xref: /freebsd/crypto/openssl/crypto/des/asm/des-586.pl (revision 780fb4a2)
1#!/usr/local/bin/perl
2#
3# The inner loop instruction sequence and the IP/FP modifications are from
4# Svend Olaf Mikkelsen <svolaf@inet.uni-c.dk>
5#
6
7$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
8push(@INC,"${dir}","${dir}../../perlasm");
9require "x86asm.pl";
10require "cbc.pl";
11require "desboth.pl";
12
13# base code is in microsft
14# op dest, source
15# format.
16#
17
18&asm_init($ARGV[0],"des-586.pl");
19
20$L="edi";
21$R="esi";
22$trans="ebp";
23$small_footprint=1 if (grep(/\-DOPENSSL_SMALL_FOOTPRINT/,@ARGV));
24# one can discuss setting this variable to 1 unconditionally, as
25# the folded loop is only 3% slower than unrolled, but >7 times smaller
26
27&public_label("DES_SPtrans");
28&static_label("des_sptrans");
29
30&DES_encrypt_internal();
31&DES_decrypt_internal();
32&DES_encrypt("DES_encrypt1",1);
33&DES_encrypt("DES_encrypt2",0);
34&DES_encrypt3("DES_encrypt3",1);
35&DES_encrypt3("DES_decrypt3",0);
36&cbc("DES_ncbc_encrypt","DES_encrypt1","DES_encrypt1",0,4,5,3,5,-1);
37&cbc("DES_ede3_cbc_encrypt","DES_encrypt3","DES_decrypt3",0,6,7,3,4,5);
38&DES_SPtrans();
39
40&asm_finish();
41
42sub DES_encrypt_internal()
43	{
44	&function_begin_B("_x86_DES_encrypt");
45
46	if ($small_footprint)
47	    {
48	    &lea("edx",&DWP(128,"ecx"));
49	    &push("edx");
50	    &push("ecx");
51	    &set_label("eloop");
52		&D_ENCRYPT(0,$L,$R,0,$trans,"eax","ebx","ecx","edx",&swtmp(0));
53		&comment("");
54		&D_ENCRYPT(1,$R,$L,2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
55		&comment("");
56		&add("ecx",16);
57		&cmp("ecx",&swtmp(1));
58		&mov(&swtmp(0),"ecx");
59		&jb(&label("eloop"));
60	    &add("esp",8);
61	    }
62	else
63	    {
64	    &push("ecx");
65	    for ($i=0; $i<16; $i+=2)
66		{
67		&comment("Round $i");
68		&D_ENCRYPT($i,$L,$R,$i*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
69		&comment("Round ".sprintf("%d",$i+1));
70		&D_ENCRYPT($i+1,$R,$L,($i+1)*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
71		}
72	    &add("esp",4);
73	}
74	&ret();
75
76	&function_end_B("_x86_DES_encrypt");
77	}
78
79sub DES_decrypt_internal()
80	{
81	&function_begin_B("_x86_DES_decrypt");
82
83	if ($small_footprint)
84	    {
85	    &push("ecx");
86	    &lea("ecx",&DWP(128,"ecx"));
87	    &push("ecx");
88	    &set_label("dloop");
89		&D_ENCRYPT(0,$L,$R,-2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
90		&comment("");
91		&D_ENCRYPT(1,$R,$L,-4,$trans,"eax","ebx","ecx","edx",&swtmp(0));
92		&comment("");
93		&sub("ecx",16);
94		&cmp("ecx",&swtmp(1));
95		&mov(&swtmp(0),"ecx");
96		&ja(&label("dloop"));
97	    &add("esp",8);
98	    }
99	else
100	    {
101	    &push("ecx");
102	    for ($i=15; $i>0; $i-=2)
103		{
104		&comment("Round $i");
105		&D_ENCRYPT(15-$i,$L,$R,$i*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
106		&comment("Round ".sprintf("%d",$i-1));
107		&D_ENCRYPT(15-$i+1,$R,$L,($i-1)*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
108		}
109	    &add("esp",4);
110	    }
111	&ret();
112
113	&function_end_B("_x86_DES_decrypt");
114	}
115
116sub DES_encrypt
117	{
118	local($name,$do_ip)=@_;
119
120	&function_begin_B($name);
