aes-x86_64.pl revision f48372ded3bb76c2598392aa58abe6e2eb7432d2 
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# Version 1.2.
10#
11# aes-*-cbc benchmarks are improved by >70% [compared to gcc 3.3.2 on
12# Opteron 240 CPU] plus all the bells-n-whistles from 32-bit version
13# [you'll notice a lot of resemblance], such as compressed S-boxes
14# in little-endian byte order, prefetch of these tables in CBC mode,
15# as well as avoiding L1 cache aliasing between stack frame and key
16# schedule and already mentioned tables, compressed Td4...
17#
18# Performance in number of cycles per processed byte for 128-bit key:
19#
20#		ECB		CBC encrypt
21# AMD64		13.7		13.0(*)
22# EM64T		20.2		18.6(*)
23#
24# (*)	CBC benchmarks are better than ECB thanks to custom ABI used
25#	by the private block encryption function.
26
27$verticalspin=1;	# unlike 32-bit version $verticalspin performs
28			# ~15% better on both AMD and Intel cores
29$output=shift;
30open STDOUT,"| $^X ../perlasm/x86_64-xlate.pl $output";
31
32$code=".text\n";
33
34$s0="%eax";
35$s1="%ebx";
36$s2="%ecx";
37$s3="%edx";
38$acc0="%esi";
39$acc1="%edi";
40$acc2="%ebp";
41$inp="%r8";
42$out="%r9";
43$t0="%r10d";
44$t1="%r11d";
45$t2="%r12d";
46$rnds="%r13d";
47$sbox="%r14";
48$key="%r15";
49
50sub hi() { my $r=shift;	$r =~ s/%[er]([a-d])x/%\1h/;	$r; }
51sub lo() { my $r=shift;	$r =~ s/%[er]([a-d])x/%\1l/;
52			$r =~ s/%[er]([sd]i)/%\1l/;
53			$r =~ s/%(r[0-9]+)[d]?/%\1b/;	$r; }
54sub _data_word()
55{ my $i;
56    while(defined($i=shift)) { $code.=sprintf".long\t0x%08x,0x%08x\n",$i,$i; }
57}
58sub data_word()
59{ my $i;
60  my $last=pop(@_);
61    $code.=".long\t";
62    while(defined($i=shift)) { $code.=sprintf"0x%08x,",$i; }
63    $code.=sprintf"0x%08x\n",$last;
64}
65
66sub data_byte()
67{ my $i;
68  my $last=pop(@_);
69    $code.=".byte\t";
70    while(defined($i=shift)) { $code.=sprintf"0x%02x,",$i&0xff; }
71    $code.=sprintf"0x%02x\n",$last&0xff;
72}
73
74sub encvert()
75{ my $t3="%r8d";	# zaps $inp!
76
77$code.=<<___;
78	# favor 3-way issue Opteron pipeline...
79	movzb	`&lo("$s0")`,$acc0
80	movzb	`&lo("$s1")`,$acc1
81	movzb	`&lo("$s2")`,$acc2
82	mov	0($sbox,$acc0,8),$t0
83	mov	0($sbox,$acc1,8),$t1
84	mov	0($sbox,$acc2,8),$t2
85
86	movzb	`&hi("$s1")`,$acc0
87	movzb	`&hi("$s2")`,$acc1
88	movzb	`&lo("$s3")`,$acc2
89	xor	3($sbox,$acc0,8),$t0
90	xor	3($sbox,$acc1,8),$t1
91	mov	0($sbox,$acc2,8),$t3
92
93	movzb	`&hi("$s3")`,$acc0
94	shr	\$16,$s2
95	movzb	`&hi("$s0")`,$acc2
96	xor	3($sbox,$acc0,8),$t2
97	shr	\$16,$s3
98	xor	3($sbox,$acc2,8),$t3
99
100	shr	\$16,$s1
101	lea	16($key),$key
102	shr	\$16,$s0
103
104	movzb	`&lo("$s2")`,$acc0
105	movzb	`&lo("$s3")`,$acc1
106	movzb	`&lo("$s0")`,$acc2
107	xor	2($sbox,$acc0,8),$t0
108	xor	2($sbox,$acc1,8),$t1
109	xor	2($sbox,$acc2,8),$t2
110
111	movzb	`&hi("$s3")`,$acc0
112	movzb	`&hi("$s0")`,$acc1
113	movzb	`&lo("$s1")`,$acc2
114	xor	1($sbox,$acc0,8),$t0
115	xor	1($sbox,$acc1,8),$t1
116	xor	2($sbox,$acc2,8),$t3
117
118	mov	12($key),$s3
119	movzb	`&hi("$s1")`,$acc1
120	movzb	`&hi("$s2")`,$acc2
121	mov	0($key),$s0
122	xor	1($sbox,$acc1,8),$t2
123	xor	1($sbox,$acc2,8),$t3
124
125	mov	4($key),$s1
126	mov	8($key),$s2
127	xor	$t0,$s0
128	xor	$t1,$s1
129	xor	$t2,$s2
130	xor	$t3,$s3
131___
132}
133
134sub enclastvert()
135{ my $t3="%r8d";	# zaps $inp!
136
137$code.=<<___;
138	movzb	`&lo("$s0")`,$acc0
139	movzb	`&lo("$s1")`,$acc1
140	movzb	`&lo("$s2")`,$acc2
141	mov	2($sbox,$acc0,8),$t0
142	mov	2($sbox,$acc1,8),$t1
143	mov	2($sbox,$acc2,8),$t2
144
145	and	\$0x000000ff,$t0
146	and	\$0x000000ff,$t1
147	and	\$0x000000ff,$t2
148
149	movzb	`&lo("$s3")`,$acc0
150	movzb	`&hi("$s1")`,$acc1
151	movzb	`&hi("$s2")`,$acc2
152	mov	2($sbox,$acc0,8),$t3
153	mov	0($sbox,$acc1,8),$acc1	#$t0
154	mov	0($sbox,$acc2,8),$acc2	#$t1
155
156	and	\$0x000000ff,$t3
157	and	\$0x0000ff00,$acc1
158	and	\$0x0000ff00,$acc2
159
160	xor	$acc1,$t0
161	xor	$acc2,$t1
162	shr	\$16,$s2
163
164	movzb	`&hi("$s3")`,$acc0
165	movzb	`&hi("$s0")`,$acc1
166	shr	\$16,$s3
167	mov	0($sbox,$acc0,8),$acc0	#$t2
168	mov	0($sbox,$acc1,8),$acc1	#$t3
169
170	and	\$0x0000ff00,$acc0
171	and	\$0x0000ff00,$acc1
172	shr	\$16,$s1
173	xor	$acc0,$t2
174	xor	$acc1,$t3
175	shr	\$16,$s0
176
177	movzb	`&lo("$s2")`,$acc0
178	movzb	`&lo("$s3")`,$acc1
179	movzb	`&lo("$s0")`,$acc2
180	mov	0($sbox,$acc0,8),$acc0	#$t0
181	mov	0($sbox,$acc1,8),$acc1	#$t1
182	mov	0($sbox,$acc2,8),$acc2	#$t2
183
184	and	\$0x00ff0000,$acc0
185	and	\$0x00ff0000,$acc1
186	and	\$0x00ff0000,$acc2
187
188	xor	$acc0,$t0
189	xor	$acc1,$t1
190	xor	$acc2,$t2
191
192	movzb	`&lo("$s1")`,$acc0
193	movzb	`&hi("$s3")`,$acc1
194	movzb	`&hi("$s0")`,$acc2
195	mov	0($sbox,$acc0,8),$acc0	#$t3
196	mov	2($sbox,$acc1,8),$acc1	#$t0
197	mov	2($sbox,$acc2,8),$acc2	#$t1
198
199	and	\$0x00ff0000,$acc0
200	and	\$0xff000000,$acc1
201	and	\$0xff000000,$acc2
202
203	xor	$acc0,$t3
204	xor	$acc1,$t0
205	xor	$acc2,$t1
206
207	movzb	`&hi("$s1")`,$acc0
208	movzb	`&hi("$s2")`,$acc1
209	mov	16+12($key),$s3
210	mov	2($sbox,$acc0,8),$acc0	#$t2
211	mov	2($sbox,$acc1,8),$acc1	#$t3
212	mov	16+0($key),$s0
213
214	and	\$0xff000000,$acc0
215	and	\$0xff000000,$acc1
216
217	xor	$acc0,$t2
218	xor	$acc1,$t3
219
220	mov	16+4($key),$s1
221	mov	16+8($key),$s2
222	xor	$t0,$s0
223	xor	$t1,$s1
224	xor	$t2,$s2
