summaryrefslogtreecommitdiff
path: root/src/lib/libcrypto/rc4/asm
diff options
context:
space:
mode:
Diffstat (limited to 'src/lib/libcrypto/rc4/asm')
-rw-r--r--src/lib/libcrypto/rc4/asm/rc4-586.pl229
-rwxr-xr-xsrc/lib/libcrypto/rc4/asm/rc4-amd64.pl227
2 files changed, 456 insertions, 0 deletions
diff --git a/src/lib/libcrypto/rc4/asm/rc4-586.pl b/src/lib/libcrypto/rc4/asm/rc4-586.pl
new file mode 100644
index 0000000000..d6e98f0811
--- /dev/null
+++ b/src/lib/libcrypto/rc4/asm/rc4-586.pl
@@ -0,0 +1,229 @@
1#!/usr/local/bin/perl
2
3# At some point it became apparent that the original SSLeay RC4
4# assembler implementation performs suboptimaly on latest IA-32
5# microarchitectures. After re-tuning performance has changed as
6# following:
7#
8# Pentium +0%
9# Pentium III +17%
10# AMD +52%(*)
11# P4 +180%(**)
12#
13# (*) This number is actually a trade-off:-) It's possible to
14# achieve +72%, but at the cost of -48% off PIII performance.
15# In other words code performing further 13% faster on AMD
16# would perform almost 2 times slower on Intel PIII...
17# For reference! This code delivers ~80% of rc4-amd64.pl
18# performance on the same Opteron machine.
19# (**) This number requires compressed key schedule set up by
20# RC4_set_key and therefore doesn't apply to 0.9.7 [option for
21# compressed key schedule is implemented in 0.9.8 and later,
22# see commentary section in rc4_skey.c for further details].
23#
24# <appro@fy.chalmers.se>
25
26push(@INC,"perlasm","../../perlasm");
27require "x86asm.pl";
28
29&asm_init($ARGV[0],"rc4-586.pl");
30
31$x="eax";
32$y="ebx";
33$tx="ecx";
34$ty="edx";
35$in="esi";
36$out="edi";
37$d="ebp";
38
39&RC4("RC4");
40
41&asm_finish();
42
43sub RC4_loop
44 {
45 local($n,$p,$char)=@_;
46
47 &comment("Round $n");
48
49 if ($char)
50 {
51 if ($p >= 0)
52 {
53 &mov($ty, &swtmp(2));
54 &cmp($ty, $in);
55 &jbe(&label("finished"));
56 &inc($in);
57 }
58 else
59 {
60 &add($ty, 8);
61 &inc($in);
62 &cmp($ty, $in);
63 &jb(&label("finished"));
64 &mov(&swtmp(2), $ty);
65 }
66 }
67 # Moved out
68 # &mov( $tx, &DWP(0,$d,$x,4)) if $p < 0;
69
70 &add( &LB($y), &LB($tx));
71 &mov( $ty, &DWP(0,$d,$y,4));
72 # XXX
73 &mov( &DWP(0,$d,$x,4),$ty);
74 &add( $ty, $tx);
75 &mov( &DWP(0,$d,$y,4),$tx);
76 &and( $ty, 0xff);
77 &inc( &LB($x)); # NEXT ROUND
78 &mov( $tx, &DWP(0,$d,$x,4)) if $p < 1; # NEXT ROUND
79 &mov( $ty, &DWP(0,$d,$ty,4));
80
81 if (!$char)
82 {
83 #moved up into last round
84 if ($p >= 1)
85 {
86 &add( $out, 8)
87 }
88 &movb( &BP($n,"esp","",0), &LB($ty));
89 }
90 else
91 {
92 # Note in+=8 has occured
93 &movb( &HB($ty), &BP(-1,$in,"",0));
94 # XXX
95 &xorb(&LB($ty), &HB($ty));
96 # XXX
97 &movb(&BP($n,$out,"",0),&LB($ty));
98 }
99 }
100
101
102sub RC4
103 {
104 local($name)=@_;
105
106 &function_begin_B($name,"");
107
108 &mov($ty,&wparam(1)); # len
109 &cmp($ty,0);
110 &jne(&label("proceed"));
111 &ret();
112 &set_label("proceed");
113
114 &comment("");
115
116 &push("ebp");
117 &push("ebx");
118 &push("esi");
119 &xor( $x, $x); # avoid partial register stalls
120 &push("edi");
121 &xor( $y, $y); # avoid partial register stalls
122 &mov( $d, &wparam(0)); # key
123 &mov( $in, &wparam(2));
124
125 &movb( &LB($x), &BP(0,$d,"",1));
126 &movb( &LB($y), &BP(4,$d,"",1));
127
128 &mov( $out, &wparam(3));
129 &inc( &LB($x));
130
131 &stack_push(3); # 3 temp variables
132 &add( $d, 8);
133
134 # detect compressed schedule, see commentary section in rc4_skey.c...
