aes-mips.pl@ 99371

最後變更在這個檔案從99371是 99366,由 vboxsync 提交於 23 月前
openssl-3.1.0: Applied and adjusted our OpenSSL changes to 3.0.7. bugref:10418
檔案大小: 53.2 KB

行
1	#! /usr/bin/env perl
2	# Copyright 2010-2020 The OpenSSL Project Authors. All Rights Reserved.
3	#
4	# Licensed under the Apache License 2.0 (the "License"). You may not use
5	# this file except in compliance with the License. You can obtain a copy
6	# in the file LICENSE in the source distribution or at
7	# https://www.openssl.org/source/license.html
8
9
10	# ====================================================================
11	# Written by Andy Polyakov <[email protected]> for the OpenSSL
12	# project. The module is, however, dual licensed under OpenSSL and
13	# CRYPTOGAMS licenses depending on where you obtain it. For further
14	# details see http://www.openssl.org/~appro/cryptogams/.
15	# ====================================================================
16
17	# AES for MIPS
18
19	# October 2010
20	#
21	# Code uses 1K[+256B] S-box and on single-issue core [such as R5000]
22	# spends ~68 cycles per byte processed with 128-bit key. This is ~16%
23	# faster than gcc-generated code, which is not very impressive. But
24	# recall that compressed S-box requires extra processing, namely
25	# additional rotations. Rotations are implemented with lwl/lwr pairs,
26	# which is normally used for loading unaligned data. Another cool
27	# thing about this module is its endian neutrality, which means that
28	# it processes data without ever changing byte order...
29
30	# September 2012
31	#
32	# Add MIPS32R2 (~10% less instructions) and SmartMIPS ASE (further
33	# ~25% less instructions) code. Note that there is no run-time switch,
34	# instead, code path is chosen upon pre-process time, pass -mips32r2
35	# or/and -msmartmips.
36
37	# February 2019
38	#
39	# Normalize MIPS32R2 AES table address calculation by always using EXT
40	# instruction. This reduces the standard codebase by another 10%.
41
42	######################################################################
43	# There is a number of MIPS ABI in use, O32 and N32/64 are most
44	# widely used. Then there is a new contender: NUBI. It appears that if
45	# one picks the latter, it's possible to arrange code in ABI neutral
46	# manner. Therefore let's stick to NUBI register layout:
47	#
48	($zero,$at,$t0,$t1,$t2)=map("\$$_",(0..2,24,25));
49	($a0,$a1,$a2,$a3,$a4,$a5,$a6,$a7)=map("\$$_",(4..11));
50	($s0,$s1,$s2,$s3,$s4,$s5,$s6,$s7,$s8,$s9,$s10,$s11)=map("\$$_",(12..23));
51	($gp,$tp,$sp,$fp,$ra)=map("\$$_",(3,28..31));
52	#
53	# The return value is placed in $a0. Following coding rules facilitate
54	# interoperability:
55	#
56	# - never ever touch $tp, "thread pointer", former $gp;
57	# - copy return value to $t0, former $v0 [or to $a0 if you're adapting
58	# old code];
59	# - on O32 populate $a4-$a7 with 'lw $aN,4*N($sp)' if necessary;
60	#
61	# For reference here is register layout for N32/64 MIPS ABIs:
62	#
63	# ($zero,$at,$v0,$v1)=map("\$$_",(0..3));
64	# ($a0,$a1,$a2,$a3,$a4,$a5,$a6,$a7)=map("\$$_",(4..11));
65	# ($t0,$t1,$t2,$t3,$t8,$t9)=map("\$$_",(12..15,24,25));
66	# ($s0,$s1,$s2,$s3,$s4,$s5,$s6,$s7)=map("\$$_",(16..23));
67	# ($gp,$sp,$fp,$ra)=map("\$$_",(28..31));
68
69	# $output is the last argument if it looks like a file (it has an extension)
70	# $flavour is the first argument if it doesn't look like a file
71	$output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m\|\.\w+$\| ? pop : undef;
72	$flavour = $#ARGV >= 0 && $ARGV[0] !~ m\|\.\| ? shift : undef;
73	$flavour \|\|= "o32"; # supported flavours are o32,n32,64,nubi32,nubi64
74
75	if ($flavour =~ /64\|n32/i) {
76	$PTR_LA="dla";
77	$PTR_ADD="daddu"; # incidentally works even on n32
78	$PTR_SUB="dsubu"; # incidentally works even on n32
79	$PTR_INS="dins";
80	$REG_S="sd";
81	$REG_L="ld";
82	$PTR_SLL="dsll"; # incidentally works even on n32
83	$SZREG=8;
84	} else {
85	$PTR_LA="la";
86	$PTR_ADD="addu";
87	$PTR_SUB="subu";
88	$PTR_INS="ins";
89	$REG_S="sw";
90	$REG_L="lw";
91	$PTR_SLL="sll";
92	$SZREG=4;
93	}
94	$pf = ($flavour =~ /nubi/i) ? $t0 : $t2;
95	#
96	# <[email protected]>
97	#
98	######################################################################
99
100	$big_endian=(`echo MIPSEB \| $ENV{CC} -E -`=~/MIPSEB/)?0:1 if ($ENV{CC});
101
102	if (!defined($big_endian))
103	{ $big_endian=(unpack('L',pack('N',1))==1); }
104
105	my ($MSB,$LSB)=(0,3); # automatically converted to little-endian
106
107	$output and open STDOUT,">$output";
108
109	$code.=<<___;
110	#include "mips_arch.h"
111
112	.text
113	#if !defined(__mips_eabi) && (!defined(__vxworks) \|\| defined(__pic__))
114	.option pic2
115	#endif
116	.set noat
117	___
118
119
120	{{{
121	my $FRAMESIZE=16*$SZREG;
122	my $SAVED_REGS_MASK = ($flavour =~ /nubi/i) ? "0xc0fff008" : "0xc0ff0000";
123
124	my ($inp,$out,$key,$Tbl,$s0,$s1,$s2,$s3)=($a0,$a1,$a2,$a3,$a4,$a5,$a6,$a7);
125	my ($i0,$i1,$i2,$i3)=($at,$t0,$t1,$t2);
126	my ($t0,$t1,$t2,$t3,$t4,$t5,$t6,$t7,$t8,$t9,$t10,$t11) = map("\$$_",(12..23));
127	my ($key0,$cnt)=($gp,$fp);
128
129	# instruction ordering is "stolen" from output from MIPSpro assembler
130	# invoked with -mips3 -O3 arguments...