225	xor	$t3,$s3
226___
227}
228
229sub encstep()
230{ my ($i,@s) = @_;
231  my $tmp0=$acc0;
232  my $tmp1=$acc1;
233  my $tmp2=$acc2;
234  my $out=($t0,$t1,$t2,$s[0])[$i];
235
236	if ($i==3) {
237		$tmp0=$s[1];
238		$tmp1=$s[2];
239		$tmp2=$s[3];
240	}
241	$code.="	movzb	".&lo($s[0]).",$out\n";
242	$code.="	mov	$s[2],$tmp1\n"		if ($i!=3);
243	$code.="	lea	16($key),$key\n"	if ($i==0);
244
245	$code.="	movzb	".&hi($s[1]).",$tmp0\n";
246	$code.="	mov	0($sbox,$out,8),$out\n";
247
248	$code.="	shr	\$16,$tmp1\n";
249	$code.="	mov	$s[3],$tmp2\n"		if ($i!=3);
250	$code.="	xor	3($sbox,$tmp0,8),$out\n";
251
252	$code.="	movzb	".&lo($tmp1).",$tmp1\n";
253	$code.="	shr	\$24,$tmp2\n";
254	$code.="	xor	4*$i($key),$out\n";
255
256	$code.="	xor	2($sbox,$tmp1,8),$out\n";
257	$code.="	xor	1($sbox,$tmp2,8),$out\n";
258
259	$code.="	mov	$t0,$s[1]\n"		if ($i==3);
260	$code.="	mov	$t1,$s[2]\n"		if ($i==3);
261	$code.="	mov	$t2,$s[3]\n"		if ($i==3);
262	$code.="\n";
263}
264
265sub enclast()
266{ my ($i,@s)=@_;
267  my $tmp0=$acc0;
268  my $tmp1=$acc1;
269  my $tmp2=$acc2;
270  my $out=($t0,$t1,$t2,$s[0])[$i];
271
272	if ($i==3) {
273		$tmp0=$s[1];
274		$tmp1=$s[2];
275		$tmp2=$s[3];
276	}
277	$code.="	movzb	".&lo($s[0]).",$out\n";
278	$code.="	mov	$s[2],$tmp1\n"		if ($i!=3);
279
280	$code.="	mov	2($sbox,$out,8),$out\n";
281	$code.="	shr	\$16,$tmp1\n";
282	$code.="	mov	$s[3],$tmp2\n"		if ($i!=3);
283
284	$code.="	and	\$0x000000ff,$out\n";
285	$code.="	movzb	".&hi($s[1]).",$tmp0\n";
286	$code.="	movzb	".&lo($tmp1).",$tmp1\n";
287	$code.="	shr	\$24,$tmp2\n";
288
289	$code.="	mov	0($sbox,$tmp0,8),$tmp0\n";
290	$code.="	mov	0($sbox,$tmp1,8),$tmp1\n";
291	$code.="	mov	2($sbox,$tmp2,8),$tmp2\n";
292
293	$code.="	and	\$0x0000ff00,$tmp0\n";
294	$code.="	and	\$0x00ff0000,$tmp1\n";
295	$code.="	and	\$0xff000000,$tmp2\n";
296
297	$code.="	xor	$tmp0,$out\n";
298	$code.="	mov	$t0,$s[1]\n"		if ($i==3);
299	$code.="	xor	$tmp1,$out\n";
300	$code.="	mov	$t1,$s[2]\n"		if ($i==3);
301	$code.="	xor	$tmp2,$out\n";
302	$code.="	mov	$t2,$s[3]\n"		if ($i==3);
303	$code.="\n";
304}
305
306$code.=<<___;
307.type	_x86_64_AES_encrypt,\@abi-omnipotent
308.align	16
309_x86_64_AES_encrypt:
310	xor	0($key),$s0			# xor with key
311	xor	4($key),$s1
312	xor	8($key),$s2
313	xor	12($key),$s3
314
315	mov	240($key),$rnds			# load key->rounds
316	sub	\$1,$rnds
317	jmp	.Lenc_loop
318.align	16
319.Lenc_loop:
320___
321	if ($verticalspin) { &encvert(); }
322	else {	&encstep(0,$s0,$s1,$s2,$s3);
323		&encstep(1,$s1,$s2,$s3,$s0);
324		&encstep(2,$s2,$s3,$s0,$s1);
325		&encstep(3,$s3,$s0,$s1,$s2);
326	}
327$code.=<<___;
328	sub	\$1,$rnds
329	jnz	.Lenc_loop
330___
331	if ($verticalspin) { &enclastvert(); }
332	else {	&enclast(0,$s0,$s1,$s2,$s3);
333		&enclast(1,$s1,$s2,$s3,$s0);
334		&enclast(2,$s2,$s3,$s0,$s1);
335		&enclast(3,$s3,$s0,$s1,$s2);
336		$code.=<<___;
337		xor	16+0($key),$s0		# xor with key
338		xor	16+4($key),$s1
339		xor	16+8($key),$s2
340		xor	16+12($key),$s3
341___
342	}
343$code.=<<___;
344	.byte	0xf3,0xc3			# rep ret
345.size	_x86_64_AES_encrypt,.-_x86_64_AES_encrypt
346___
347
348# void AES_encrypt (const void *inp,void *out,const AES_KEY *key);
349$code.=<<___;
350.globl	AES_encrypt
351.type	AES_encrypt,\@function,3
352.align	16
353AES_encrypt:
354	push	%rbx
355	push	%rbp
356	push	%r12
357	push	%r13
358	push	%r14
359	push	%r15
360
361	mov	%rdx,$key
362	mov	%rdi,$inp
363	mov	%rsi,$out
364
365	.picmeup	$sbox
366	lea	AES_Te-.($sbox),$sbox
367
368	mov	0($inp),$s0
369	mov	4($inp),$s1
370	mov	8($inp),$s2
371	mov	12($inp),$s3
372
373	call	_x86_64_AES_encrypt
374
375	mov	$s0,0($out)
376	mov	$s1,4($out)
377	mov	$s2,8($out)
378	mov	$s3,12($out)
379
380	pop	%r15
381	pop	%r14
382	pop	%r13
383	pop	%r12
384	pop	%rbp
385	pop	%rbx
386	ret
387.size	AES_encrypt,.-AES_encrypt
388___
389
390#------------------------------------------------------------------#
391
392sub decvert()
393{ my $t3="%r8d";	# zaps $inp!
394
395$code.=<<___;
396	# favor 3-way issue Opteron pipeline...
397	movzb	`&lo("$s0")`,$acc0
398	movzb	`&lo("$s1")`,$acc1
399	movzb	`&lo("$s2")`,$acc2
400	mov	0($sbox,$acc0,8),$t0
401	mov	0($sbox,$acc1,8),$t1
402	mov	0($sbox,$acc2,8),$t2
403
404	movzb	`&hi("$s3")`,$acc0
405	movzb	`&hi("$s0")`,$acc1
406	movzb	`&lo("$s3")`,$acc2
407	xor	3($sbox,$acc0,8),$t0
408	xor	3($sbox,$acc1,8),$t1
409	mov	0($sbox,$acc2,8),$t3
410
411	movzb	`&hi("$s1")`,$acc0
412	shr	\$16,$s0
413	movzb	`&hi("$s2")`,$acc2
414	xor	3($sbox,$acc0,8),$t2
415	shr	\$16,$s3
416	xor	3($sbox,$acc2,8),$t3
417
418	shr	\$16,$s1
419	lea	16($key),$key
420	shr	\$16,$s2
421
422	movzb	`&lo("$s2")`,$acc0
423	movzb	`&lo("$s3")`,$acc1
424	movzb	`&lo("$s0")`,$acc2
425	xor	2($sbox,$acc0,8),$t0
426	xor	2($sbox,$acc1,8),$t1
427	xor	2($sbox,$acc2,8),$t2
428
429	movzb	`&hi("$s1")`,$acc0
430	movzb	`&hi("$s2")`,$acc1
431	movzb	`&lo("$s1")`,$acc2
432	xor	1($sbox,$acc0,8),$t0
433	xor	1($sbox,$acc1,8),$t1
434	xor	2($sbox,$acc2,8),$t3
435
436	movzb	`&hi("$s3")`,$acc0
437	mov	12($key),$s3
438	movzb	`&hi("$s0")`,$acc2
439	xor	1($sbox,$acc0,8),$t2
440	mov	0($key),$s0
441	xor	1($sbox,$acc2,8),$t3
442
443	xor	$t0,$s0
444	mov	4($key),$s1
445	mov	8($key),$s2
446	xor	$t2,$s2
447	xor	$t1,$s1
448	xor	$t3,$s3
449___
450}
451
452sub declastvert()
453{ my $t3="%r8d";	# zaps $inp!