135 # in 0.9.7 context ~50 bytes below RC4_CHAR label remain redundant,
136 # as compressed key schedule is set up in 0.9.8 and later.
137 &cmp(&DWP(256,$d),-1);
138 &je(&label("RC4_CHAR"));
139
140 &lea( $ty, &DWP(-8,$ty,$in));
141
142 # check for 0 length input
143
144 &mov( &swtmp(2), $ty); # this is now address to exit at
145 &mov( $tx, &DWP(0,$d,$x,4));
146
147 &cmp( $ty, $in);
148 &jb( &label("end")); # less than 8 bytes
149
150 &set_label("start");
151
152 # filling DELAY SLOT
153 &add( $in, 8);
154
155 &RC4_loop(0,-1,0);
156 &RC4_loop(1,0,0);
157 &RC4_loop(2,0,0);
158 &RC4_loop(3,0,0);
159 &RC4_loop(4,0,0);
160 &RC4_loop(5,0,0);
161 &RC4_loop(6,0,0);
162 &RC4_loop(7,1,0);
163
164 &comment("apply the cipher text");
165 # xor the cipher data with input
166
167 #&add( $out, 8); #moved up into last round
168
169 &mov( $tx, &swtmp(0));
170 &mov( $ty, &DWP(-8,$in,"",0));
171 &xor( $tx, $ty);
172 &mov( $ty, &DWP(-4,$in,"",0));
173 &mov( &DWP(-8,$out,"",0), $tx);
174 &mov( $tx, &swtmp(1));
175 &xor( $tx, $ty);
176 &mov( $ty, &swtmp(2)); # load end ptr;
177 &mov( &DWP(-4,$out,"",0), $tx);
178 &mov( $tx, &DWP(0,$d,$x,4));
179 &cmp($in, $ty);
180 &jbe(&label("start"));
181
182 &set_label("end");
183
184 # There is quite a bit of extra crap in RC4_loop() for this
185 # first round
186 &RC4_loop(0,-1,1);
187 &RC4_loop(1,0,1);
188 &RC4_loop(2,0,1);
189 &RC4_loop(3,0,1);
190 &RC4_loop(4,0,1);
191 &RC4_loop(5,0,1);
192 &RC4_loop(6,1,1);
193
194 &jmp(&label("finished"));
195
196 &align(16);
197 # this is essentially Intel P4 specific codepath, see rc4_skey.c,
198 # and is engaged in 0.9.8 and later context...
199 &set_label("RC4_CHAR");
200
201 &lea ($ty,&DWP(0,$in,$ty));
202 &mov (&swtmp(2),$ty);
203
204 # strangely enough unrolled loop performs over 20% slower...
205 &set_label("RC4_CHAR_loop");
206 &movz ($tx,&BP(0,$d,$x));
207 &add (&LB($y),&LB($tx));
208 &movz ($ty,&BP(0,$d,$y));
209 &movb (&BP(0,$d,$y),&LB($tx));
210 &movb (&BP(0,$d,$x),&LB($ty));
211 &add (&LB($ty),&LB($tx));
212 &movz ($ty,&BP(0,$d,$ty));
213 &xorb (&LB($ty),&BP(0,$in));
214 &movb (&BP(0,$out),&LB($ty));
215 &inc (&LB($x));
216 &inc ($in);
217 &inc ($out);
218 &cmp ($in,&swtmp(2));
219 &jb (&label("RC4_CHAR_loop"));
220
221 &set_label("finished");
222 &dec( $x);
223 &stack_pop(3);
224 &movb( &BP(-4,$d,"",0),&LB($y));
225 &movb( &BP(-8,$d,"",0),&LB($x));
226
227 &function_end($name);
228 }
229
diff --git a/src/lib/libcrypto/rc4/asm/rc4-amd64.pl b/src/lib/libcrypto/rc4/asm/rc4-amd64.pl
new file mode 100755
index 0000000000..9e0da8af99
--- /dev/null
+++ b/src/lib/libcrypto/rc4/asm/rc4-amd64.pl
@@ -0,0 +1,227 @@
1#!/usr/bin/env perl
2#
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# ====================================================================
8#
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 $255,%r8'
17# I simply 'inc %r8b'. Even though optimization manual discourages
18# to operate on partial registers, it turned out to be the best bet.
