1#!/usr/bin/env perl
2
3# ====================================================================
4# [Re]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# ====================================================================
9
10# At some point it became apparent that the original SSLeay RC4
11# assembler implementation performs suboptimally on latest IA-32
12# microarchitectures. After re-tuning performance has changed as
13# following:
14#
15# Pentium	-10%
16# Pentium III	+12%
17# AMD		+50%(*)
18# P4		+250%(**)
19#
20# (*)	This number is actually a trade-off:-) It's possible to
21#	achieve	+72%, but at the cost of -48% off PIII performance.
22#	In other words code performing further 13% faster on AMD
23#	would perform almost 2 times slower on Intel PIII...
24#	For reference! This code delivers ~80% of rc4-amd64.pl
25#	performance on the same Opteron machine.
26# (**)	This number requires compressed key schedule set up by
27#	RC4_set_key [see commentary below for further details].
28#
29#					<appro@fy.chalmers.se>
30
31# May 2011
32#
33# Optimize for Core2 and Westmere [and incidentally Opteron]. Current
34# performance in cycles per processed byte (less is better) and
35# improvement relative to previous version of this module is:
36#
37# Pentium	10.2			# original numbers
38# Pentium III	7.8(*)
39# Intel P4	7.5
40#
41# Opteron	6.1/+20%		# new MMX numbers
42# Core2		5.3/+67%(**)
43# Westmere	5.1/+94%(**)
44# Sandy Bridge	5.0/+8%
45# Atom		12.6/+6%
46#
47# (*)	PIII can actually deliver 6.6 cycles per byte with MMX code,
48#	but this specific code performs poorly on Core2. And vice
49#	versa, below MMX/SSE code delivering 5.8/7.1 on Core2 performs
50#	poorly on PIII, at 8.0/14.5:-( As PIII is not a "hot" CPU
51#	[anymore], I chose to discard PIII-specific code path and opt
52#	for original IALU-only code, which is why MMX/SSE code path
53#	is guarded by SSE2 bit (see below), not MMX/SSE.
54# (**)	Performance vs. block size on Core2 and Westmere had a maximum
55#	at ... 64 bytes block size. And it was quite a maximum, 40-60%
56#	in comparison to largest 8KB block size. Above improvement
57#	coefficients are for the largest block size.
58
59$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
60push(@INC,"${dir}","${dir}../../perlasm");
61require "x86asm.pl";
62
63&asm_init($ARGV[0],"rc4-586.pl",$x86only = $ARGV[$#ARGV] eq "386");
64
65$xx="eax";
66$yy="ebx";
67$tx="ecx";
68$ty="edx";
69$inp="esi";
70$out="ebp";
71$dat="edi";
72
73sub RC4_loop {
74  my $i=shift;
75  my $func = ($i==0)?*mov:*or;
76
77	&add	(&LB($yy),&LB($tx));
78	&mov	($ty,&DWP(0,$dat,$yy,4));
79	&mov	(&DWP(0,$dat,$yy,4),$tx);
80	&mov	(&DWP(0,$dat,$xx,4),$ty);
81	&add	($ty,$tx);
82	&inc	(&LB($xx));
83	&and	($ty,0xff);
84	&ror	($out,8)	if ($i!=0);
85	if ($i<3) {
86	  &mov	($tx,&DWP(0,$dat,$xx,4));
87	} else {
88	  &mov	($tx,&wparam(3));	# reload [re-biased] out
89	}
90	&$func	($out,&DWP(0,$dat,$ty,4));
91}
92
93if ($alt=0) {
94  # >20% faster on Atom and Sandy Bridge[!], 8% faster on Opteron,
95  # but ~40% slower on Core2 and Westmere... Attempt to add movz
96  # brings down Opteron by 25%, Atom and Sandy Bridge by 15%, yet
97  # on Core2 with movz it's almost 20% slower than below alternative
98  # code... Yes, it's a total mess...
99  my @XX=($xx,$out);
100  $RC4_loop_mmx = sub {		# SSE actually...