454
455$code.=<<___;
456	movzb	`&lo("$s0")`,$acc0
457	movzb	`&lo("$s1")`,$acc1
458	movzb	`&lo("$s2")`,$acc2
459	movzb	2048($sbox,$acc0,1),$t0
460	movzb	2048($sbox,$acc1,1),$t1
461	movzb	2048($sbox,$acc2,1),$t2
462
463	movzb	`&lo("$s3")`,$acc0
464	movzb	`&hi("$s3")`,$acc1
465	movzb	`&hi("$s0")`,$acc2
466	movzb	2048($sbox,$acc0,1),$t3
467	movzb	2048($sbox,$acc1,1),$acc1	#$t0
468	movzb	2048($sbox,$acc2,1),$acc2	#$t1
469
470	shl	\$8,$acc1
471	shl	\$8,$acc2
472
473	xor	$acc1,$t0
474	xor	$acc2,$t1
475	shr	\$16,$s3
476
477	movzb	`&hi("$s1")`,$acc0
478	movzb	`&hi("$s2")`,$acc1
479	shr	\$16,$s0
480	movzb	2048($sbox,$acc0,1),$acc0	#$t2
481	movzb	2048($sbox,$acc1,1),$acc1	#$t3
482
483	shl	\$8,$acc0
484	shl	\$8,$acc1
485	shr	\$16,$s1
486	xor	$acc0,$t2
487	xor	$acc1,$t3
488	shr	\$16,$s2
489
490	movzb	`&lo("$s2")`,$acc0
491	movzb	`&lo("$s3")`,$acc1
492	movzb	`&lo("$s0")`,$acc2
493	movzb	2048($sbox,$acc0,1),$acc0	#$t0
494	movzb	2048($sbox,$acc1,1),$acc1	#$t1
495	movzb	2048($sbox,$acc2,1),$acc2	#$t2
496
497	shl	\$16,$acc0
498	shl	\$16,$acc1
499	shl	\$16,$acc2
500
501	xor	$acc0,$t0
502	xor	$acc1,$t1
503	xor	$acc2,$t2
504
505	movzb	`&lo("$s1")`,$acc0
506	movzb	`&hi("$s1")`,$acc1
507	movzb	`&hi("$s2")`,$acc2
508	movzb	2048($sbox,$acc0,1),$acc0	#$t3
509	movzb	2048($sbox,$acc1,1),$acc1	#$t0
510	movzb	2048($sbox,$acc2,1),$acc2	#$t1
511
512	shl	\$16,$acc0
513	shl	\$24,$acc1
514	shl	\$24,$acc2
515
516	xor	$acc0,$t3
517	xor	$acc1,$t0
518	xor	$acc2,$t1
519
520	movzb	`&hi("$s3")`,$acc0
521	movzb	`&hi("$s0")`,$acc1
522	mov	16+12($key),$s3
523	movzb	2048($sbox,$acc0,1),$acc0	#$t2
524	movzb	2048($sbox,$acc1,1),$acc1	#$t3
525	mov	16+0($key),$s0
526
527	shl	\$24,$acc0
528	shl	\$24,$acc1
529
530	xor	$acc0,$t2
531	xor	$acc1,$t3
532
533	mov	16+4($key),$s1
534	mov	16+8($key),$s2
535	xor	$t0,$s0
536	xor	$t1,$s1
537	xor	$t2,$s2
538	xor	$t3,$s3
539___
540}
541
542sub decstep()
543{ my ($i,@s) = @_;
544  my $tmp0=$acc0;
545  my $tmp1=$acc1;
546  my $tmp2=$acc2;
547  my $out=($t0,$t1,$t2,$s[0])[$i];
548
549	$code.="	mov	$s[0],$out\n"		if ($i!=3);
550			$tmp1=$s[2]			if ($i==3);
551	$code.="	mov	$s[2],$tmp1\n"		if ($i!=3);
552	$code.="	and	\$0xFF,$out\n";
553
554	$code.="	mov	0($sbox,$out,8),$out\n";
555	$code.="	shr	\$16,$tmp1\n";
556			$tmp2=$s[3]			if ($i==3);
557	$code.="	mov	$s[3],$tmp2\n"		if ($i!=3);
558
559			$tmp0=$s[1]			if ($i==3);
560	$code.="	movzb	".&hi($s[1]).",$tmp0\n";
561	$code.="	and	\$0xFF,$tmp1\n";
562	$code.="	shr	\$24,$tmp2\n";
563
564	$code.="	xor	3($sbox,$tmp0,8),$out\n";
565	$code.="	xor	2($sbox,$tmp1,8),$out\n";
566	$code.="	xor	1($sbox,$tmp2,8),$out\n";
567
568	$code.="	mov	$t2,$s[1]\n"		if ($i==3);
569	$code.="	mov	$t1,$s[2]\n"		if ($i==3);
570	$code.="	mov	$t0,$s[3]\n"		if ($i==3);
571	$code.="\n";
572}
573
574sub declast()
575{ my ($i,@s)=@_;
576  my $tmp0=$acc0;
577  my $tmp1=$acc1;
578  my $tmp2=$acc2;
579  my $out=($t0,$t1,$t2,$s[0])[$i];
580
581	$code.="	mov	$s[0],$out\n"		if ($i!=3);
582			$tmp1=$s[2]			if ($i==3);
583	$code.="	mov	$s[2],$tmp1\n"		if ($i!=3);
584	$code.="	and	\$0xFF,$out\n";
585
586	$code.="	movzb	2048($sbox,$out,1),$out\n";
587	$code.="	shr	\$16,$tmp1\n";
588			$tmp2=$s[3]			if ($i==3);
589	$code.="	mov	$s[3],$tmp2\n"		if ($i!=3);
590
591			$tmp0=$s[1]			if ($i==3);
592	$code.="	movzb	".&hi($s[1]).",$tmp0\n";
593	$code.="	and	\$0xFF,$tmp1\n";
594	$code.="	shr	\$24,$tmp2\n";
595
596	$code.="	movzb	2048($sbox,$tmp0,1),$tmp0\n";
597	$code.="	movzb	2048($sbox,$tmp1,1),$tmp1\n";
598	$code.="	movzb	2048($sbox,$tmp2,1),$tmp2\n";
599
600	$code.="	shl	\$8,$tmp0\n";
601	$code.="	shl	\$16,$tmp1\n";
602	$code.="	shl	\$24,$tmp2\n";
603
604	$code.="	xor	$tmp0,$out\n";
605	$code.="	mov	$t2,$s[1]\n"		if ($i==3);
606	$code.="	xor	$tmp1,$out\n";
607	$code.="	mov	$t1,$s[2]\n"		if ($i==3);
608	$code.="	xor	$tmp2,$out\n";
609	$code.="	mov	$t0,$s[3]\n"		if ($i==3);
610	$code.="\n";
611}
612
613$code.=<<___;
614.type	_x86_64_AES_decrypt,\@abi-omnipotent
615.align	16
616_x86_64_AES_decrypt:
617	xor	0($key),$s0			# xor with key
618	xor	4($key),$s1
619	xor	8($key),$s2
620	xor	12($key),$s3
621
622	mov	240($key),$rnds			# load key->rounds
623	sub	\$1,$rnds
624	jmp	.Ldec_loop
625.align	16
626.Ldec_loop:
627___
628	if ($verticalspin) { &decvert(); }
629	else {	&decstep(0,$s0,$s3,$s2,$s1);
630		&decstep(1,$s1,$s0,$s3,$s2);
631		&decstep(2,$s2,$s1,$s0,$s3);
632		&decstep(3,$s3,$s2,$s1,$s0);
633		$code.=<<___;
634		lea	16($key),$key
635		xor	0($key),$s0			# xor with key
636		xor	4($key),$s1
637		xor	8($key),$s2
638		xor	12($key),$s3
639___
640	}
641$code.=<<___;
642	sub	\$1,$rnds
643	jnz	.Ldec_loop
644___
645	if ($verticalspin) { &declastvert(); }
646	else {	&declast(0,$s0,$s3,$s2,$s1);
647		&declast(1,$s1,$s0,$s3,$s2);
648		&declast(2,$s2,$s1,$s0,$s3);
649		&declast(3,$s3,$s2,$s1,$s0);
650		$code.=<<___;
651		xor	16+0($key),$s0			# xor with key
652		xor	16+4($key),$s1
653		xor	16+8($key),$s2
654		xor	16+12($key),$s3
655___
656	}
657$code.=<<___;
658	.byte	0xf3,0xc3			# rep ret
659.size	_x86_64_AES_decrypt,.-_x86_64_AES_decrypt