131	$code.=<<___;
132	.align 5
133	.ent _mips_AES_encrypt
134	_mips_AES_encrypt:
135	.frame $sp,0,$ra
136	.set reorder
137	lw $t0,0($key)
138	lw $t1,4($key)
139	lw $t2,8($key)
140	lw $t3,12($key)
141	lw $cnt,240($key)
142	$PTR_ADD $key0,$key,16
143
144	xor $s0,$t0
145	xor $s1,$t1
146	xor $s2,$t2
147	xor $s3,$t3
148
149	subu $cnt,1
150	#if defined(__mips_smartmips)
151	ext $i0,$s1,16,8
152	.Loop_enc:
153	ext $i1,$s2,16,8
154	ext $i2,$s3,16,8
155	ext $i3,$s0,16,8
156	lwxs $t0,$i0($Tbl) # Te1[s1>>16]
157	ext $i0,$s2,8,8
158	lwxs $t1,$i1($Tbl) # Te1[s2>>16]
159	ext $i1,$s3,8,8
160	lwxs $t2,$i2($Tbl) # Te1[s3>>16]
161	ext $i2,$s0,8,8
162	lwxs $t3,$i3($Tbl) # Te1[s0>>16]
163	ext $i3,$s1,8,8
164
165	lwxs $t4,$i0($Tbl) # Te2[s2>>8]
166	ext $i0,$s3,0,8
167	lwxs $t5,$i1($Tbl) # Te2[s3>>8]
168	ext $i1,$s0,0,8
169	lwxs $t6,$i2($Tbl) # Te2[s0>>8]
170	ext $i2,$s1,0,8
171	lwxs $t7,$i3($Tbl) # Te2[s1>>8]
172	ext $i3,$s2,0,8
173
174	lwxs $t8,$i0($Tbl) # Te3[s3]
175	ext $i0,$s0,24,8
176	lwxs $t9,$i1($Tbl) # Te3[s0]
177	ext $i1,$s1,24,8
178	lwxs $t10,$i2($Tbl) # Te3[s1]
179	ext $i2,$s2,24,8
180	lwxs $t11,$i3($Tbl) # Te3[s2]
181	ext $i3,$s3,24,8
182
183	rotr $t0,$t0,8
184	rotr $t1,$t1,8
185	rotr $t2,$t2,8
186	rotr $t3,$t3,8
187
188	rotr $t4,$t4,16
189	rotr $t5,$t5,16
190	rotr $t6,$t6,16
191	rotr $t7,$t7,16
192
193	xor $t0,$t4
194	lwxs $t4,$i0($Tbl) # Te0[s0>>24]
195	xor $t1,$t5
196	lwxs $t5,$i1($Tbl) # Te0[s1>>24]
197	xor $t2,$t6
198	lwxs $t6,$i2($Tbl) # Te0[s2>>24]
199	xor $t3,$t7
200	lwxs $t7,$i3($Tbl) # Te0[s3>>24]
201
202	rotr $t8,$t8,24
203	lw $s0,0($key0)
204	rotr $t9,$t9,24
205	lw $s1,4($key0)
206	rotr $t10,$t10,24
207	lw $s2,8($key0)
208	rotr $t11,$t11,24
209	lw $s3,12($key0)
210
211	xor $t0,$t8
212	xor $t1,$t9
213	xor $t2,$t10
214	xor $t3,$t11
215
216	xor $t0,$t4
217	xor $t1,$t5
218	xor $t2,$t6
219	xor $t3,$t7
220
221	subu $cnt,1
222	$PTR_ADD $key0,16
223	xor $s0,$t0
224	xor $s1,$t1
225	xor $s2,$t2
226	xor $s3,$t3
227	.set noreorder
228	bnez $cnt,.Loop_enc
229	ext $i0,$s1,16,8
230
231	_xtr $i0,$s1,16-2
232	#else
233	#if defined(_MIPS_ARCH_MIPS32R2) \|\| defined(_MIPS_ARCH_MIPS64R2)
234	move $i0,$Tbl
235	move $i1,$Tbl
236	move $i2,$Tbl
237	move $i3,$Tbl
238	ext $t0,$s1,16,8
239	.Loop_enc:
240	ext $t1,$s2,16,8
241	ext $t2,$s3,16,8
242	ext $t3,$s0,16,8
243	$PTR_INS $i0,$t0,2,8
244	$PTR_INS $i1,$t1,2,8
245	$PTR_INS $i2,$t2,2,8
246	$PTR_INS $i3,$t3,2,8
247	lw $t0,0($i0) # Te1[s1>>16]
248	ext $t4,$s2,8,8
249	lw $t1,0($i1) # Te1[s2>>16]
250	ext $t5,$s3,8,8
251	lw $t2,0($i2) # Te1[s3>>16]
252	ext $t6,$s0,8,8
253	lw $t3,0($i3) # Te1[s0>>16]
254	ext $t7,$s1,8,8
255	$PTR_INS $i0,$t4,2,8
256	$PTR_INS $i1,$t5,2,8
257	$PTR_INS $i2,$t6,2,8
258	$PTR_INS $i3,$t7,2,8
259	#else
260	_xtr $i0,$s1,16-2
261	.Loop_enc:
262	_xtr $i1,$s2,16-2
263	_xtr $i2,$s3,16-2
264	_xtr $i3,$s0,16-2
265	and $i0,0x3fc
266	and $i1,0x3fc
267	and $i2,0x3fc
268	and $i3,0x3fc
269	$PTR_ADD $i0,$Tbl
270	$PTR_ADD $i1,$Tbl
271	$PTR_ADD $i2,$Tbl
272	$PTR_ADD $i3,$Tbl
273	lwl $t0,3($i0) # Te1[s1>>16]
274	lwl $t1,3($i1) # Te1[s2>>16]
275	lwl $t2,3($i2) # Te1[s3>>16]
276	lwl $t3,3($i3) # Te1[s0>>16]
277	lwr $t0,2($i0) # Te1[s1>>16]
278	_xtr $i0,$s2,8-2
279	lwr $t1,2($i1) # Te1[s2>>16]
280	_xtr $i1,$s3,8-2
281	lwr $t2,2($i2) # Te1[s3>>16]
282	_xtr $i2,$s0,8-2
283	lwr $t3,2($i3) # Te1[s0>>16]
284	_xtr $i3,$s1,8-2
285	and $i0,0x3fc
286	and $i1,0x3fc
287	and $i2,0x3fc
288	and $i3,0x3fc
289	$PTR_ADD $i0,$Tbl
290	$PTR_ADD $i1,$Tbl
291	$PTR_ADD $i2,$Tbl
292	$PTR_ADD $i3,$Tbl
