rc4-586.pl revision 194206 
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),$ty);
203	&movz	($tx,&BP(0,$d,$x));
204
205	# strangely enough unrolled loop performs over 20% slower...
206	&set_label("RC4_CHAR_loop");
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		&add	(&LB($x),1);
214		&xorb	(&LB($ty),&BP(0,$in));
215		&lea	($in,&DWP(1,$in));
216		&movz	($tx,&BP(0,$d,$x));
217		&cmp	($in,&swtmp(2));
218		&movb	(&BP(0,$out),&LB($ty));
219		&lea	($out,&DWP(1,$out));
220	&jb	(&label("RC4_CHAR_loop"));
221
222	&set_label("finished");
223	&dec(	$x);
224	 &stack_pop(3);
225	&movb(	&BP(-4,$d,"",0),&LB($y));
226	 &movb(	&BP(-8,$d,"",0),&LB($x));
227
228	&function_end($name);
229	}
230
231