660___
661
662# void AES_decrypt (const void *inp,void *out,const AES_KEY *key);
663$code.=<<___;
664.globl	AES_decrypt
665.type	AES_decrypt,\@function,3
666.align	16
667AES_decrypt:
668	push	%rbx
669	push	%rbp
670	push	%r12
671	push	%r13
672	push	%r14
673	push	%r15
674
675	mov	%rdx,$key
676	mov	%rdi,$inp
677	mov	%rsi,$out
678
679	.picmeup	$sbox
680	lea	AES_Td-.($sbox),$sbox
681
682	# prefetch Td4
683	lea	2048+128($sbox),$sbox;
684	mov	0-128($sbox),$s0
685	mov	32-128($sbox),$s1
686	mov	64-128($sbox),$s2
687	mov	96-128($sbox),$s3
688	mov	128-128($sbox),$s0
689	mov	160-128($sbox),$s1
690	mov	192-128($sbox),$s2
691	mov	224-128($sbox),$s3
692	lea	-2048-128($sbox),$sbox;
693
694	mov	0($inp),$s0
695	mov	4($inp),$s1
696	mov	8($inp),$s2
697	mov	12($inp),$s3
698
699	call	_x86_64_AES_decrypt
700
701	mov	$s0,0($out)
702	mov	$s1,4($out)
703	mov	$s2,8($out)
704	mov	$s3,12($out)
705
706	pop	%r15
707	pop	%r14
708	pop	%r13
709	pop	%r12
710	pop	%rbp
711	pop	%rbx
712	ret
713.size	AES_decrypt,.-AES_decrypt
714___
715#------------------------------------------------------------------#
716
717sub enckey()
718{
719$code.=<<___;
720	movz	%dl,%esi		# rk[i]>>0
721	mov	2(%rbp,%rsi,8),%ebx
722	movz	%dh,%esi		# rk[i]>>8
723	and	\$0xFF000000,%ebx
724	xor	%ebx,%eax
725
726	mov	2(%rbp,%rsi,8),%ebx
727	shr	\$16,%edx
728	and	\$0x000000FF,%ebx
729	movz	%dl,%esi		# rk[i]>>16
730	xor	%ebx,%eax
731
732	mov	0(%rbp,%rsi,8),%ebx
733	movz	%dh,%esi		# rk[i]>>24
734	and	\$0x0000FF00,%ebx
735	xor	%ebx,%eax
736
737	mov	0(%rbp,%rsi,8),%ebx
738	and	\$0x00FF0000,%ebx
739	xor	%ebx,%eax
740
741	xor	2048(%rbp,%rcx,4),%eax		# rcon
742___
743}
744
745# int AES_set_encrypt_key(const unsigned char *userKey, const int bits,
746#                        AES_KEY *key)
747$code.=<<___;
748.globl	AES_set_encrypt_key
749.type	AES_set_encrypt_key,\@function,3
750.align	16
751AES_set_encrypt_key:
752	push	%rbx
753	push	%rbp
754
755	mov	%esi,%ecx			# %ecx=bits
756	mov	%rdi,%rsi			# %rsi=userKey
757	mov	%rdx,%rdi			# %rdi=key
758
759	test	\$-1,%rsi
760	jz	.Lbadpointer
761	test	\$-1,%rdi
762	jz	.Lbadpointer
763
764	.picmeup %rbp
765	lea	AES_Te-.(%rbp),%rbp
766
767	cmp	\$128,%ecx
768	je	.L10rounds
769	cmp	\$192,%ecx
770	je	.L12rounds
771	cmp	\$256,%ecx
772	je	.L14rounds
773	mov	\$-2,%rax			# invalid number of bits
774	jmp	.Lexit
775
776.L10rounds:
777	mov	0(%rsi),%eax			# copy first 4 dwords
778	mov	4(%rsi),%ebx
779	mov	8(%rsi),%ecx
780	mov	12(%rsi),%edx
781	mov	%eax,0(%rdi)
782	mov	%ebx,4(%rdi)
783	mov	%ecx,8(%rdi)
784	mov	%edx,12(%rdi)
785
786	xor	%ecx,%ecx
787	jmp	.L10shortcut
788.align	4
789.L10loop:
790		mov	0(%rdi),%eax			# rk[0]
791		mov	12(%rdi),%edx			# rk[3]
792.L10shortcut:
793___
794		&enckey	();
795$code.=<<___;
796		mov	%eax,16(%rdi)			# rk[4]
797		xor	4(%rdi),%eax
798		mov	%eax,20(%rdi)			# rk[5]
799		xor	8(%rdi),%eax
800		mov	%eax,24(%rdi)			# rk[6]
801		xor	12(%rdi),%eax
802		mov	%eax,28(%rdi)			# rk[7]
803		add	\$1,%ecx
804		lea	16(%rdi),%rdi
805		cmp	\$10,%ecx
806	jl	.L10loop
807
808	movl	\$10,80(%rdi)			# setup number of rounds
809	xor	%rax,%rax
810	jmp	.Lexit
811
812.L12rounds:
813	mov	0(%rsi),%eax			# copy first 6 dwords
814	mov	4(%rsi),%ebx
815	mov	8(%rsi),%ecx
816	mov	12(%rsi),%edx
817	mov	%eax,0(%rdi)
818	mov	%ebx,4(%rdi)
819	mov	%ecx,8(%rdi)
820	mov	%edx,12(%rdi)
821	mov	16(%rsi),%ecx
822	mov	20(%rsi),%edx
823	mov	%ecx,16(%rdi)
824	mov	%edx,20(%rdi)
825
826	xor	%ecx,%ecx
827	jmp	.L12shortcut
828.align	4
829.L12loop:
830		mov	0(%rdi),%eax			# rk[0]
831		mov	20(%rdi),%edx			# rk[5]
832.L12shortcut:
833___
834		&enckey	();
835$code.=<<___;
836		mov	%eax,24(%rdi)			# rk[6]
837		xor	4(%rdi),%eax
838		mov	%eax,28(%rdi)			# rk[7]
839		xor	8(%rdi),%eax
840		mov	%eax,32(%rdi)			# rk[8]
841		xor	12(%rdi),%eax
842		mov	%eax,36(%rdi)			# rk[9]
843
844		cmp	\$7,%ecx
845		je	.L12break
846		add	\$1,%ecx
847
848		xor	16(%rdi),%eax
849		mov	%eax,40(%rdi)			# rk[10]
850		xor	20(%rdi),%eax
851		mov	%eax,44(%rdi)			# rk[11]
852
853		lea	24(%rdi),%rdi
854	jmp	.L12loop
855.L12break:
856	movl	\$12,72(%rdi)		# setup number of rounds
857	xor	%rax,%rax
858	jmp	.Lexit
859
860.L14rounds:
861	mov	0(%rsi),%eax			# copy first 8 dwords
862	mov	4(%rsi),%ebx
863	mov	8(%rsi),%ecx
864	mov	12(%rsi),%edx
865	mov	%eax,0(%rdi)
866	mov	%ebx,4(%rdi)
867	mov	%ecx,8(%rdi)
868	mov	%edx,12(%rdi)
869	mov	16(%rsi),%eax
870	mov	20(%rsi),%ebx
871	mov	24(%rsi),%ecx
872	mov	28(%rsi),%edx
873	mov	%eax,16(%rdi)
874	mov	%ebx,20(%rdi)
875	mov	%ecx,24(%rdi)
876	mov	%edx,28(%rdi)
877
878	xor	%ecx,%ecx
879	jmp	.L14shortcut
880.align	4
881.L14loop:
882		mov	28(%rdi),%edx			# rk[4]
883.L14shortcut:
884		mov	0(%rdi),%eax			# rk[0]
885___
886		&enckey	();
887$code.=<<___;
888		mov	%eax,32(%rdi)			# rk[8]
889		xor	4(%rdi),%eax
890		mov	%eax,36(%rdi)			# rk[9]
891		xor	8(%rdi),%eax
892		mov	%eax,40(%rdi)			# rk[10]
893		xor	12(%rdi),%eax
894		mov	%eax,44(%rdi)			# rk[11]
895
896		cmp	\$6,%ecx
897		je	.L14break
898		add	\$1,%ecx
899
900		mov	%eax,%edx
901		mov	16(%rdi),%eax			# rk[4]
902		movz	%dl,%esi			# rk[11]>>0