293	#endif
294	#if defined(_MIPS_ARCH_MIPS32R2) \|\| defined(_MIPS_ARCH_MIPS64R2)
295	rotr $t0,$t0,8
296	rotr $t1,$t1,8
297	rotr $t2,$t2,8
298	rotr $t3,$t3,8
299	# if defined(_MIPSEL)
300	lw $t4,0($i0) # Te2[s2>>8]
301	ext $t8,$s3,0,8
302	lw $t5,0($i1) # Te2[s3>>8]
303	ext $t9,$s0,0,8
304	lw $t6,0($i2) # Te2[s0>>8]
305	ext $t10,$s1,0,8
306	lw $t7,0($i3) # Te2[s1>>8]
307	ext $t11,$s2,0,8
308	$PTR_INS $i0,$t8,2,8
309	$PTR_INS $i1,$t9,2,8
310	$PTR_INS $i2,$t10,2,8
311	$PTR_INS $i3,$t11,2,8
312
313	lw $t8,0($i0) # Te3[s3]
314	$PTR_INS $i0,$s0,2,8
315	lw $t9,0($i1) # Te3[s0]
316	$PTR_INS $i1,$s1,2,8
317	lw $t10,0($i2) # Te3[s1]
318	$PTR_INS $i2,$s2,2,8
319	lw $t11,0($i3) # Te3[s2]
320	$PTR_INS $i3,$s3,2,8
321	# else
322	lw $t4,0($i0) # Te2[s2>>8]
323	$PTR_INS $i0,$s3,2,8
324	lw $t5,0($i1) # Te2[s3>>8]
325	$PTR_INS $i1,$s0,2,8
326	lw $t6,0($i2) # Te2[s0>>8]
327	$PTR_INS $i2,$s1,2,8
328	lw $t7,0($i3) # Te2[s1>>8]
329	$PTR_INS $i3,$s2,2,8
330
331	lw $t8,0($i0) # Te3[s3]
332	_xtr $i0,$s0,24-2
333	lw $t9,0($i1) # Te3[s0]
334	_xtr $i1,$s1,24-2
335	lw $t10,0($i2) # Te3[s1]
336	_xtr $i2,$s2,24-2
337	lw $t11,0($i3) # Te3[s2]
338	_xtr $i3,$s3,24-2
339
340	and $i0,0x3fc
341	and $i1,0x3fc
342	and $i2,0x3fc
343	and $i3,0x3fc
344	$PTR_ADD $i0,$Tbl
345	$PTR_ADD $i1,$Tbl
346	$PTR_ADD $i2,$Tbl
347	$PTR_ADD $i3,$Tbl
348	# endif
349	rotr $t4,$t4,16
350	rotr $t5,$t5,16
351	rotr $t6,$t6,16
352	rotr $t7,$t7,16
353
354	rotr $t8,$t8,24
355	rotr $t9,$t9,24
356	rotr $t10,$t10,24
357	rotr $t11,$t11,24
358	#else
359	lwl $t4,2($i0) # Te2[s2>>8]
360	lwl $t5,2($i1) # Te2[s3>>8]
361	lwl $t6,2($i2) # Te2[s0>>8]
362	lwl $t7,2($i3) # Te2[s1>>8]
363	lwr $t4,1($i0) # Te2[s2>>8]
364	_xtr $i0,$s3,0-2
365	lwr $t5,1($i1) # Te2[s3>>8]
366	_xtr $i1,$s0,0-2
367	lwr $t6,1($i2) # Te2[s0>>8]
368	_xtr $i2,$s1,0-2
369	lwr $t7,1($i3) # Te2[s1>>8]
370	_xtr $i3,$s2,0-2
371
372	and $i0,0x3fc
373	and $i1,0x3fc
374	and $i2,0x3fc
375	and $i3,0x3fc
376	$PTR_ADD $i0,$Tbl
377	$PTR_ADD $i1,$Tbl
378	$PTR_ADD $i2,$Tbl
379	$PTR_ADD $i3,$Tbl
380	lwl $t8,1($i0) # Te3[s3]
381	lwl $t9,1($i1) # Te3[s0]
382	lwl $t10,1($i2) # Te3[s1]
383	lwl $t11,1($i3) # Te3[s2]
384	lwr $t8,0($i0) # Te3[s3]
385	_xtr $i0,$s0,24-2
386	lwr $t9,0($i1) # Te3[s0]
387	_xtr $i1,$s1,24-2
388	lwr $t10,0($i2) # Te3[s1]
389	_xtr $i2,$s2,24-2
390	lwr $t11,0($i3) # Te3[s2]
391	_xtr $i3,$s3,24-2
392
393	and $i0,0x3fc
394	and $i1,0x3fc
395	and $i2,0x3fc
396	and $i3,0x3fc
397	$PTR_ADD $i0,$Tbl
398	$PTR_ADD $i1,$Tbl
399	$PTR_ADD $i2,$Tbl
400	$PTR_ADD $i3,$Tbl
401	#endif
402	xor $t0,$t4
403	lw $t4,0($i0) # Te0[s0>>24]
404	xor $t1,$t5
405	lw $t5,0($i1) # Te0[s1>>24]
406	xor $t2,$t6
407	lw $t6,0($i2) # Te0[s2>>24]
408	xor $t3,$t7
409	lw $t7,0($i3) # Te0[s3>>24]
410
411	xor $t0,$t8
412	lw $s0,0($key0)
413	xor $t1,$t9
414	lw $s1,4($key0)
415	xor $t2,$t10
416	lw $s2,8($key0)
417	xor $t3,$t11
418	lw $s3,12($key0)
419
420	xor $t0,$t4
421	xor $t1,$t5
422	xor $t2,$t6
423	xor $t3,$t7
424
425	subu $cnt,1
426	$PTR_ADD $key0,16
427	xor $s0,$t0
428	xor $s1,$t1
429	xor $s2,$t2
430	xor $s3,$t3
431	.set noreorder
432	bnez $cnt,.Loop_enc
433	#if defined(_MIPS_ARCH_MIPS32R2) \|\| defined(_MIPS_ARCH_MIPS64R2)
434	ext $t0,$s1,16,8
435	#endif
436	_xtr $i0,$s1,16-2
437	#endif
438
439	.set reorder
440	_xtr $i1,$s2,16-2
441	_xtr $i2,$s3,16-2
442	_xtr $i3,$s0,16-2
443	and $i0,0x3fc
444	and $i1,0x3fc
445	and $i2,0x3fc
446	and $i3,0x3fc
447	$PTR_ADD $i0,$Tbl
448	$PTR_ADD $i1,$Tbl
449	$PTR_ADD $i2,$Tbl
450	$PTR_ADD $i3,$Tbl