101    my $i=shift;
102    my $j=$i<=0?0:$i>>1;
103    my $mm=$i<=0?"mm0":"mm".($i&1);
104
105	&add	(&LB($yy),&LB($tx));
106	&lea	(@XX[1],&DWP(1,@XX[0]));
107	&pxor	("mm2","mm0")				if ($i==0);
108	&psllq	("mm1",8)				if ($i==0);
109	&and	(@XX[1],0xff);
110	&pxor	("mm0","mm0")				if ($i<=0);
111	&mov	($ty,&DWP(0,$dat,$yy,4));
112	&mov	(&DWP(0,$dat,$yy,4),$tx);
113	&pxor	("mm1","mm2")				if ($i==0);
114	&mov	(&DWP(0,$dat,$XX[0],4),$ty);
115	&add	(&LB($ty),&LB($tx));
116	&movd	(@XX[0],"mm7")				if ($i==0);
117	&mov	($tx,&DWP(0,$dat,@XX[1],4));
118	&pxor	("mm1","mm1")				if ($i==1);
119	&movq	("mm2",&QWP(0,$inp))			if ($i==1);
120	&movq	(&QWP(-8,(@XX[0],$inp)),"mm1")		if ($i==0);
121	&pinsrw	($mm,&DWP(0,$dat,$ty,4),$j);
122
123	push	(@XX,shift(@XX))			if ($i>=0);
124  }
125} else {
126  # Using pinsrw here improves performane on Intel CPUs by 2-3%, but
127  # brings down AMD by 7%...
128  $RC4_loop_mmx = sub {
129    my $i=shift;
130
131	&add	(&LB($yy),&LB($tx));
132	&psllq	("mm1",8*(($i-1)&7))			if (abs($i)!=1);
133	&mov	($ty,&DWP(0,$dat,$yy,4));
134	&mov	(&DWP(0,$dat,$yy,4),$tx);
135	&mov	(&DWP(0,$dat,$xx,4),$ty);
136	&inc	($xx);
137	&add	($ty,$tx);
138	&movz	($xx,&LB($xx));				# (*)
139	&movz	($ty,&LB($ty));				# (*)
140	&pxor	("mm2",$i==1?"mm0":"mm1")		if ($i>=0);
141	&movq	("mm0",&QWP(0,$inp))			if ($i<=0);
142	&movq	(&QWP(-8,($out,$inp)),"mm2")		if ($i==0);
143	&mov	($tx,&DWP(0,$dat,$xx,4));
144	&movd	($i>0?"mm1":"mm2",&DWP(0,$dat,$ty,4));
145
146	# (*)	This is the key to Core2 and Westmere performance.
147	#	Whithout movz out-of-order execution logic confuses
148	#	itself and fails to reorder loads and stores. Problem
149	#	appears to be fixed in Sandy Bridge...
150  }
151}
152
153&external_label("OPENSSL_ia32cap_P");
154
155# void RC4(RC4_KEY *key,size_t len,const unsigned char *inp,unsigned char *out);
156&function_begin("RC4");
157	&mov	($dat,&wparam(0));	# load key schedule pointer
158	&mov	($ty, &wparam(1));	# load len
159	&mov	($inp,&wparam(2));	# load inp
160	&mov	($out,&wparam(3));	# load out
161
162	&xor	($xx,$xx);		# avoid partial register stalls
163	&xor	($yy,$yy);
164
165	&cmp	($ty,0);		# safety net
166	&je	(&label("abort"));
167
168	&mov	(&LB($xx),&BP(0,$dat));	# load key->x
169	&mov	(&LB($yy),&BP(4,$dat));	# load key->y
170	&add	($dat,8);
171
172	&lea	($tx,&DWP(0,$inp,$ty));
173	&sub	($out,$inp);		# re-bias out
174	&mov	(&wparam(1),$tx);	# save input+len
175
176	&inc	(&LB($xx));
177
178	# detect compressed key schedule...
179	&cmp	(&DWP(256,$dat),-1);
180	&je	(&label("RC4_CHAR"));
181
182	&mov	($tx,&DWP(0,$dat,$xx,4));
183
184	&and	($ty,-4);		# how many 4-byte chunks?
185	&jz	(&label("loop1"));
186
187	&mov	(&wparam(3),$out);	# $out as accumulator in these loops
188					if ($x86only) {
189	&jmp	(&label("go4loop4"));
190					} else {
191	&test	($ty,-8);
192	&jz	(&label("go4loop4"));
193
194	&picmeup($out,"OPENSSL_ia32cap_P");
195	&bt	(&DWP(0,$out),26);	# check SSE2 bit [could have been MMX]
196	&jnc	(&label("go4loop4"));
197
198	&mov	($out,&wparam(3))	if (!$alt);
199	&movd	("mm7",&wparam(3))	if ($alt);
200	&and	($ty,-8);
201	&lea	($ty,&DWP(-8,$inp,$ty));
202	&mov	(&DWP(-4,$dat),$ty);	# save input+(len/8)*8-8
203
204	&$RC4_loop_mmx(-1);
205	&jmp(&label("loop_mmx_enter"));
206
207	&set_label("loop_mmx",16);
208		&$RC4_loop_mmx(0);
209	&set_label("loop_mmx_enter");
210		for 	($i=1;$i<8;$i++) { &$RC4_loop_mmx($i); }
211		&mov	($ty,$yy);
212		&xor	($yy,$yy);		# this is second key to Core2
213		&mov	(&LB($yy),&LB($ty));	# and Westmere performance...