903		mov	2(%rbp,%rsi,8),%ebx
904		movz	%dh,%esi			# rk[11]>>8
905		and	\$0x000000FF,%ebx
906		xor	%ebx,%eax
907
908		mov	0(%rbp,%rsi,8),%ebx
909		shr	\$16,%edx
910		and	\$0x0000FF00,%ebx
911		movz	%dl,%esi			# rk[11]>>16
912		xor	%ebx,%eax
913
914		mov	0(%rbp,%rsi,8),%ebx
915		movz	%dh,%esi			# rk[11]>>24
916		and	\$0x00FF0000,%ebx
917		xor	%ebx,%eax
918
919		mov	2(%rbp,%rsi,8),%ebx
920		and	\$0xFF000000,%ebx
921		xor	%ebx,%eax
922
923		mov	%eax,48(%rdi)			# rk[12]
924		xor	20(%rdi),%eax
925		mov	%eax,52(%rdi)			# rk[13]
926		xor	24(%rdi),%eax
927		mov	%eax,56(%rdi)			# rk[14]
928		xor	28(%rdi),%eax
929		mov	%eax,60(%rdi)			# rk[15]
930
931		lea	32(%rdi),%rdi
932	jmp	.L14loop
933.L14break:
934	movl	\$14,48(%rdi)		# setup number of rounds
935	xor	%rax,%rax
936	jmp	.Lexit
937
938.Lbadpointer:
939	mov	\$-1,%rax
940.Lexit:
941	pop	%rbp
942	pop	%rbx
943	ret
944.size	AES_set_encrypt_key,.-AES_set_encrypt_key
945___
946
947sub deckey()
948{ my ($i,$ptr,$te,$td) = @_;
949$code.=<<___;
950	mov	$i($ptr),%eax
951	mov	%eax,%edx
952	movz	%ah,%ebx
953	shr	\$16,%edx
954	and	\$0xFF,%eax
955	movzb	2($te,%rax,8),%rax
956	movzb	2($te,%rbx,8),%rbx
957	mov	0($td,%rax,8),%eax
958	xor	3($td,%rbx,8),%eax
959	movzb	%dh,%ebx
960	and	\$0xFF,%edx
961	movzb	2($te,%rdx,8),%rdx
962	movzb	2($te,%rbx,8),%rbx
963	xor	2($td,%rdx,8),%eax
964	xor	1($td,%rbx,8),%eax
965	mov	%eax,$i($ptr)
966___
967}
968
969# int AES_set_decrypt_key(const unsigned char *userKey, const int bits,
970#                        AES_KEY *key)
971$code.=<<___;
972.globl	AES_set_decrypt_key
973.type	AES_set_decrypt_key,\@function,3
974.align	16
975AES_set_decrypt_key:
976	push	%rdx
977	call	AES_set_encrypt_key
978	cmp	\$0,%eax
979	je	.Lproceed
980	lea	24(%rsp),%rsp
981	ret
982.Lproceed:
983	mov	(%rsp),%r8		# restore key schedule
984	mov	%rbx,(%rsp)
985
986	mov	240(%r8),%ecx		# pull number of rounds
987	xor	%rdi,%rdi
988	lea	(%rdi,%rcx,4),%rcx
989	mov	%r8,%rsi
990	lea	(%r8,%rcx,4),%rdi	# pointer to last chunk
991.align	4
992.Linvert:
993		mov	0(%rsi),%rax
994		mov	8(%rsi),%rbx
995		mov	0(%rdi),%rcx
996		mov	8(%rdi),%rdx
997		mov	%rax,0(%rdi)
998		mov	%rbx,8(%rdi)
999		mov	%rcx,0(%rsi)
1000		mov	%rdx,8(%rsi)
1001		lea	16(%rsi),%rsi
1002		lea	-16(%rdi),%rdi
1003		cmp	%rsi,%rdi
1004	jne	.Linvert
1005
1006	.picmeup %r9
1007	lea	AES_Td-.(%r9),%rdi
1008	lea	AES_Te-AES_Td(%rdi),%r9
1009
1010	mov	%r8,%rsi
1011	mov	240(%r8),%ecx		# pull number of rounds
1012	sub	\$1,%ecx
1013.align	4
1014.Lpermute:
1015		lea	16(%rsi),%rsi
1016___
1017		&deckey	(0,"%rsi","%r9","%rdi");
1018		&deckey	(4,"%rsi","%r9","%rdi");
1019		&deckey	(8,"%rsi","%r9","%rdi");
1020		&deckey	(12,"%rsi","%r9","%rdi");
1021$code.=<<___;
1022		sub	\$1,%ecx
1023	jnz	.Lpermute
1024
1025	xor	%rax,%rax
1026	pop	%rbx
1027	ret
1028.size	AES_set_decrypt_key,.-AES_set_decrypt_key
1029___
1030
1031# void AES_cbc_encrypt (const void char *inp, unsigned char *out,
1032#			size_t length, const AES_KEY *key,
1033#			unsigned char *ivp,const int enc);
1034{
1035# stack frame layout
1036# -8(%rsp)		return address
1037my $_rsp="0(%rsp)";		# saved %rsp
1038my $_len="8(%rsp)";		# copy of 3rd parameter, length
1039my $_key="16(%rsp)";		# copy of 4th parameter, key
1040my $_ivp="24(%rsp)";		# copy of 5th parameter, ivp
1041my $keyp="32(%rsp)";		# one to pass as $key
1042my $ivec="40(%rsp)";		# ivec[16]
1043my $aes_key="56(%rsp)";		# copy of aes_key
1044my $mark="56+240(%rsp)";	# copy of aes_key->rounds
1045
1046$code.=<<___;
1047.globl	AES_cbc_encrypt
1048.type	AES_cbc_encrypt,\@function,6
1049.align	16
1050AES_cbc_encrypt:
1051	cmp	\$0,%rdx	# check length
1052	je	.Lcbc_just_ret
1053	push	%rbx
1054	push	%rbp
1055	push	%r12
1056	push	%r13
1057	push	%r14
1058	push	%r15
1059	pushfq
1060	cld
1061	mov	%r9d,%r9d	# clear upper half of enc
1062
1063	.picmeup $sbox
1064.Lcbc_pic_point:
1065
1066	cmp	\$0,%r9
1067	je	.LDECRYPT
1068
1069	lea	AES_Te-.Lcbc_pic_point($sbox),$sbox
1070
1071	# allocate aligned stack frame...
1072	lea	-64-248(%rsp),$key
1073	and	\$-64,$key
1074
1075	# ... and make it doesn't alias with AES_Te modulo 4096
1076	mov	$sbox,%r10
1077	lea	2048($sbox),%r11
1078	mov	$key,%r12
1079	and	\$0xFFF,%r10	# s = $sbox&0xfff
1080	and	\$0xFFF,%r11	# e = ($sbox+2048)&0xfff
1081	and	\$0xFFF,%r12	# p = %rsp&0xfff
1082
1083	cmp	%r11,%r12	# if (p=>e) %rsp =- (p-e);
1084	jb	.Lcbc_te_break_out
1085	sub	%r11,%r12
1086	sub	%r12,$key
1087	jmp	.Lcbc_te_ok
1088.Lcbc_te_break_out:		# else %rsp -= (p-s)&0xfff + framesz
1089	sub	%r10,%r12
1090	and	\$0xFFF,%r12
1091	add	\$320,%r12
1092	sub	%r12,$key
1093.align	4
1094.Lcbc_te_ok:
1095
1096	xchg	%rsp,$key
1097	add	\$8,%rsp	# reserve for return address!
1098	mov	$key,$_rsp	# save %rsp
1099	mov	%rdx,$_len	# save copy of len
1100	mov	%rcx,$_key	# save copy of key
1101	mov	%r8,$_ivp	# save copy of ivp
1102	movl	\$0,$mark	# copy of aes_key->rounds = 0;
1103	mov	%r8,%rbp	# rearrange input arguments
1104	mov	%rsi,$out
1105	mov	%rdi,$inp
1106	mov	%rcx,$key
1107
1108	# do we copy key schedule to stack?