451	lbu $t0,2($i0) # Te4[s1>>16]
452	_xtr $i0,$s2,8-2
453	lbu $t1,2($i1) # Te4[s2>>16]
454	_xtr $i1,$s3,8-2
455	lbu $t2,2($i2) # Te4[s3>>16]
456	_xtr $i2,$s0,8-2
457	lbu $t3,2($i3) # Te4[s0>>16]
458	_xtr $i3,$s1,8-2
459
460	and $i0,0x3fc
461	and $i1,0x3fc
462	and $i2,0x3fc
463	and $i3,0x3fc
464	$PTR_ADD $i0,$Tbl
465	$PTR_ADD $i1,$Tbl
466	$PTR_ADD $i2,$Tbl
467	$PTR_ADD $i3,$Tbl
468	#if defined(_MIPS_ARCH_MIPS32R2) \|\| defined(_MIPS_ARCH_MIPS64R2)
469	# if defined(_MIPSEL)
470	lbu $t4,2($i0) # Te4[s2>>8]
471	$PTR_INS $i0,$s0,2,8
472	lbu $t5,2($i1) # Te4[s3>>8]
473	$PTR_INS $i1,$s1,2,8
474	lbu $t6,2($i2) # Te4[s0>>8]
475	$PTR_INS $i2,$s2,2,8
476	lbu $t7,2($i3) # Te4[s1>>8]
477	$PTR_INS $i3,$s3,2,8
478
479	lbu $t8,2($i0) # Te4[s0>>24]
480	_xtr $i0,$s3,0-2
481	lbu $t9,2($i1) # Te4[s1>>24]
482	_xtr $i1,$s0,0-2
483	lbu $t10,2($i2) # Te4[s2>>24]
484	_xtr $i2,$s1,0-2
485	lbu $t11,2($i3) # Te4[s3>>24]
486	_xtr $i3,$s2,0-2
487
488	and $i0,0x3fc
489	and $i1,0x3fc
490	and $i2,0x3fc
491	and $i3,0x3fc
492	$PTR_ADD $i0,$Tbl
493	$PTR_ADD $i1,$Tbl
494	$PTR_ADD $i2,$Tbl
495	$PTR_ADD $i3,$Tbl
496	# else
497	lbu $t4,2($i0) # Te4[s2>>8]
498	_xtr $i0,$s0,24-2
499	lbu $t5,2($i1) # Te4[s3>>8]
500	_xtr $i1,$s1,24-2
501	lbu $t6,2($i2) # Te4[s0>>8]
502	_xtr $i2,$s2,24-2
503	lbu $t7,2($i3) # Te4[s1>>8]
504	_xtr $i3,$s3,24-2
505
506	and $i0,0x3fc
507	and $i1,0x3fc
508	and $i2,0x3fc
509	and $i3,0x3fc
510	$PTR_ADD $i0,$Tbl
511	$PTR_ADD $i1,$Tbl
512	$PTR_ADD $i2,$Tbl
513	$PTR_ADD $i3,$Tbl
514	lbu $t8,2($i0) # Te4[s0>>24]
515	$PTR_INS $i0,$s3,2,8
516	lbu $t9,2($i1) # Te4[s1>>24]
517	$PTR_INS $i1,$s0,2,8
518	lbu $t10,2($i2) # Te4[s2>>24]
519	$PTR_INS $i2,$s1,2,8
520	lbu $t11,2($i3) # Te4[s3>>24]
521	$PTR_INS $i3,$s2,2,8
522	# endif
523	_ins $t0,16
524	_ins $t1,16
525	_ins $t2,16
526	_ins $t3,16
527
528	_ins2 $t0,$t4,8
529	lbu $t4,2($i0) # Te4[s3]
530	_ins2 $t1,$t5,8
531	lbu $t5,2($i1) # Te4[s0]
532	_ins2 $t2,$t6,8
533	lbu $t6,2($i2) # Te4[s1]
534	_ins2 $t3,$t7,8
535	lbu $t7,2($i3) # Te4[s2]
536
537	_ins2 $t0,$t8,24
538	lw $s0,0($key0)
539	_ins2 $t1,$t9,24
540	lw $s1,4($key0)
541	_ins2 $t2,$t10,24
542	lw $s2,8($key0)
543	_ins2 $t3,$t11,24
544	lw $s3,12($key0)
545
546	_ins2 $t0,$t4,0
547	_ins2 $t1,$t5,0
548	_ins2 $t2,$t6,0
549	_ins2 $t3,$t7,0
550	#else
551	lbu $t4,2($i0) # Te4[s2>>8]
552	_xtr $i0,$s0,24-2
553	lbu $t5,2($i1) # Te4[s3>>8]
554	_xtr $i1,$s1,24-2
555	lbu $t6,2($i2) # Te4[s0>>8]
556	_xtr $i2,$s2,24-2
557	lbu $t7,2($i3) # Te4[s1>>8]
558	_xtr $i3,$s3,24-2
559
560	and $i0,0x3fc
561	and $i1,0x3fc
562	and $i2,0x3fc
563	and $i3,0x3fc
564	$PTR_ADD $i0,$Tbl
565	$PTR_ADD $i1,$Tbl
566	$PTR_ADD $i2,$Tbl
567	$PTR_ADD $i3,$Tbl
568	lbu $t8,2($i0) # Te4[s0>>24]
569	_xtr $i0,$s3,0-2
570	lbu $t9,2($i1) # Te4[s1>>24]
571	_xtr $i1,$s0,0-2
572	lbu $t10,2($i2) # Te4[s2>>24]
573	_xtr $i2,$s1,0-2
574	lbu $t11,2($i3) # Te4[s3>>24]
575	_xtr $i3,$s2,0-2
576
577	and $i0,0x3fc
578	and $i1,0x3fc
579	and $i2,0x3fc
580	and $i3,0x3fc
581	$PTR_ADD $i0,$Tbl
582	$PTR_ADD $i1,$Tbl
583	$PTR_ADD $i2,$Tbl
584	$PTR_ADD $i3,$Tbl
585
586	_ins $t0,16
587	_ins $t1,16
588	_ins $t2,16
589	_ins $t3,16
590
591	_ins $t4,8
592	_ins $t5,8
593	_ins $t6,8
594	_ins $t7,8
595
596	xor $t0,$t4
597	lbu $t4,2($i0) # Te4[s3]
598	xor $t1,$t5
599	lbu $t5,2($i1) # Te4[s0]
600	xor $t2,$t6
601	lbu $t6,2($i2) # Te4[s1]
602	xor $t3,$t7
603	lbu $t7,2($i3) # Te4[s2]
604
605	_ins $t8,24
606	lw $s0,0($key0)
607	_ins $t9,24
608	lw $s1,4($key0)
609	_ins $t10,24
610	lw $s2,8($key0)
611	_ins $t11,24