121
122	&push("esi");
123	&push("edi");
124
125	&comment("");
126	&comment("Load the 2 words");
127
128	if ($do_ip)
129		{
130		&mov($R,&wparam(0));
131		 &xor(	"ecx",		"ecx"		);
132
133		&push("ebx");
134		&push("ebp");
135
136		&mov("eax",&DWP(0,$R,"",0));
137		 &mov("ebx",&wparam(2));	# get encrypt flag
138		&mov($L,&DWP(4,$R,"",0));
139		&comment("");
140		&comment("IP");
141		&IP_new("eax",$L,$R,3);
142		}
143	else
144		{
145		&mov("eax",&wparam(0));
146		 &xor(	"ecx",		"ecx"		);
147
148		&push("ebx");
149		&push("ebp");
150
151		&mov($R,&DWP(0,"eax","",0));
152		 &mov("ebx",&wparam(2));	# get encrypt flag
153		&rotl($R,3);
154		&mov($L,&DWP(4,"eax","",0));
155		&rotl($L,3);
156		}
157
158	# PIC-ification:-)
159	&call	(&label("pic_point"));
160	&set_label("pic_point");
161	&blindpop($trans);
162	&lea	($trans,&DWP(&label("des_sptrans")."-".&label("pic_point"),$trans));
163
164	&mov(	"ecx",	&wparam(1)	);
165
166	&cmp("ebx","0");
167	&je(&label("decrypt"));
168	&call("_x86_DES_encrypt");
169	&jmp(&label("done"));
170	&set_label("decrypt");
171	&call("_x86_DES_decrypt");
172	&set_label("done");
173
174	if ($do_ip)
175		{
176		&comment("");
177		&comment("FP");
178		&mov("edx",&wparam(0));
179		&FP_new($L,$R,"eax",3);
180
181		&mov(&DWP(0,"edx","",0),"eax");
182		&mov(&DWP(4,"edx","",0),$R);
183		}
184	else
185		{
186		&comment("");
187		&comment("Fixup");
188		&rotr($L,3);		# r
189		 &mov("eax",&wparam(0));
190		&rotr($R,3);		# l
191		 &mov(&DWP(0,"eax","",0),$L);
192		 &mov(&DWP(4,"eax","",0),$R);
193		}
194
195	&pop("ebp");
196	&pop("ebx");
197	&pop("edi");
198	&pop("esi");
199	&ret();
200
201	&function_end_B($name);
202	}
203
204sub D_ENCRYPT
205	{
206	local($r,$L,$R,$S,$trans,$u,$tmp1,$tmp2,$t,$wp1)=@_;
207
208	 &mov(	$u,		&DWP(&n2a($S*4),$tmp2,"",0));
209	&xor(	$tmp1,		$tmp1);
210	 &mov(	$t,		&DWP(&n2a(($S+1)*4),$tmp2,"",0));
211	&xor(	$u,		$R);
212	&xor(	$tmp2,		$tmp2);
213	 &xor(	$t,		$R);
214	&and(	$u,		"0xfcfcfcfc"	);
215	 &and(	$t,		"0xcfcfcfcf"	);
216	&movb(	&LB($tmp1),	&LB($u)	);
217	 &movb(	&LB($tmp2),	&HB($u)	);
218	&rotr(	$t,		4		);
219	&xor(	$L,		&DWP("     ",$trans,$tmp1,0));
220	 &movb(	&LB($tmp1),	&LB($t)	);
221	 &xor(	$L,		&DWP("0x200",$trans,$tmp2,0));
222	 &movb(	&LB($tmp2),	&HB($t)	);
223	&shr(	$u,		16);
224	 &xor(	$L,		&DWP("0x100",$trans,$tmp1,0));
225	 &movb(	&LB($tmp1),	&HB($u)	);
226	&shr(	$t,		16);
227	 &xor(	$L,		&DWP("0x300",$trans,$tmp2,0));
228	&movb(	&LB($tmp2),	&HB($t)	);
229	 &and(	$u,		"0xff"	);
230	&and(	$t,		"0xff"	);
231	 &xor(	$L,		&DWP("0x600",$trans,$tmp1,0));
232	 &xor(	$L,		&DWP("0x700",$trans,$tmp2,0));
233	&mov(	$tmp2,		$wp1	);
234	 &xor(	$L,		&DWP("0x400",$trans,$u,0));
235	 &xor(	$L,		&DWP("0x500",$trans,$t,0));
236	}
237
238sub n2a
239	{
240	sprintf("%d",$_[0]);
241	}
242
243# now has a side affect of rotating $a by $shift
244sub R_PERM_OP
245	{
246	local($a,$b,$tt,$shift,$mask,$last)=@_;
247
248	&rotl(	$a,		$shift		) if ($shift != 0);
249	&mov(	$tt,		$a		);
250	&xor(	$a,		$b		);
251	&and(	$a,		$mask		);
252	# This can never succeed, and besides it is difficult to see what the
253	# idea was - Ben 13 Feb 99
254	if (!$last eq $b)
255		{
256		&xor(	$b,		$a		);
257		&xor(	$tt,		$a		);