214		&cmp	($inp,&DWP(-4,$dat));
215		&lea	($inp,&DWP(8,$inp));
216	&jb	(&label("loop_mmx"));
217
218    if ($alt) {
219	&movd	($out,"mm7");
220	&pxor	("mm2","mm0");
221	&psllq	("mm1",8);
222	&pxor	("mm1","mm2");
223	&movq	(&QWP(-8,$out,$inp),"mm1");
224    } else {
225	&psllq	("mm1",56);
226	&pxor	("mm2","mm1");
227	&movq	(&QWP(-8,$out,$inp),"mm2");
228    }
229	&emms	();
230
231	&cmp	($inp,&wparam(1));	# compare to input+len
232	&je	(&label("done"));
233	&jmp	(&label("loop1"));
234					}
235
236&set_label("go4loop4",16);
237	&lea	($ty,&DWP(-4,$inp,$ty));
238	&mov	(&wparam(2),$ty);	# save input+(len/4)*4-4
239
240	&set_label("loop4");
241		for ($i=0;$i<4;$i++) { RC4_loop($i); }
242		&ror	($out,8);
243		&xor	($out,&DWP(0,$inp));
244		&cmp	($inp,&wparam(2));	# compare to input+(len/4)*4-4
245		&mov	(&DWP(0,$tx,$inp),$out);# $tx holds re-biased out here
246		&lea	($inp,&DWP(4,$inp));
247		&mov	($tx,&DWP(0,$dat,$xx,4));
248	&jb	(&label("loop4"));
249
250	&cmp	($inp,&wparam(1));	# compare to input+len
251	&je	(&label("done"));
252	&mov	($out,&wparam(3));	# restore $out
253
254	&set_label("loop1",16);
255		&add	(&LB($yy),&LB($tx));
256		&mov	($ty,&DWP(0,$dat,$yy,4));
257		&mov	(&DWP(0,$dat,$yy,4),$tx);
258		&mov	(&DWP(0,$dat,$xx,4),$ty);
259		&add	($ty,$tx);
260		&inc	(&LB($xx));
261		&and	($ty,0xff);
262		&mov	($ty,&DWP(0,$dat,$ty,4));
263		&xor	(&LB($ty),&BP(0,$inp));
264		&lea	($inp,&DWP(1,$inp));
265		&mov	($tx,&DWP(0,$dat,$xx,4));
266		&cmp	($inp,&wparam(1));	# compare to input+len
267		&mov	(&BP(-1,$out,$inp),&LB($ty));
268	&jb	(&label("loop1"));
269
270	&jmp	(&label("done"));
271
272# this is essentially Intel P4 specific codepath...
273&set_label("RC4_CHAR",16);
274	&movz	($tx,&BP(0,$dat,$xx));
275	# strangely enough unrolled loop performs over 20% slower...
276	&set_label("cloop1");
277		&add	(&LB($yy),&LB($tx));
278		&movz	($ty,&BP(0,$dat,$yy));
279		&mov	(&BP(0,$dat,$yy),&LB($tx));
280		&mov	(&BP(0,$dat,$xx),&LB($ty));
281		&add	(&LB($ty),&LB($tx));
282		&movz	($ty,&BP(0,$dat,$ty));
283		&add	(&LB($xx),1);
284		&xor	(&LB($ty),&BP(0,$inp));
285		&lea	($inp,&DWP(1,$inp));
286		&movz	($tx,&BP(0,$dat,$xx));
287		&cmp	($inp,&wparam(1));
288		&mov	(&BP(-1,$out,$inp),&LB($ty));
289	&jb	(&label("cloop1"));
290
291&set_label("done");
292	&dec	(&LB($xx));
293	&mov	(&DWP(-4,$dat),$yy);		# save key->y
294	&mov	(&BP(-8,$dat),&LB($xx));	# save key->x
295&set_label("abort");
296&function_end("RC4");
297
298########################################################################
299
300$inp="esi";
301$out="edi";
302$idi="ebp";
303$ido="ecx";
304$idx="edx";
305
306# void RC4_set_key(RC4_KEY *key,int len,const unsigned char *data);
307&function_begin("private_RC4_set_key");
308	&mov	($out,&wparam(0));		# load key
309	&mov	($idi,&wparam(1));		# load len
310	&mov	($inp,&wparam(2));		# load data
311	&picmeup($idx,"OPENSSL_ia32cap_P");
312
313	&lea	($out,&DWP(2*4,$out));		# &key->data
314	&lea	($inp,&DWP(0,$inp,$idi));	# $inp to point at the end
315	&neg	($idi);
316	&xor	("eax","eax");
317	&mov	(&DWP(-4,$out),$idi);		# borrow key->y
318