1109	mov	$key,%r10
1110	sub	$sbox,%r10
1111	and	\$0xfff,%r10
1112	cmp	\$2048,%r10
1113	jb	.Lcbc_do_ecopy
1114	cmp	\$4096-248,%r10
1115	jb	.Lcbc_skip_ecopy
1116.align	4
1117.Lcbc_do_ecopy:
1118		mov	$key,%rsi
1119		lea	$aes_key,%rdi
1120		lea	$aes_key,$key
1121		mov	\$240/8,%ecx
1122		.long	0x90A548F3	# rep movsq
1123		mov	(%rsi),%eax	# copy aes_key->rounds
1124		mov	%eax,(%rdi)
1125.Lcbc_skip_ecopy:
1126	mov	$key,$keyp	# save key pointer
1127
1128	mov	\$16,%ecx
1129.align	4
1130.Lcbc_prefetch_te:
1131		mov	0($sbox),%r10
1132		mov	32($sbox),%r11
1133		mov	64($sbox),%r12
1134		mov	96($sbox),%r13
1135		lea	128($sbox),$sbox
1136		sub	\$1,%ecx
1137	jnz	.Lcbc_prefetch_te
1138	sub	\$2048,$sbox
1139
1140	test	\$-16,%rdx		# check upon length
1141	mov	%rdx,%r10
1142	mov	0(%rbp),$s0		# load iv
1143	mov	4(%rbp),$s1
1144	mov	8(%rbp),$s2
1145	mov	12(%rbp),$s3
1146	jz	.Lcbc_enc_tail		# short input...
1147
1148.align	4
1149.Lcbc_enc_loop:
1150		xor	0($inp),$s0
1151		xor	4($inp),$s1
1152		xor	8($inp),$s2
1153		xor	12($inp),$s3
1154		mov	$inp,$ivec	# if ($verticalspin) save inp
1155
1156		mov	$keyp,$key	# restore key
1157		call	_x86_64_AES_encrypt
1158
1159		mov	$ivec,$inp	# if ($verticalspin) restore inp
1160		mov	$s0,0($out)
1161		mov	$s1,4($out)
1162		mov	$s2,8($out)
1163		mov	$s3,12($out)
1164
1165		mov	$_len,%r10
1166		lea	16($inp),$inp
1167		lea	16($out),$out
1168		sub	\$16,%r10
1169		test	\$-16,%r10
1170		mov	%r10,$_len
1171	jnz	.Lcbc_enc_loop
1172	test	\$15,%r10
1173	jnz	.Lcbc_enc_tail
1174	mov	$_ivp,%rbp	# restore ivp
1175	mov	$s0,0(%rbp)	# save ivec
1176	mov	$s1,4(%rbp)
1177	mov	$s2,8(%rbp)
1178	mov	$s3,12(%rbp)
1179
1180.align	4
1181.Lcbc_cleanup:
1182	cmpl	\$0,$mark	# was the key schedule copied?
1183	lea	$aes_key,%rdi
1184	mov	$_rsp,%rsp
1185	je	.Lcbc_exit
1186		mov	\$240/8,%ecx
1187		xor	%rax,%rax
1188		.long	0x90AB48F3	# rep stosq
1189.Lcbc_exit:
1190	popfq
1191	pop	%r15
1192	pop	%r14
1193	pop	%r13
1194	pop	%r12
1195	pop	%rbp
1196	pop	%rbx
1197.Lcbc_just_ret:
1198	ret
1199.align	4
1200.Lcbc_enc_tail:
1201	cmp	$inp,$out
1202	je	.Lcbc_enc_in_place
1203	mov	%r10,%rcx
1204	mov	$inp,%rsi
1205	mov	$out,%rdi
1206	.long	0xF689A4F3		# rep movsb
1207.Lcbc_enc_in_place:
1208	mov	\$16,%rcx		# zero tail
1209	sub	%r10,%rcx
1210	xor	%rax,%rax
1211	.long	0xF689AAF3		# rep stosb
1212	mov	$out,$inp		# this is not a mistake!
1213	movq	\$16,$_len		# len=16
1214	jmp	.Lcbc_enc_loop		# one more spin...
1215#----------------------------- DECRYPT -----------------------------#
1216.align	16
1217.LDECRYPT:
1218	lea	AES_Td-.Lcbc_pic_point($sbox),$sbox
1219
1220	# allocate aligned stack frame...
1221	lea	-64-248(%rsp),$key
1222	and	\$-64,$key
1223
1224	# ... and make it doesn't alias with AES_Td modulo 4096
1225	mov	$sbox,%r10
1226	lea	2304($sbox),%r11
1227	mov	$key,%r12
1228	and	\$0xFFF,%r10	# s = $sbox&0xfff
1229	and	\$0xFFF,%r11	# e = ($sbox+2048+256)&0xfff
1230	and	\$0xFFF,%r12	# p = %rsp&0xfff
1231
1232	cmp	%r11,%r12	# if (p=>e) %rsp =- (p-e);
1233	jb	.Lcbc_td_break_out
1234	sub	%r11,%r12
1235	sub	%r12,$key
1236	jmp	.Lcbc_td_ok
1237.Lcbc_td_break_out:		# else %rsp -= (p-s)&0xfff + framesz
1238	sub	%r10,%r12
1239	and	\$0xFFF,%r12
1240	add	\$320,%r12
1241	sub	%r12,$key
1242.align	4
1243.Lcbc_td_ok:
1244
1245	xchg	%rsp,$key
1246	add	\$8,%rsp	# reserve for return address!
1247	mov	$key,$_rsp	# save %rsp
1248	mov	%rdx,$_len	# save copy of len
1249	mov	%rcx,$_key	# save copy of key
1250	mov	%r8,$_ivp	# save copy of ivp
1251	movl	\$0,$mark	# copy of aes_key->rounds = 0;
1252	mov	%r8,%rbp	# rearrange input arguments
1253	mov	%rsi,$out
1254	mov	%rdi,$inp
1255	mov	%rcx,$key
1256
1257	# do we copy key schedule to stack?
1258	mov	$key,%r10
1259	sub	$sbox,%r10
1260	and	\$0xfff,%r10
1261	cmp	\$2304,%r10
1262	jb	.Lcbc_do_dcopy
1263	cmp	\$4096-248,%r10
1264	jb	.Lcbc_skip_dcopy
1265.align	4
1266.Lcbc_do_dcopy:
1267		mov	$key,%rsi
1268		lea	$aes_key,%rdi
1269		lea	$aes_key,$key
1270		mov	\$240/8,%ecx
1271		.long	0x90A548F3	# rep movsq
1272		mov	(%rsi),%eax	# copy aes_key->rounds
1273		mov	%eax,(%rdi)
1274.Lcbc_skip_dcopy:
1275	mov	$key,$keyp	# save key pointer
1276
1277	mov	\$18,%ecx
1278.align	4
1279.Lcbc_prefetch_td:
1280		mov	0($sbox),%r10
1281		mov	32($sbox),%r11
1282		mov	64($sbox),%r12
1283		mov	96($sbox),%r13
1284		lea	128($sbox),$sbox
1285		sub	\$1,%ecx
1286	jnz	.Lcbc_prefetch_td
1287	sub	\$2304,$sbox
1288
1289	cmp	$inp,$out
1290	je	.Lcbc_dec_in_place
1291
1292	mov	%rbp,$ivec
1293.align	4
1294.Lcbc_dec_loop:
1295		mov	0($inp),$s0		# read input
1296		mov	4($inp),$s1
1297		mov	8($inp),$s2
1298		mov	12($inp),$s3
1299		mov	$inp,8+$ivec	# if ($verticalspin) save inp
1300
1301		mov	$keyp,$key	# restore key
1302		call	_x86_64_AES_decrypt
1303
1304		mov	$ivec,%rbp	# load ivp
1305		mov	8+$ivec,$inp	# if ($verticalspin) restore inp
1306		xor	0(%rbp),$s0	# xor iv
1307		xor	4(%rbp),$s1
1308		xor	8(%rbp),$s2
1309		xor	12(%rbp),$s3
1310		mov	$inp,%rbp	# current input, next iv
1311
1312		mov	$_len,%r10	# load len
1313		sub	\$16,%r10
1314		jc	.Lcbc_dec_partial
1315		mov	%r10,$_len	# update len
1316		mov	%rbp,$ivec	# update ivp
1317
1318		mov	$s0,0($out)	# write output
1319		mov	$s1,4($out)
1320		mov	$s2,8($out)
1321		mov	$s3,12($out)
1322
1323		lea	16($inp),$inp
1324		lea	16($out),$out
1325	jnz	.Lcbc_dec_loop
1326.Lcbc_dec_end:
1327	mov	$_ivp,%r12		# load user ivp
1328	mov	0(%rbp),%r10		# load iv
1329	mov	8(%rbp),%r11
1330	mov	%r10,0(%r12)		# copy back to user
1331	mov	%r11,8(%r12)
1332	jmp	.Lcbc_cleanup
1333
1334.align	4
1335.Lcbc_dec_partial:
1336	mov	$s0,0+$ivec		# dump output to stack
1337	mov	$s1,4+$ivec
1338	mov	$s2,8+$ivec
1339	mov	$s3,12+$ivec
1340	mov	$out,%rdi
1341	lea	$ivec,%rsi
1342	mov	\$16,%rcx
1343	add	%r10,%rcx		# number of bytes to copy
1344	.long	0xF689A4F3		# rep movsb
1345	jmp	.Lcbc_dec_end
1346
1347.align	16
1348.Lcbc_dec_in_place:
1349		mov	0($inp),$s0	# load input
1350		mov	4($inp),$s1
1351		mov	8($inp),$s2
1352		mov	12($inp),$s3