258		}
259	else
260		{
261		&xor(	$tt,		$a		);
262		&xor(	$b,		$a		);
263		}
264	&comment("");
265	}
266
267sub IP_new
268	{
269	local($l,$r,$tt,$lr)=@_;
270
271	&R_PERM_OP($l,$r,$tt, 4,"0xf0f0f0f0",$l);
272	&R_PERM_OP($r,$tt,$l,20,"0xfff0000f",$l);
273	&R_PERM_OP($l,$tt,$r,14,"0x33333333",$r);
274	&R_PERM_OP($tt,$r,$l,22,"0x03fc03fc",$r);
275	&R_PERM_OP($l,$r,$tt, 9,"0xaaaaaaaa",$r);
276
277	if ($lr != 3)
278		{
279		if (($lr-3) < 0)
280			{ &rotr($tt,	3-$lr); }
281		else	{ &rotl($tt,	$lr-3); }
282		}
283	if ($lr != 2)
284		{
285		if (($lr-2) < 0)
286			{ &rotr($r,	2-$lr); }
287		else	{ &rotl($r,	$lr-2); }
288		}
289	}
290
291sub FP_new
292	{
293	local($l,$r,$tt,$lr)=@_;
294
295	if ($lr != 2)
296		{
297		if (($lr-2) < 0)
298			{ &rotl($r,	2-$lr); }
299		else	{ &rotr($r,	$lr-2); }
300		}
301	if ($lr != 3)
302		{
303		if (($lr-3) < 0)
304			{ &rotl($l,	3-$lr); }
305		else	{ &rotr($l,	$lr-3); }
306		}
307
308	&R_PERM_OP($l,$r,$tt, 0,"0xaaaaaaaa",$r);
309	&R_PERM_OP($tt,$r,$l,23,"0x03fc03fc",$r);
310	&R_PERM_OP($l,$r,$tt,10,"0x33333333",$l);
311	&R_PERM_OP($r,$tt,$l,18,"0xfff0000f",$l);
312	&R_PERM_OP($l,$tt,$r,12,"0xf0f0f0f0",$r);
313	&rotr($tt	, 4);
314	}
315
316sub DES_SPtrans
317	{
318	&set_label("DES_SPtrans",64);
319	&set_label("des_sptrans");
320	&data_word(0x02080800, 0x00080000, 0x02000002, 0x02080802);
321	&data_word(0x02000000, 0x00080802, 0x00080002, 0x02000002);
322	&data_word(0x00080802, 0x02080800, 0x02080000, 0x00000802);
323	&data_word(0x02000802, 0x02000000, 0x00000000, 0x00080002);
324	&data_word(0x00080000, 0x00000002, 0x02000800, 0x00080800);
325	&data_word(0x02080802, 0x02080000, 0x00000802, 0x02000800);
326	&data_word(0x00000002, 0x00000800, 0x00080800, 0x02080002);
327	&data_word(0x00000800, 0x02000802, 0x02080002, 0x00000000);
328	&data_word(0x00000000, 0x02080802, 0x02000800, 0x00080002);
329	&data_word(0x02080800, 0x00080000, 0x00000802, 0x02000800);
330	&data_word(0x02080002, 0x00000800, 0x00080800, 0x02000002);
331	&data_word(0x00080802, 0x00000002, 0x02000002, 0x02080000);
332	&data_word(0x02080802, 0x00080800, 0x02080000, 0x02000802);
333	&data_word(0x02000000, 0x00000802, 0x00080002, 0x00000000);
334	&data_word(0x00080000, 0x02000000, 0x02000802, 0x02080800);
335	&data_word(0x00000002, 0x02080002, 0x00000800, 0x00080802);
336	# nibble 1
337	&data_word(0x40108010, 0x00000000, 0x00108000, 0x40100000);
338	&data_word(0x40000010, 0x00008010, 0x40008000, 0x00108000);
339	&data_word(0x00008000, 0x40100010, 0x00000010, 0x40008000);
340	&data_word(0x00100010, 0x40108000, 0x40100000, 0x00000010);
341	&data_word(0x00100000, 0x40008010, 0x40100010, 0x00008000);
342	&data_word(0x00108010, 0x40000000, 0x00000000, 0x00100010);
343	&data_word(0x40008010, 0x00108010, 0x40108000, 0x40000010);
344	&data_word(0x40000000, 0x00100000, 0x00008010, 0x40108010);
345	&data_word(0x00100010, 0x40108000, 0x40008000, 0x00108010);
346	&data_word(0x40108010, 0x00100010, 0x40000010, 0x00000000);
347	&data_word(0x40000000, 0x00008010, 0x00100000, 0x40100010);
348	&data_word(0x00008000, 0x40000000, 0x00108010, 0x40008010);
349	&data_word(0x40108000, 0x00008000, 0x00000000, 0x40000010);
350	&data_word(0x00000010, 0x40108010, 0x00108000, 0x40100000);
