # achieves respectful 432MBps on 2.8GHz processor now. For reference.
46 # If executed on Xeon, current RC4_CHAR code-path is 2.7x faster than
47 # RC4_INT code-path. While if executed on Opteron, it's only 25%
48 # slower slower than the RC4_INT one [meaning that if detecting CPU
49 # is not desired, then RC4_CHAR code-path should be preferred, as it
50 # provides better *all-round* performance].
53 open STDOUT,"| $^X ../perlasm/x86_64-xlate.pl $output";
69 .type RC4,\@function,4
79 movl -8($dat),$XX[0]#d
84 movl ($dat,$XX[0],4),$TX[0]#d
91 for ($i=0;$i<8;$i++) {
95 movl ($dat,$YY,4),$TY#d
96 ror \$8,%rax # ror is redundant when $i=0
98 movl ($dat,$XX[1],4),$TX[1]#d
100 movl $TX[0]#d,($dat,$YY,4)
102 movl $TY#d,($dat,$XX[0],4)
104 movb ($dat,$TY,4),%al
106 push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers
125 movl $XX[0]#d,-8($dat)
134 movl ($dat,$YY,4),$TY#d
135 movl $TX[0]#d,($dat,$YY,4)
136 movl $TY#d,($dat,$XX[0],4)
139 movl ($dat,$TX[0],4),$TY#d
140 movl ($dat,$XX[0],4),$TX[0]#d
152 movzb ($dat,$XX[0]),$TX[0]#d
162 # unroll 2x4-wise, because 64-bit rotates kill Intel P4...
163 for ($i=0;$i<4;$i++) {
167 movzb ($dat,$YY),$TY#d
168 movzb $XX[1]#b,$XX[1]#d
169 movzb ($dat,$XX[1]),$TX[1]#d
170 movb $TX[0]#b,($dat,$YY)
172 movb $TY#b,($dat,$XX[0])
173 jne .Lcmov$i # Intel cmov is sloooow...
180 push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers
182 for ($i=4;$i<8;$i++) {
186 movzb ($dat,$YY),$TY#d
187 movzb $XX[1]#b,$XX[1]
188 movzb ($dat,$XX[1]),$TX[1]#d
189 movb $TX[0]#b,($dat,$YY)
191 movb $TY#b,($dat,$XX[0])
192 jne .Lcmov$i # Intel cmov is sloooow...
199 push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers
219 movzb ($dat,$YY),$TY#d
220 movb $TX[0]#b,($dat,$YY)
221 movb $TY#b,($dat,$XX[0])
224 movzb ($dat,$TY),$TY#d
225 movzb ($dat,$XX[0]),$TX[0]#d
236 $code =~ s/#([bwd])/$1/gm;