19# At least for AMD... How IA32E would perform remains to be seen...
20
21# As was shown by Marc Bevand reordering of couple of load operations
22# results in even higher performance gain of 3.3x:-) At least on
23# Opteron... For reference, 1x in this case is RC4_CHAR C-code
24# compiled with gcc 3.3.2, which performs at ~54MBps per 1GHz clock.
25# Latter means that if you want to *estimate* what to expect from
26# *your* CPU, then multiply 54 by 3.3 and clock frequency in GHz.
27
28# Intel P4 EM64T core was found to run the AMD64 code really slow...
29# The only way to achieve comparable performance on P4 is to keep
30# RC4_CHAR. Kind of ironic, huh? As it's apparently impossible to
31# compose blended code, which would perform even within 30% marginal
32# on either AMD and Intel platforms, I implement both cases. See
33# rc4_skey.c for further details... This applies to 0.9.8 and later.
34# In 0.9.7 context RC4_CHAR codepath is never engaged and ~70 bytes
35# of code remain redundant.
36
37$output=shift;
38
39$win64a=1 if ($output =~ /win64a.[s|asm]/);
40
41open STDOUT,">$output" || die "can't open $output: $!";
42
43if (defined($win64a)) {
44 $dat="%rcx"; # arg1
45 $len="%rdx"; # arg2
46 $inp="%rsi"; # r8, arg3 moves here
47 $out="%rdi"; # r9, arg4 moves here
48} else {
49 $dat="%rdi"; # arg1
50 $len="%rsi"; # arg2
51 $inp="%rdx"; # arg3
52 $out="%rcx"; # arg4
53}
54
55$XX="%r10";
56$TX="%r8";
57$YY="%r11";
58$TY="%r9";
59
60sub PTR() {
61 my $ret=shift;
62 if (defined($win64a)) {
63 $ret =~ s/\[([\S]+)\+([\S]+)\]/[$2+$1]/g; # [%rN+%rM*4]->[%rM*4+%rN]
64 $ret =~ s/:([^\[]+)\[([^\]]+)\]/:[$2+$1]/g; # :off[ea]->:[ea+off]
65 } else {
66 $ret =~ s/[\+\*]/,/g; # [%rN+%rM*4]->[%rN,%rM,4]
67 $ret =~ s/\[([^\]]+)\]/($1)/g; # [%rN]->(%rN)
68 }
69 $ret;
70}
71
72$code=<<___ if (!defined($win64a));
73.text
74
75.globl RC4
76.type RC4,\@function
77.align 16
78RC4: or $len,$len
79 jne .Lentry
80 repret
81.Lentry:
82___
83$code=<<___ if (defined($win64a));
84_TEXT SEGMENT
85PUBLIC RC4
86ALIGN 16
87RC4 PROC
88 or $len,$len
89 jne .Lentry
90 repret
91.Lentry:
92 push %rdi
93 push %rsi
94 sub \$40,%rsp
95 mov %r8,$inp
96 mov %r9,$out
97___
98$code.=<<___;
99 add \$8,$dat
100 movl `&PTR("DWORD:-8[$dat]")`,$XX#d
101 movl `&PTR("DWORD:-4[$dat]")`,$YY#d
102 cmpl \$-1,`&PTR("DWORD:256[$dat]")`
103 je .LRC4_CHAR
104 test \$-8,$len
105 jz .Lloop1
106.align 16
107.Lloop8:
108 inc $XX#b
109 movl `&PTR("DWORD:[$dat+$XX*4]")`,$TX#d
110 add $TX#b,$YY#b
111 movl `&PTR("DWORD:[$dat+$YY*4]")`,$TY#d
112 movl $TX#d,`&PTR("DWORD:[$dat+$YY*4]")`
113 movl $TY#d,`&PTR("DWORD:[$dat+$XX*4]")`
114 add $TX#b,$TY#b
115 inc $XX#b
116 movl `&PTR("DWORD:[$dat+$XX*4]")`,$TX#d
117 movb `&PTR("BYTE:[$dat+$TY*4]")`,%al