1353
1354		mov	$inp,$ivec	# if ($verticalspin) save inp
1355		mov	$keyp,$key
1356		call	_x86_64_AES_decrypt
1357
1358		mov	$ivec,$inp	# if ($verticalspin) restore inp
1359		mov	$_ivp,%rbp
1360		xor	0(%rbp),$s0
1361		xor	4(%rbp),$s1
1362		xor	8(%rbp),$s2
1363		xor	12(%rbp),$s3
1364
1365		mov	0($inp),%r10	# copy input to iv
1366		mov	8($inp),%r11
1367		mov	%r10,0(%rbp)
1368		mov	%r11,8(%rbp)
1369
1370		mov	$s0,0($out)	# save output [zaps input]
1371		mov	$s1,4($out)
1372		mov	$s2,8($out)
1373		mov	$s3,12($out)
1374
1375		mov	$_len,%rcx
1376		lea	16($inp),$inp
1377		lea	16($out),$out
1378		sub	\$16,%rcx
1379		jc	.Lcbc_dec_in_place_partial
1380		mov	%rcx,$_len
1381	jnz	.Lcbc_dec_in_place
1382	jmp	.Lcbc_cleanup
1383
1384.align	4
1385.Lcbc_dec_in_place_partial:
1386	# one can argue if this is actually required
1387	lea	($out,%rcx),%rdi
1388	lea	(%rbp,%rcx),%rsi
1389	neg	%rcx
1390	.long	0xF689A4F3	# rep movsb	# restore tail
1391	jmp	.Lcbc_cleanup
1392.size	AES_cbc_encrypt,.-AES_cbc_encrypt
1393___
1394}
1395
1396$code.=<<___;
1397.globl	AES_Te
1398.align	64
1399AES_Te:
1400___
1401	&_data_word(0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6);
1402	&_data_word(0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591);
1403	&_data_word(0x50303060, 0x03010102, 0xa96767ce, 0x7d2b2b56);
1404	&_data_word(0x19fefee7, 0x62d7d7b5, 0xe6abab4d, 0x9a7676ec);
1405	&_data_word(0x45caca8f, 0x9d82821f, 0x40c9c989, 0x877d7dfa);
1406	&_data_word(0x15fafaef, 0xeb5959b2, 0xc947478e, 0x0bf0f0fb);
1407	&_data_word(0xecadad41, 0x67d4d4b3, 0xfda2a25f, 0xeaafaf45);
1408	&_data_word(0xbf9c9c23, 0xf7a4a453, 0x967272e4, 0x5bc0c09b);
1409	&_data_word(0xc2b7b775, 0x1cfdfde1, 0xae93933d, 0x6a26264c);
1410	&_data_word(0x5a36366c, 0x413f3f7e, 0x02f7f7f5, 0x4fcccc83);
1411	&_data_word(0x5c343468, 0xf4a5a551, 0x34e5e5d1, 0x08f1f1f9);
1412	&_data_word(0x937171e2, 0x73d8d8ab, 0x53313162, 0x3f15152a);
1413	&_data_word(0x0c040408, 0x52c7c795, 0x65232346, 0x5ec3c39d);
1414	&_data_word(0x28181830, 0xa1969637, 0x0f05050a, 0xb59a9a2f);
1415	&_data_word(0x0907070e, 0x36121224, 0x9b80801b, 0x3de2e2df);
1416	&_data_word(0x26ebebcd, 0x6927274e, 0xcdb2b27f, 0x9f7575ea);
1417	&_data_word(0x1b090912, 0x9e83831d, 0x742c2c58, 0x2e1a1a34);
1418	&_data_word(0x2d1b1b36, 0xb26e6edc, 0xee5a5ab4, 0xfba0a05b);
1419	&_data_word(0xf65252a4, 0x4d3b3b76, 0x61d6d6b7, 0xceb3b37d);
1420	&_data_word(0x7b292952, 0x3ee3e3dd, 0x712f2f5e, 0x97848413);
1421	&_data_word(0xf55353a6, 0x68d1d1b9, 0x00000000, 0x2cededc1);
1422	&_data_word(0x60202040, 0x1ffcfce3, 0xc8b1b179, 0xed5b5bb6);
1423	&_data_word(0xbe6a6ad4, 0x46cbcb8d, 0xd9bebe67, 0x4b393972);
1424	&_data_word(0xde4a4a94, 0xd44c4c98, 0xe85858b0, 0x4acfcf85);
1425	&_data_word(0x6bd0d0bb, 0x2aefefc5, 0xe5aaaa4f, 0x16fbfbed);
1426	&_data_word(0xc5434386, 0xd74d4d9a, 0x55333366, 0x94858511);
1427	&_data_word(0xcf45458a, 0x10f9f9e9, 0x06020204, 0x817f7ffe);
1428	&_data_word(0xf05050a0, 0x443c3c78, 0xba9f9f25, 0xe3a8a84b);
1429	&_data_word(0xf35151a2, 0xfea3a35d, 0xc0404080, 0x8a8f8f05);
1430	&_data_word(0xad92923f, 0xbc9d9d21, 0x48383870, 0x04f5f5f1);
1431	&_data_word(0xdfbcbc63, 0xc1b6b677, 0x75dadaaf, 0x63212142);
1432	&_data_word(0x30101020, 0x1affffe5, 0x0ef3f3fd, 0x6dd2d2bf);
1433	&_data_word(0x4ccdcd81, 0x140c0c18, 0x35131326, 0x2fececc3);
1434	&_data_word(0xe15f5fbe, 0xa2979735, 0xcc444488, 0x3917172e);
1435	&_data_word(0x57c4c493, 0xf2a7a755, 0x827e7efc, 0x473d3d7a);
1436	&_data_word(0xac6464c8, 0xe75d5dba, 0x2b191932, 0x957373e6);
1437	&_data_word(0xa06060c0, 0x98818119, 0xd14f4f9e, 0x7fdcdca3);
1438	&_data_word(0x66222244, 0x7e2a2a54, 0xab90903b, 0x8388880b);
1439	&_data_word(0xca46468c, 0x29eeeec7, 0xd3b8b86b, 0x3c141428);
1440	&_data_word(0x79dedea7, 0xe25e5ebc, 0x1d0b0b16, 0x76dbdbad);
1441	&_data_word(0x3be0e0db, 0x56323264, 0x4e3a3a74, 0x1e0a0a14);
1442	&_data_word(0xdb494992, 0x0a06060c, 0x6c242448, 0xe45c5cb8);
1443	&_data_word(0x5dc2c29f, 0x6ed3d3bd, 0xefacac43, 0xa66262c4);
1444	&_data_word(0xa8919139, 0xa4959531, 0x37e4e4d3, 0x8b7979f2);
1445	&_data_word(0x32e7e7d5, 0x43c8c88b, 0x5937376e, 0xb76d6dda);
1446	&_data_word(0x8c8d8d01, 0x64d5d5b1, 0xd24e4e9c, 0xe0a9a949);
1447	&_data_word(0xb46c6cd8, 0xfa5656ac, 0x07f4f4f3, 0x25eaeacf);
1448	&_data_word(0xaf6565ca, 0x8e7a7af4, 0xe9aeae47, 0x18080810);
1449	&_data_word(0xd5baba6f, 0x887878f0, 0x6f25254a, 0x722e2e5c);
1450	&_data_word(0x241c1c38, 0xf1a6a657, 0xc7b4b473, 0x51c6c697);
1451	&_data_word(0x23e8e8cb, 0x7cdddda1, 0x9c7474e8, 0x211f1f3e);
1452	&_data_word(0xdd4b4b96, 0xdcbdbd61, 0x868b8b0d, 0x858a8a0f);
1453	&_data_word(0x907070e0, 0x423e3e7c, 0xc4b5b571, 0xaa6666cc);
1454	&_data_word(0xd8484890, 0x05030306, 0x01f6f6f7, 0x120e0e1c);
1455	&_data_word(0xa36161c2, 0x5f35356a, 0xf95757ae, 0xd0b9b969);
1456	&_data_word(0x91868617, 0x58c1c199, 0x271d1d3a, 0xb99e9e27);
1457	&_data_word(0x38e1e1d9, 0x13f8f8eb, 0xb398982b, 0x33111122);
1458	&_data_word(0xbb6969d2, 0x70d9d9a9, 0x898e8e07, 0xa7949433);
1459	&_data_word(0xb69b9b2d, 0x221e1e3c, 0x92878715, 0x20e9e9c9);
1460	&_data_word(0x49cece87, 0xff5555aa, 0x78282850, 0x7adfdfa5);
1461	&_data_word(0x8f8c8c03, 0xf8a1a159, 0x80898909, 0x170d0d1a);
1462	&_data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0);
1463	&_data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e);
1464	&_data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c);
1465#rcon:
1466$code.=<<___;
1467	.long	0x00000001, 0x00000002, 0x00000004, 0x00000008
1468	.long	0x00000010, 0x00000020, 0x00000040, 0x00000080
1469	.long	0x0000001b, 0x00000036, 0, 0, 0, 0, 0, 0
1470___
1471$code.=<<___;
1472.globl	AES_Td
1473.align	64
1474AES_Td:
1475___
1476	&_data_word(0x50a7f451, 0x5365417e, 0xc3a4171a, 0x965e273a);
1477	&_data_word(0xcb6bab3b, 0xf1459d1f, 0xab58faac, 0x9303e34b);
1478	&_data_word(0x55fa3020, 0xf66d76ad, 0x9176cc88, 0x254c02f5);
1479	&_data_word(0xfcd7e54f, 0xd7cb2ac5, 0x80443526, 0x8fa362b5);