351	&data_word(0x40100010, 0x00100000, 0x00008010, 0x40008000);
352	&data_word(0x40008010, 0x00000010, 0x40100000, 0x00108000);
353	# nibble 2
354	&data_word(0x04000001, 0x04040100, 0x00000100, 0x04000101);
355	&data_word(0x00040001, 0x04000000, 0x04000101, 0x00040100);
356	&data_word(0x04000100, 0x00040000, 0x04040000, 0x00000001);
357	&data_word(0x04040101, 0x00000101, 0x00000001, 0x04040001);
358	&data_word(0x00000000, 0x00040001, 0x04040100, 0x00000100);
359	&data_word(0x00000101, 0x04040101, 0x00040000, 0x04000001);
360	&data_word(0x04040001, 0x04000100, 0x00040101, 0x04040000);
361	&data_word(0x00040100, 0x00000000, 0x04000000, 0x00040101);
362	&data_word(0x04040100, 0x00000100, 0x00000001, 0x00040000);
363	&data_word(0x00000101, 0x00040001, 0x04040000, 0x04000101);
364	&data_word(0x00000000, 0x04040100, 0x00040100, 0x04040001);
365	&data_word(0x00040001, 0x04000000, 0x04040101, 0x00000001);
366	&data_word(0x00040101, 0x04000001, 0x04000000, 0x04040101);
367	&data_word(0x00040000, 0x04000100, 0x04000101, 0x00040100);
368	&data_word(0x04000100, 0x00000000, 0x04040001, 0x00000101);
369	&data_word(0x04000001, 0x00040101, 0x00000100, 0x04040000);
370	# nibble 3
371	&data_word(0x00401008, 0x10001000, 0x00000008, 0x10401008);
372	&data_word(0x00000000, 0x10400000, 0x10001008, 0x00400008);
373	&data_word(0x10401000, 0x10000008, 0x10000000, 0x00001008);
374	&data_word(0x10000008, 0x00401008, 0x00400000, 0x10000000);
375	&data_word(0x10400008, 0x00401000, 0x00001000, 0x00000008);
376	&data_word(0x00401000, 0x10001008, 0x10400000, 0x00001000);
377	&data_word(0x00001008, 0x00000000, 0x00400008, 0x10401000);
378	&data_word(0x10001000, 0x10400008, 0x10401008, 0x00400000);
379	&data_word(0x10400008, 0x00001008, 0x00400000, 0x10000008);
380	&data_word(0x00401000, 0x10001000, 0x00000008, 0x10400000);
381	&data_word(0x10001008, 0x00000000, 0x00001000, 0x00400008);
382	&data_word(0x00000000, 0x10400008, 0x10401000, 0x00001000);
383	&data_word(0x10000000, 0x10401008, 0x00401008, 0x00400000);
384	&data_word(0x10401008, 0x00000008, 0x10001000, 0x00401008);
385	&data_word(0x00400008, 0x00401000, 0x10400000, 0x10001008);
386	&data_word(0x00001008, 0x10000000, 0x10000008, 0x10401000);
387	# nibble 4
388	&data_word(0x08000000, 0x00010000, 0x00000400, 0x08010420);
389	&data_word(0x08010020, 0x08000400, 0x00010420, 0x08010000);
390	&data_word(0x00010000, 0x00000020, 0x08000020, 0x00010400);
391	&data_word(0x08000420, 0x08010020, 0x08010400, 0x00000000);
392	&data_word(0x00010400, 0x08000000, 0x00010020, 0x00000420);
393	&data_word(0x08000400, 0x00010420, 0x00000000, 0x08000020);
394	&data_word(0x00000020, 0x08000420, 0x08010420, 0x00010020);
395	&data_word(0x08010000, 0x00000400, 0x00000420, 0x08010400);
396	&data_word(0x08010400, 0x08000420, 0x00010020, 0x08010000);
397	&data_word(0x00010000, 0x00000020, 0x08000020, 0x08000400);
398	&data_word(0x08000000, 0x00010400, 0x08010420, 0x00000000);
399	&data_word(0x00010420, 0x08000000, 0x00000400, 0x00010020);
400	&data_word(0x08000420, 0x00000400, 0x00000000, 0x08010420);
401	&data_word(0x08010020, 0x08010400, 0x00000420, 0x00010000);
402	&data_word(0x00010400, 0x08010020, 0x08000400, 0x00000420);