612	lw $s3,12($key0)
613
614	xor $t0,$t8
615	xor $t1,$t9
616	xor $t2,$t10
617	xor $t3,$t11
618
619	_ins $t4,0
620	_ins $t5,0
621	_ins $t6,0
622	_ins $t7,0
623
624	xor $t0,$t4
625	xor $t1,$t5
626	xor $t2,$t6
627	xor $t3,$t7
628	#endif
629	xor $s0,$t0
630	xor $s1,$t1
631	xor $s2,$t2
632	xor $s3,$t3
633
634	jr $ra
635	.end _mips_AES_encrypt
636
637	.align 5
638	.globl AES_encrypt
639	.ent AES_encrypt
640	AES_encrypt:
641	.frame $sp,$FRAMESIZE,$ra
642	.mask $SAVED_REGS_MASK,-$SZREG
643	.set noreorder
644	___
645	$code.=<<___ if ($flavour =~ /o32/i); # o32 PIC-ification
646	.cpload $pf
647	___
648	$code.=<<___;
649	$PTR_SUB $sp,$FRAMESIZE
650	$REG_S $ra,$FRAMESIZE-1*$SZREG($sp)
651	$REG_S $fp,$FRAMESIZE-2*$SZREG($sp)
652	$REG_S $s11,$FRAMESIZE-3*$SZREG($sp)
653	$REG_S $s10,$FRAMESIZE-4*$SZREG($sp)
654	$REG_S $s9,$FRAMESIZE-5*$SZREG($sp)
655	$REG_S $s8,$FRAMESIZE-6*$SZREG($sp)
656	$REG_S $s7,$FRAMESIZE-7*$SZREG($sp)
657	$REG_S $s6,$FRAMESIZE-8*$SZREG($sp)
658	$REG_S $s5,$FRAMESIZE-9*$SZREG($sp)
659	$REG_S $s4,$FRAMESIZE-10*$SZREG($sp)
660	___
661	$code.=<<___ if ($flavour =~ /nubi/i); # optimize non-nubi prologue
662	$REG_S \$15,$FRAMESIZE-11*$SZREG($sp)
663	$REG_S \$14,$FRAMESIZE-12*$SZREG($sp)
664	$REG_S \$13,$FRAMESIZE-13*$SZREG($sp)
665	$REG_S \$12,$FRAMESIZE-14*$SZREG($sp)
666	$REG_S $gp,$FRAMESIZE-15*$SZREG($sp)
667	___
668	$code.=<<___ if ($flavour !~ /o32/i); # non-o32 PIC-ification
669	.cplocal $Tbl
670	.cpsetup $pf,$zero,AES_encrypt
671	___
672	$code.=<<___;
673	.set reorder
674	$PTR_LA $Tbl,AES_Te # PIC-ified 'load address'
675
676	#if defined(_MIPS_ARCH_MIPS32R6) \|\| defined(_MIPS_ARCH_MIPS64R6)
677	lw $s0,0($inp)
678	lw $s1,4($inp)
679	lw $s2,8($inp)
680	lw $s3,12($inp)
681	#else
682	lwl $s0,0+$MSB($inp)
683	lwl $s1,4+$MSB($inp)
684	lwl $s2,8+$MSB($inp)
685	lwl $s3,12+$MSB($inp)
686	lwr $s0,0+$LSB($inp)
687	lwr $s1,4+$LSB($inp)
688	lwr $s2,8+$LSB($inp)
689	lwr $s3,12+$LSB($inp)
690	#endif
691
692	bal _mips_AES_encrypt
693
694	#if defined(_MIPS_ARCH_MIPS32R6) \|\| defined(_MIPS_ARCH_MIPS64R6)
695	sw $s0,0($out)
696	sw $s1,4($out)
697	sw $s2,8($out)
698	sw $s3,12($out)
699	#else
700	swr $s0,0+$LSB($out)
701	swr $s1,4+$LSB($out)
702	swr $s2,8+$LSB($out)
703	swr $s3,12+$LSB($out)
704	swl $s0,0+$MSB($out)
705	swl $s1,4+$MSB($out)
706	swl $s2,8+$MSB($out)
707	swl $s3,12+$MSB($out)
708	#endif
709
710	.set noreorder
711	$REG_L $ra,$FRAMESIZE-1*$SZREG($sp)
712	$REG_L $fp,$FRAMESIZE-2*$SZREG($sp)
713	$REG_L $s11,$FRAMESIZE-3*$SZREG($sp)
714	$REG_L $s10,$FRAMESIZE-4*$SZREG($sp)
715	$REG_L $s9,$FRAMESIZE-5*$SZREG($sp)
716	$REG_L $s8,$FRAMESIZE-6*$SZREG($sp)
717	$REG_L $s7,$FRAMESIZE-7*$SZREG($sp)
718	$REG_L $s6,$FRAMESIZE-8*$SZREG($sp)
719	$REG_L $s5,$FRAMESIZE-9*$SZREG($sp)
720	$REG_L $s4,$FRAMESIZE-10*$SZREG($sp)
721	___
722	$code.=<<___ if ($flavour =~ /nubi/i);
723	$REG_L \$15,$FRAMESIZE-11*$SZREG($sp)
724	$REG_L \$14,$FRAMESIZE-12*$SZREG($sp)
725	$REG_L \$13,$FRAMESIZE-13*$SZREG($sp)
726	$REG_L \$12,$FRAMESIZE-14*$SZREG($sp)
727	$REG_L $gp,$FRAMESIZE-15*$SZREG($sp)
728	___
729	$code.=<<___;
730	jr $ra
731	$PTR_ADD $sp,$FRAMESIZE
732	.end AES_encrypt
733	___
734
735
736	$code.=<<___;
737	.align 5
738	.ent _mips_AES_decrypt
739	_mips_AES_decrypt:
740	.frame $sp,0,$ra
741	.set reorder
742	lw $t0,0($key)
743	lw $t1,4($key)
744	lw $t2,8($key)
745	lw $t3,12($key)
746	lw $cnt,240($key)
747	$PTR_ADD $key0,$key,16
748
749	xor $s0,$t0
750	xor $s1,$t1
751	xor $s2,$t2
752	xor $s3,$t3
753
754	subu $cnt,1