118___
119for ($i=1;$i<=6;$i++) {
120$code.=<<___;
121 add $TX#b,$YY#b
122 ror \$8,%rax
123 movl `&PTR("DWORD:[$dat+$YY*4]")`,$TY#d
124 movl $TX#d,`&PTR("DWORD:[$dat+$YY*4]")`
125 movl $TY#d,`&PTR("DWORD:[$dat+$XX*4]")`
126 add $TX#b,$TY#b
127 inc $XX#b
128 movl `&PTR("DWORD:[$dat+$XX*4]")`,$TX#d
129 movb `&PTR("BYTE:[$dat+$TY*4]")`,%al
130___
131}
132$code.=<<___;
133 add $TX#b,$YY#b
134 ror \$8,%rax
135 movl `&PTR("DWORD:[$dat+$YY*4]")`,$TY#d
136 movl $TX#d,`&PTR("DWORD:[$dat+$YY*4]")`
137 movl $TY#d,`&PTR("DWORD:[$dat+$XX*4]")`
138 sub \$8,$len
139 add $TY#b,$TX#b
140 movb `&PTR("BYTE:[$dat+$TX*4]")`,%al
141 ror \$8,%rax
142 add \$8,$inp
143 add \$8,$out
144
145 xor `&PTR("QWORD:-8[$inp]")`,%rax
146 mov %rax,`&PTR("QWORD:-8[$out]")`
147
148 test \$-8,$len
149 jnz .Lloop8
150 cmp \$0,$len
151 jne .Lloop1
152.Lexit:
153 movl $XX#d,`&PTR("DWORD:-8[$dat]")`
154 movl $YY#d,`&PTR("DWORD:-4[$dat]")`
155___
156$code.=<<___ if (defined($win64a));
157 add \$40,%rsp
158 pop %rsi
159 pop %rdi
160___
161$code.=<<___;
162 repret
163.align 16
164.Lloop1:
165 movzb `&PTR("BYTE:[$inp]")`,%eax
166 inc $XX#b
167 movl `&PTR("DWORD:[$dat+$XX*4]")`,$TX#d
168 add $TX#b,$YY#b
169 movl `&PTR("DWORD:[$dat+$YY*4]")`,$TY#d
170 movl $TX#d,`&PTR("DWORD:[$dat+$YY*4]")`
171 movl $TY#d,`&PTR("DWORD:[$dat+$XX*4]")`
172 add $TY#b,$TX#b
173 movl `&PTR("DWORD:[$dat+$TX*4]")`,$TY#d
174 xor $TY,%rax
175 inc $inp
176 movb %al,`&PTR("BYTE:[$out]")`
177 inc $out
178 dec $len
179 jnz .Lloop1
180 jmp .Lexit
181
182.align 16
183.LRC4_CHAR:
184 inc $XX#b
185 movzb `&PTR("BYTE:[$dat+$XX]")`,$TX#d
186 add $TX#b,$YY#b
187 movzb `&PTR("BYTE:[$dat+$YY]")`,$TY#d
188 movb $TX#b,`&PTR("BYTE:[$dat+$YY]")`
189 movb $TY#b,`&PTR("BYTE:[$dat+$XX]")`
190 add $TX#b,$TY#b
191 movzb `&PTR("BYTE:[$dat+$TY]")`,$TY#d
192 xorb `&PTR("BYTE:[$inp]")`,$TY#b
193 movb $TY#b,`&PTR("BYTE:[$out]")`
194 inc $inp
195 inc $out
196 dec $len
197 jnz .LRC4_CHAR
198 jmp .Lexit
199___
200$code.=<<___ if (defined($win64a));
201RC4 ENDP
202_TEXT ENDS
203END
204___
205$code.=<<___ if (!defined($win64a));
206.size RC4,.-RC4
207___
208
209$code =~ s/#([bwd])/$1/gm;
210$code =~ s/\`([^\`]*)\`/eval $1/gem;
211
212if (defined($win64a)) {
213 $code =~ s/\.align/ALIGN/gm;
214 $code =~ s/[\$%]//gm;
215 $code =~ s/\.L/\$L/gm;
216 $code =~ s/([\w]+)([\s]+)([\S]+),([\S]+)/$1$2$4,$3/gm;
217 $code =~ s/([QD]*WORD|BYTE):/$1 PTR/gm;
218 $code =~ s/mov[bwlq]/mov/gm;
219 $code =~ s/movzb/movzx/gm;
220 $code =~ s/repret/DB\t0F3h,0C3h/gm;
221 $code =~ s/cmpl/cmp/gm;
222 $code =~ s/xorb/xor/gm;
223} else {
224 $code =~ s/([QD]*WORD|BYTE)://gm;
225 $code =~ s/repret/.byte\t0xF3,0xC3/gm;
226}
227print $code;
generated by cgit v1.2.3-55-g6feb (git 2.39.1) at 2025-08-22 04:06:21 +0000