403	&data_word(0x00000020, 0x00010420, 0x08010000, 0x08000020);
404	# nibble 5
405	&data_word(0x80000040, 0x00200040, 0x00000000, 0x80202000);
406	&data_word(0x00200040, 0x00002000, 0x80002040, 0x00200000);
407	&data_word(0x00002040, 0x80202040, 0x00202000, 0x80000000);
408	&data_word(0x80002000, 0x80000040, 0x80200000, 0x00202040);
409	&data_word(0x00200000, 0x80002040, 0x80200040, 0x00000000);
410	&data_word(0x00002000, 0x00000040, 0x80202000, 0x80200040);
411	&data_word(0x80202040, 0x80200000, 0x80000000, 0x00002040);
412	&data_word(0x00000040, 0x00202000, 0x00202040, 0x80002000);
413	&data_word(0x00002040, 0x80000000, 0x80002000, 0x00202040);
414	&data_word(0x80202000, 0x00200040, 0x00000000, 0x80002000);
415	&data_word(0x80000000, 0x00002000, 0x80200040, 0x00200000);
416	&data_word(0x00200040, 0x80202040, 0x00202000, 0x00000040);
417	&data_word(0x80202040, 0x00202000, 0x00200000, 0x80002040);
418	&data_word(0x80000040, 0x80200000, 0x00202040, 0x00000000);
419	&data_word(0x00002000, 0x80000040, 0x80002040, 0x80202000);
420	&data_word(0x80200000, 0x00002040, 0x00000040, 0x80200040);
421	# nibble 6
422	&data_word(0x00004000, 0x00000200, 0x01000200, 0x01000004);
423	&data_word(0x01004204, 0x00004004, 0x00004200, 0x00000000);
424	&data_word(0x01000000, 0x01000204, 0x00000204, 0x01004000);
425	&data_word(0x00000004, 0x01004200, 0x01004000, 0x00000204);
426	&data_word(0x01000204, 0x00004000, 0x00004004, 0x01004204);
427	&data_word(0x00000000, 0x01000200, 0x01000004, 0x00004200);
428	&data_word(0x01004004, 0x00004204, 0x01004200, 0x00000004);
429	&data_word(0x00004204, 0x01004004, 0x00000200, 0x01000000);
430	&data_word(0x00004204, 0x01004000, 0x01004004, 0x00000204);
431	&data_word(0x00004000, 0x00000200, 0x01000000, 0x01004004);
432	&data_word(0x01000204, 0x00004204, 0x00004200, 0x00000000);
433	&data_word(0x00000200, 0x01000004, 0x00000004, 0x01000200);
434	&data_word(0x00000000, 0x01000204, 0x01000200, 0x00004200);
435	&data_word(0x00000204, 0x00004000, 0x01004204, 0x01000000);
436	&data_word(0x01004200, 0x00000004, 0x00004004, 0x01004204);
437	&data_word(0x01000004, 0x01004200, 0x01004000, 0x00004004);
438	# nibble 7
439	&data_word(0x20800080, 0x20820000, 0x00020080, 0x00000000);
440	&data_word(0x20020000, 0x00800080, 0x20800000, 0x20820080);
441	&data_word(0x00000080, 0x20000000, 0x00820000, 0x00020080);
442	&data_word(0x00820080, 0x20020080, 0x20000080, 0x20800000);
443	&data_word(0x00020000, 0x00820080, 0x00800080, 0x20020000);
444	&data_word(0x20820080, 0x20000080, 0x00000000, 0x00820000);
445	&data_word(0x20000000, 0x00800000, 0x20020080, 0x20800080);
446	&data_word(0x00800000, 0x00020000, 0x20820000, 0x00000080);
447	&data_word(0x00800000, 0x00020000, 0x20000080, 0x20820080);
448	&data_word(0x00020080, 0x20000000, 0x00000000, 0x00820000);
449	&data_word(0x20800080, 0x20020080, 0x20020000, 0x00800080);
450	&data_word(0x20820000, 0x00000080, 0x00800080, 0x20020000);
451	&data_word(0x20820080, 0x00800000, 0x20800000, 0x20000080);
452	&data_word(0x00820000, 0x00020080, 0x20020080, 0x20800000);
453	&data_word(0x00000080, 0x20820000, 0x00820080, 0x00000000);
454	&data_word(0x20000000, 0x20800080, 0x00020000, 0x00820080);
455	}
456