755	#if defined(__mips_smartmips)
756	ext $i0,$s3,16,8
757	.Loop_dec:
758	ext $i1,$s0,16,8
759	ext $i2,$s1,16,8
760	ext $i3,$s2,16,8
761	lwxs $t0,$i0($Tbl) # Td1[s3>>16]
762	ext $i0,$s2,8,8
763	lwxs $t1,$i1($Tbl) # Td1[s0>>16]
764	ext $i1,$s3,8,8
765	lwxs $t2,$i2($Tbl) # Td1[s1>>16]
766	ext $i2,$s0,8,8
767	lwxs $t3,$i3($Tbl) # Td1[s2>>16]
768	ext $i3,$s1,8,8
769
770	lwxs $t4,$i0($Tbl) # Td2[s2>>8]
771	ext $i0,$s1,0,8
772	lwxs $t5,$i1($Tbl) # Td2[s3>>8]
773	ext $i1,$s2,0,8
774	lwxs $t6,$i2($Tbl) # Td2[s0>>8]
775	ext $i2,$s3,0,8
776	lwxs $t7,$i3($Tbl) # Td2[s1>>8]
777	ext $i3,$s0,0,8
778
779	lwxs $t8,$i0($Tbl) # Td3[s1]
780	ext $i0,$s0,24,8
781	lwxs $t9,$i1($Tbl) # Td3[s2]
782	ext $i1,$s1,24,8
783	lwxs $t10,$i2($Tbl) # Td3[s3]
784	ext $i2,$s2,24,8
785	lwxs $t11,$i3($Tbl) # Td3[s0]
786	ext $i3,$s3,24,8
787
788	rotr $t0,$t0,8
789	rotr $t1,$t1,8
790	rotr $t2,$t2,8
791	rotr $t3,$t3,8
792
793	rotr $t4,$t4,16
794	rotr $t5,$t5,16
795	rotr $t6,$t6,16
796	rotr $t7,$t7,16
797
798	xor $t0,$t4
799	lwxs $t4,$i0($Tbl) # Td0[s0>>24]
800	xor $t1,$t5
801	lwxs $t5,$i1($Tbl) # Td0[s1>>24]
802	xor $t2,$t6
803	lwxs $t6,$i2($Tbl) # Td0[s2>>24]
804	xor $t3,$t7
805	lwxs $t7,$i3($Tbl) # Td0[s3>>24]
806
807	rotr $t8,$t8,24
808	lw $s0,0($key0)
809	rotr $t9,$t9,24
810	lw $s1,4($key0)
811	rotr $t10,$t10,24
812	lw $s2,8($key0)
813	rotr $t11,$t11,24
814	lw $s3,12($key0)
815
816	xor $t0,$t8
817	xor $t1,$t9
818	xor $t2,$t10
819	xor $t3,$t11
820
821	xor $t0,$t4
822	xor $t1,$t5
823	xor $t2,$t6
824	xor $t3,$t7
825
826	subu $cnt,1
827	$PTR_ADD $key0,16
828	xor $s0,$t0
829	xor $s1,$t1
830	xor $s2,$t2
831	xor $s3,$t3
832	.set noreorder
833	bnez $cnt,.Loop_dec
834	ext $i0,$s3,16,8
835
836	_xtr $i0,$s3,16-2
837	#else
838	#if defined(_MIPS_ARCH_MIPS32R2) \|\| defined(_MIPS_ARCH_MIPS64R2)
839	move $i0,$Tbl
840	move $i1,$Tbl
841	move $i2,$Tbl
842	move $i3,$Tbl
843	ext $t0,$s3,16,8
844	.Loop_dec:
845	ext $t1,$s0,16,8
846	ext $t2,$s1,16,8
847	ext $t3,$s2,16,8
848	$PTR_INS $i0,$t0,2,8
849	$PTR_INS $i1,$t1,2,8
850	$PTR_INS $i2,$t2,2,8
851	$PTR_INS $i3,$t3,2,8
852	lw $t0,0($i0) # Td1[s3>>16]
853	ext $t4,$s2,8,8
854	lw $t1,0($i1) # Td1[s0>>16]
855	ext $t5,$s3,8,8
856	lw $t2,0($i2) # Td1[s1>>16]
857	ext $t6,$s0,8,8
858	lw $t3,0($i3) # Td1[s2>>16]
859	ext $t7,$s1,8,8
860	$PTR_INS $i0,$t4,2,8
861	$PTR_INS $i1,$t5,2,8
862	$PTR_INS $i2,$t6,2,8
863	$PTR_INS $i3,$t7,2,8
864	#else
865	_xtr $i0,$s3,16-2
866	.Loop_dec:
867	_xtr $i1,$s0,16-2
868	_xtr $i2,$s1,16-2
869	_xtr $i3,$s2,16-2
870	and $i0,0x3fc
871	and $i1,0x3fc
872	and $i2,0x3fc
873	and $i3,0x3fc
874	$PTR_ADD $i0,$Tbl
875	$PTR_ADD $i1,$Tbl
876	$PTR_ADD $i2,$Tbl
877	$PTR_ADD $i3,$Tbl
878	lwl $t0,3($i0) # Td1[s3>>16]
879	lwl $t1,3($i1) # Td1[s0>>16]
880	lwl $t2,3($i2) # Td1[s1>>16]
881	lwl $t3,3($i3) # Td1[s2>>16]
882	lwr $t0,2($i0) # Td1[s3>>16]
883	_xtr $i0,$s2,8-2
884	lwr $t1,2($i1) # Td1[s0>>16]
885	_xtr $i1,$s3,8-2
886	lwr $t2,2($i2) # Td1[s1>>16]
887	_xtr $i2,$s0,8-2
888	lwr $t3,2($i3) # Td1[s2>>16]
889	_xtr $i3,$s1,8-2
890	and $i0,0x3fc
891	and $i1,0x3fc
892	and $i2,0x3fc
893	and $i3,0x3fc
894	$PTR_ADD $i0,$Tbl
895	$PTR_ADD $i1,$Tbl
896	$PTR_ADD $i2,$Tbl
897	$PTR_ADD $i3,$Tbl
898	#endif
899	#if defined(_MIPS_ARCH_MIPS32R2) \|\| defined(_MIPS_ARCH_MIPS64R2)
900	rotr $t0,$t0,8
901	rotr $t1,$t1,8
902	rotr $t2,$t2,8
903	rotr $t3,$t3,8
904	# if defined(_MIPSEL)
905	lw $t4,0($i0) # Td2[s2>>8]
906	ext $t8,$s1,0,8
907	lw $t5,0($i1) # Td2[s3>>8]
908	ext $t9,$s2,0,8
909	lw $t6,0($i2) # Td2[s0>>8]
910	ext $t10,$s3,0,8