1480	&_data_word(0x495ab1de, 0x671bba25, 0x980eea45, 0xe1c0fe5d);
1481	&_data_word(0x02752fc3, 0x12f04c81, 0xa397468d, 0xc6f9d36b);
1482	&_data_word(0xe75f8f03, 0x959c9215, 0xeb7a6dbf, 0xda595295);
1483	&_data_word(0x2d83bed4, 0xd3217458, 0x2969e049, 0x44c8c98e);
1484	&_data_word(0x6a89c275, 0x78798ef4, 0x6b3e5899, 0xdd71b927);
1485	&_data_word(0xb64fe1be, 0x17ad88f0, 0x66ac20c9, 0xb43ace7d);
1486	&_data_word(0x184adf63, 0x82311ae5, 0x60335197, 0x457f5362);
1487	&_data_word(0xe07764b1, 0x84ae6bbb, 0x1ca081fe, 0x942b08f9);
1488	&_data_word(0x58684870, 0x19fd458f, 0x876cde94, 0xb7f87b52);
1489	&_data_word(0x23d373ab, 0xe2024b72, 0x578f1fe3, 0x2aab5566);
1490	&_data_word(0x0728ebb2, 0x03c2b52f, 0x9a7bc586, 0xa50837d3);
1491	&_data_word(0xf2872830, 0xb2a5bf23, 0xba6a0302, 0x5c8216ed);
1492	&_data_word(0x2b1ccf8a, 0x92b479a7, 0xf0f207f3, 0xa1e2694e);
1493	&_data_word(0xcdf4da65, 0xd5be0506, 0x1f6234d1, 0x8afea6c4);
1494	&_data_word(0x9d532e34, 0xa055f3a2, 0x32e18a05, 0x75ebf6a4);
1495	&_data_word(0x39ec830b, 0xaaef6040, 0x069f715e, 0x51106ebd);
1496	&_data_word(0xf98a213e, 0x3d06dd96, 0xae053edd, 0x46bde64d);
1497	&_data_word(0xb58d5491, 0x055dc471, 0x6fd40604, 0xff155060);
1498	&_data_word(0x24fb9819, 0x97e9bdd6, 0xcc434089, 0x779ed967);
1499	&_data_word(0xbd42e8b0, 0x888b8907, 0x385b19e7, 0xdbeec879);
1500	&_data_word(0x470a7ca1, 0xe90f427c, 0xc91e84f8, 0x00000000);
1501	&_data_word(0x83868009, 0x48ed2b32, 0xac70111e, 0x4e725a6c);
1502	&_data_word(0xfbff0efd, 0x5638850f, 0x1ed5ae3d, 0x27392d36);
1503	&_data_word(0x64d90f0a, 0x21a65c68, 0xd1545b9b, 0x3a2e3624);
1504	&_data_word(0xb1670a0c, 0x0fe75793, 0xd296eeb4, 0x9e919b1b);
1505	&_data_word(0x4fc5c080, 0xa220dc61, 0x694b775a, 0x161a121c);
1506	&_data_word(0x0aba93e2, 0xe52aa0c0, 0x43e0223c, 0x1d171b12);
1507	&_data_word(0x0b0d090e, 0xadc78bf2, 0xb9a8b62d, 0xc8a91e14);
1508	&_data_word(0x8519f157, 0x4c0775af, 0xbbdd99ee, 0xfd607fa3);
1509	&_data_word(0x9f2601f7, 0xbcf5725c, 0xc53b6644, 0x347efb5b);
1510	&_data_word(0x7629438b, 0xdcc623cb, 0x68fcedb6, 0x63f1e4b8);
1511	&_data_word(0xcadc31d7, 0x10856342, 0x40229713, 0x2011c684);
1512	&_data_word(0x7d244a85, 0xf83dbbd2, 0x1132f9ae, 0x6da129c7);
1513	&_data_word(0x4b2f9e1d, 0xf330b2dc, 0xec52860d, 0xd0e3c177);
1514	&_data_word(0x6c16b32b, 0x99b970a9, 0xfa489411, 0x2264e947);
1515	&_data_word(0xc48cfca8, 0x1a3ff0a0, 0xd82c7d56, 0xef903322);
1516	&_data_word(0xc74e4987, 0xc1d138d9, 0xfea2ca8c, 0x360bd498);
1517	&_data_word(0xcf81f5a6, 0x28de7aa5, 0x268eb7da, 0xa4bfad3f);
1518	&_data_word(0xe49d3a2c, 0x0d927850, 0x9bcc5f6a, 0x62467e54);
1519	&_data_word(0xc2138df6, 0xe8b8d890, 0x5ef7392e, 0xf5afc382);
1520	&_data_word(0xbe805d9f, 0x7c93d069, 0xa92dd56f, 0xb31225cf);
1521	&_data_word(0x3b99acc8, 0xa77d1810, 0x6e639ce8, 0x7bbb3bdb);
1522	&_data_word(0x097826cd, 0xf418596e, 0x01b79aec, 0xa89a4f83);
1523	&_data_word(0x656e95e6, 0x7ee6ffaa, 0x08cfbc21, 0xe6e815ef);
1524	&_data_word(0xd99be7ba, 0xce366f4a, 0xd4099fea, 0xd67cb029);
1525	&_data_word(0xafb2a431, 0x31233f2a, 0x3094a5c6, 0xc066a235);
1526	&_data_word(0x37bc4e74, 0xa6ca82fc, 0xb0d090e0, 0x15d8a733);
1527	&_data_word(0x4a9804f1, 0xf7daec41, 0x0e50cd7f, 0x2ff69117);
1528	&_data_word(0x8dd64d76, 0x4db0ef43, 0x544daacc, 0xdf0496e4);
1529	&_data_word(0xe3b5d19e, 0x1b886a4c, 0xb81f2cc1, 0x7f516546);
1530	&_data_word(0x04ea5e9d, 0x5d358c01, 0x737487fa, 0x2e410bfb);
1531	&_data_word(0x5a1d67b3, 0x52d2db92, 0x335610e9, 0x1347d66d);
1532	&_data_word(0x8c61d79a, 0x7a0ca137, 0x8e14f859, 0x893c13eb);
1533	&_data_word(0xee27a9ce, 0x35c961b7, 0xede51ce1, 0x3cb1477a);
1534	&_data_word(0x59dfd29c, 0x3f73f255, 0x79ce1418, 0xbf37c773);
1535	&_data_word(0xeacdf753, 0x5baafd5f, 0x146f3ddf, 0x86db4478);
1536	&_data_word(0x81f3afca, 0x3ec468b9, 0x2c342438, 0x5f40a3c2);
1537	&_data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff);
1538	&_data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664);
1539	&_data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0);
1540#Td4:
1541	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1542	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1543	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1544	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1545	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1546	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1547	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1548	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1549	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1550	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1551	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1552	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1553	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1554	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1555	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1556	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1557	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1558	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1559	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1560	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1561	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1562	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1563	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1564	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1565	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1566	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1567	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1568	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1569	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1570	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1571	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1572	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1573
1574$code =~ s/\`([^\`]*)\`/eval($1)/gem;
1575
1576print $code;
1577
1578close STDOUT;
1579