319	&bt	(&DWP(0,$idx),20);		# check for bit#20
320	&jc	(&label("c1stloop"));
321
322&set_label("w1stloop",16);
323	&mov	(&DWP(0,$out,"eax",4),"eax");	# key->data[i]=i;
324	&add	(&LB("eax"),1);			# i++;
325	&jnc	(&label("w1stloop"));
326
327	&xor	($ido,$ido);
328	&xor	($idx,$idx);
329
330&set_label("w2ndloop",16);
331	&mov	("eax",&DWP(0,$out,$ido,4));
332	&add	(&LB($idx),&BP(0,$inp,$idi));
333	&add	(&LB($idx),&LB("eax"));
334	&add	($idi,1);
335	&mov	("ebx",&DWP(0,$out,$idx,4));
336	&jnz	(&label("wnowrap"));
337	  &mov	($idi,&DWP(-4,$out));
338	&set_label("wnowrap");
339	&mov	(&DWP(0,$out,$idx,4),"eax");
340	&mov	(&DWP(0,$out,$ido,4),"ebx");
341	&add	(&LB($ido),1);
342	&jnc	(&label("w2ndloop"));
343&jmp	(&label("exit"));
344
345# Unlike all other x86 [and x86_64] implementations, Intel P4 core
346# [including EM64T] was found to perform poorly with above "32-bit" key
347# schedule, a.k.a. RC4_INT. Performance improvement for IA-32 hand-coded
348# assembler turned out to be 3.5x if re-coded for compressed 8-bit one,
349# a.k.a. RC4_CHAR! It's however inappropriate to just switch to 8-bit
350# schedule for x86[_64], because non-P4 implementations suffer from
351# significant performance losses then, e.g. PIII exhibits >2x
352# deterioration, and so does Opteron. In order to assure optimal
353# all-round performance, we detect P4 at run-time and set up compressed
354# key schedule, which is recognized by RC4 procedure.
355
356&set_label("c1stloop",16);
357	&mov	(&BP(0,$out,"eax"),&LB("eax"));	# key->data[i]=i;
358	&add	(&LB("eax"),1);			# i++;
359	&jnc	(&label("c1stloop"));
360
361	&xor	($ido,$ido);
362	&xor	($idx,$idx);
363	&xor	("ebx","ebx");
364
365&set_label("c2ndloop",16);
366	&mov	(&LB("eax"),&BP(0,$out,$ido));
367	&add	(&LB($idx),&BP(0,$inp,$idi));
368	&add	(&LB($idx),&LB("eax"));
369	&add	($idi,1);
370	&mov	(&LB("ebx"),&BP(0,$out,$idx));
371	&jnz	(&label("cnowrap"));
372	  &mov	($idi,&DWP(-4,$out));
373	&set_label("cnowrap");
374	&mov	(&BP(0,$out,$idx),&LB("eax"));
375	&mov	(&BP(0,$out,$ido),&LB("ebx"));
376	&add	(&LB($ido),1);
377	&jnc	(&label("c2ndloop"));
378
379	&mov	(&DWP(256,$out),-1);		# mark schedule as compressed
380
381&set_label("exit");
382	&xor	("eax","eax");
383	&mov	(&DWP(-8,$out),"eax");		# key->x=0;
384	&mov	(&DWP(-4,$out),"eax");		# key->y=0;
385&function_end("private_RC4_set_key");
386
387# const char *RC4_options(void);
388&function_begin_B("RC4_options");
389	&call	(&label("pic_point"));
390&set_label("pic_point");
391	&blindpop("eax");
392	&lea	("eax",&DWP(&label("opts")."-".&label("pic_point"),"eax"));
393	&picmeup("edx","OPENSSL_ia32cap_P");
394	&mov	("edx",&DWP(0,"edx"));
395	&bt	("edx",20);
396	&jc	(&label("1xchar"));
397	&bt	("edx",26);
398	&jnc	(&label("ret"));
399	&add	("eax",25);
400	&ret	();
401&set_label("1xchar");
402	&add	("eax",12);
403&set_label("ret");
404	&ret	();
405&set_label("opts",64);
406&asciz	("rc4(4x,int)");
407&asciz	("rc4(1x,char)");
408&asciz	("rc4(8x,mmx)");
409&asciz	("RC4 for x86, CRYPTOGAMS by <appro\@openssl.org>");
410&align	(64);
411&function_end_B("RC4_options");
412
413&asm_finish();
414
415