911	lw $t7,0($i3) # Td2[s1>>8]
912	ext $t11,$s0,0,8
913	$PTR_INS $i0,$t8,2,8
914	$PTR_INS $i1,$t9,2,8
915	$PTR_INS $i2,$t10,2,8
916	$PTR_INS $i3,$t11,2,8
917	lw $t8,0($i0) # Td3[s1]
918	$PTR_INS $i0,$s0,2,8
919	lw $t9,0($i1) # Td3[s2]
920	$PTR_INS $i1,$s1,2,8
921	lw $t10,0($i2) # Td3[s3]
922	$PTR_INS $i2,$s2,2,8
923	lw $t11,0($i3) # Td3[s0]
924	$PTR_INS $i3,$s3,2,8
925	#else
926	lw $t4,0($i0) # Td2[s2>>8]
927	$PTR_INS $i0,$s1,2,8
928	lw $t5,0($i1) # Td2[s3>>8]
929	$PTR_INS $i1,$s2,2,8
930	lw $t6,0($i2) # Td2[s0>>8]
931	$PTR_INS $i2,$s3,2,8
932	lw $t7,0($i3) # Td2[s1>>8]
933	$PTR_INS $i3,$s0,2,8
934
935	lw $t8,0($i0) # Td3[s1]
936	_xtr $i0,$s0,24-2
937	lw $t9,0($i1) # Td3[s2]
938	_xtr $i1,$s1,24-2
939	lw $t10,0($i2) # Td3[s3]
940	_xtr $i2,$s2,24-2
941	lw $t11,0($i3) # Td3[s0]
942	_xtr $i3,$s3,24-2
943
944	and $i0,0x3fc
945	and $i1,0x3fc
946	and $i2,0x3fc
947	and $i3,0x3fc
948	$PTR_ADD $i0,$Tbl
949	$PTR_ADD $i1,$Tbl
950	$PTR_ADD $i2,$Tbl
951	$PTR_ADD $i3,$Tbl
952	#endif
953	rotr $t4,$t4,16
954	rotr $t5,$t5,16
955	rotr $t6,$t6,16
956	rotr $t7,$t7,16
957
958	rotr $t8,$t8,24
959	rotr $t9,$t9,24
960	rotr $t10,$t10,24
961	rotr $t11,$t11,24
962	#else
963	lwl $t4,2($i0) # Td2[s2>>8]
964	lwl $t5,2($i1) # Td2[s3>>8]
965	lwl $t6,2($i2) # Td2[s0>>8]
966	lwl $t7,2($i3) # Td2[s1>>8]
967	lwr $t4,1($i0) # Td2[s2>>8]
968	_xtr $i0,$s1,0-2
969	lwr $t5,1($i1) # Td2[s3>>8]
970	_xtr $i1,$s2,0-2
971	lwr $t6,1($i2) # Td2[s0>>8]
972	_xtr $i2,$s3,0-2
973	lwr $t7,1($i3) # Td2[s1>>8]
974	_xtr $i3,$s0,0-2
975
976	and $i0,0x3fc
977	and $i1,0x3fc
978	and $i2,0x3fc
979	and $i3,0x3fc
980	$PTR_ADD $i0,$Tbl
981	$PTR_ADD $i1,$Tbl
982	$PTR_ADD $i2,$Tbl
983	$PTR_ADD $i3,$Tbl
984	lwl $t8,1($i0) # Td3[s1]
985	lwl $t9,1($i1) # Td3[s2]
986	lwl $t10,1($i2) # Td3[s3]
987	lwl $t11,1($i3) # Td3[s0]
988	lwr $t8,0($i0) # Td3[s1]
989	_xtr $i0,$s0,24-2
990	lwr $t9,0($i1) # Td3[s2]
991	_xtr $i1,$s1,24-2
992	lwr $t10,0($i2) # Td3[s3]
993	_xtr $i2,$s2,24-2
994	lwr $t11,0($i3) # Td3[s0]
995	_xtr $i3,$s3,24-2
996
997	and $i0,0x3fc
998	and $i1,0x3fc
999	and $i2,0x3fc
1000	and $i3,0x3fc
1001	$PTR_ADD $i0,$Tbl
1002	$PTR_ADD $i1,$Tbl
1003	$PTR_ADD $i2,$Tbl
1004	$PTR_ADD $i3,$Tbl
1005	#endif
1006
1007	xor $t0,$t4
1008	lw $t4,0($i0) # Td0[s0>>24]
1009	xor $t1,$t5
1010	lw $t5,0($i1) # Td0[s1>>24]
1011	xor $t2,$t6
1012	lw $t6,0($i2) # Td0[s2>>24]
1013	xor $t3,$t7
1014	lw $t7,0($i3) # Td0[s3>>24]
1015
1016	xor $t0,$t8
1017	lw $s0,0($key0)
1018	xor $t1,$t9
1019	lw $s1,4($key0)
1020	xor $t2,$t10
1021	lw $s2,8($key0)
1022	xor $t3,$t11
1023	lw $s3,12($key0)
1024
1025	xor $t0,$t4
1026	xor $t1,$t5
1027	xor $t2,$t6
1028	xor $t3,$t7
1029
1030	subu $cnt,1
1031	$PTR_ADD $key0,16
1032	xor $s0,$t0
1033	xor $s1,$t1
1034	xor $s2,$t2
1035	xor $s3,$t3
1036	.set noreorder
1037	bnez $cnt,.Loop_dec
1038	#if defined(_MIPS_ARCH_MIPS32R2) \|\| defined(_MIPS_ARCH_MIPS64R2)
1039	ext $t0,$s3,16,8
1040	#endif
1041
1042	_xtr $i0,$s3,16-2
1043	#endif
1044
1045	.set reorder
1046	lw $t4,1024($Tbl) # prefetch Td4
1047	_xtr $i0,$s3,16
1048	lw $t5,1024+32($Tbl)
1049	_xtr $i1,$s0,16
1050	lw $t6,1024+64($Tbl)
1051	_xtr $i2,$s1,16
1052	lw $t7,1024+96($Tbl)
1053	_xtr $i3,$s2,16
1054	lw $t8,1024+128($Tbl)
1055	and $i0,0xff
1056	lw $t9,1024+160($Tbl)
1057	and $i1,0xff
1058	lw $t10,1024+192($Tbl)
1059	and $i2,0xff
1060	lw $t11,1024+224($Tbl)
1061	and $i3,0xff
1062
1063	$PTR_ADD $i0,$Tbl
1064	$PTR_ADD $i1,$Tbl
1065	$PTR_ADD $i2,$Tbl
1066	$PTR_ADD $i3,$Tbl
1067	lbu $t0,1024($i0) # Td4[s3>>16]
1068	_xtr $i0,$s2,8
1069	lbu $t1,1024($i1) # Td4[s0>>16]
1070	_xtr $i1,$s3,8
1071	lbu $t2,1024($i2) # Td4[s1>>16]
1072	_xtr $i2,$s0,8
1073	lbu $t3,1024($i3) # Td4[s2>>16]
1074	_xtr $i3,$s1,8
1075
1076	and $i0,0xff
1077	and $i1,0xff
1078	and $i2,0xff
1079	and $i3,0xff
1080	$PTR_ADD $i0,$Tbl
1081	$PTR_ADD $i1,$Tbl
1082	$PTR_ADD $i2,$Tbl
1083	$PTR_ADD $i3,$Tbl
1084	#if defined(_MIPS_ARCH_MIPS32R2) \|\| defined(_MIPS_ARCH_MIPS64R2)
1085	# if defined(_MIPSEL)
1086	lbu $t4,1024($i0) # Td4[s2>>8]
1087	$PTR_INS $i0,$s0,0,8
1088	lbu $t5,1024($i1) # Td4[s3>>8]
1089	$PTR_INS $i1,$s1,0,8
1090	lbu $t6,1024($i2) # Td4[s0>>8]
1091	$PTR_INS $i2,$s2,0,8
1092	lbu $t7,1024($i3) # Td4[s1>>8]
1093	$PTR_INS $i3,$s3,0,8
1094
1095	lbu $t8,1024($i0) # Td4[s0>>24]
1096	_xtr $i0,$s1,0
1097	lbu $t9,1024($i1) # Td4[s1>>24]
1098	_xtr $i1,$s2,0
1099	lbu $t10,1024($i2) # Td4[s2>>24]
1100	_xtr $i2,$s3,0
1101	lbu $t11,1024($i3) # Td4[s3>>24]
1102	_xtr $i3,$s0,0
1103
1104	$PTR_ADD $i0,$Tbl
1105	$PTR_ADD $i1,$Tbl
1106	$PTR_ADD $i2,$Tbl
1107	$PTR_ADD $i3,$Tbl
1108	# else
1109	lbu $t4,1024($i0) # Td4[s2>>8]
1110	_xtr $i0,$s0,24
1111	lbu $t5,1024($i1) # Td4[s3>>8]
1112	_xtr $i1,$s1,24
1113	lbu $t6,1024($i2) # Td4[s0>>8]
1114	_xtr $i2,$s2,24
1115	lbu $t7,1024($i3) # Td4[s1>>8]
1116	_xtr $i3,$s3,24
1117
1118	$PTR_ADD $i0,$Tbl
1119	$PTR_ADD $i1,$Tbl
1120	$PTR_ADD $i2,$Tbl
1121	$PTR_ADD $i3,$Tbl
1122	lbu $t8,1024($i0) # Td4[s0>>24]
1123	$PTR_INS $i0,$s1,0,8
1124	lbu $t9,1024($i1) # Td4[s1>>24]
1125	$PTR_INS $i1,$s2,0,8
1126	lbu $t10,1024($i2) # Td4[s2>>24]
1127	$PTR_INS $i2,$s3,0,8
1128	lbu $t11,1024($i3) # Td4[s3>>24]
1129	$PTR_INS $i3,$s0,0,8
1130	# endif
1131	_ins $t0,16
1132	_ins $t1,16
1133	_ins $t2,16
1134	_ins $t3,16
1135
1136	_ins2 $t0,$t4,8
1137	lbu $t4,1024($i0) # Td4[s1]
1138	_ins2 $t1,$t5,8
1139	lbu $t5,1024($i1) # Td4[s2]
1140	_ins2 $t2,$t6,8
1141	lbu $t6,1024($i2) # Td4[s3]
1142	_ins2 $t3,$t7,8
1143	lbu $t7,1024($i3) # Td4[s0]
1144
1145	_ins2 $t0,$t8,24
1146	lw $s0,0($key0)
1147	_ins2 $t1,$t9,24
1148	lw $s1,4($key0)
1149	_ins2 $t2,$t10,24
1150	lw $s2,8($key0)
1151	_ins2 $t3,$t11,24
1152	lw $s3,12($key0)
1153
1154	_ins2 $t0,$t4,0
1155	_ins2 $t1,$t5,0
1156	_ins2 $t2,$t6,0
1157	_ins2 $t3,$t7,0
1158	#else
1159	lbu $t4,1024($i0) # Td4[s2>>8]
1160	_xtr $i0,$s0,24
1161	lbu $t5,1024($i1) # Td4[s3>>8]
1162	_xtr $i1,$s1,24
1163	lbu $t6,1024($i2) # Td4[s0>>8]
1164	_xtr $i2,$s2,24
1165	lbu $t7,1024($i3) # Td4[s1>>8]
1166	_xtr $i3,$s3,24
1167
1168	$PTR_ADD $i0,$Tbl
1169	$PTR_ADD $i1,$Tbl
1170	$PTR_ADD $i2,$Tbl
1171	$PTR_ADD $i3,$Tbl
1172	lbu $t8,1024($i0) # Td4[s0>>24]
1173	_xtr $i0,$s1,0
1174	lbu $t9,1024($i1) # Td4[s1>>24]
1175	_xtr $i1,$s2,0
1176	lbu $t10,1024($i2) # Td4[s2>>24]
1177	_xtr $i2,$s3,0
1178	lbu $t11,1024($i3) # Td4[s3>>24]
1179	_xtr $i3,$s0,0
1180
1181	$PTR_ADD $i0,$Tbl
1182	$PTR_ADD $i1,$Tbl
1183	$PTR_ADD $i2,$Tbl
1184	$PTR_ADD $i3,$Tbl
1185
1186	_ins $t0,16
1187	_ins $t1,16
1188	_ins $t2,16
1189	_ins $t3,16
1190
1191	_ins $t4,8
1192	_ins $t5,8
1193	_ins $t6,8
1194	_ins $t7,8
1195
1196	xor $t0,$t4
1197	lbu $t4,1024($i0) # Td4[s1]
1198	xor $t1,$t5
1199	lbu $t5,1024($i1) # Td4[s2]
1200	xor $t2,$t6
1201	lbu $t6,1024($i2) # Td4[s3]
1202	xor $t3,$t7
1203	lbu $t7,1024($i3) # Td4[s0]
1204
1205	_ins $t8,24
1206	lw $s0,0($key0)
1207	_ins $t9,24
1208	lw $s1,4($key0)
1209	_ins $t10,24
1210	lw $s2,8($key0)
1211	_ins $t11,24
1212	lw $s3,12($key0)
1213
1214	xor $t0,$t8
1215	xor $t1,$t9
1216	xor $t2,$t10
1217	xor $t3,$t11
1218
1219	_ins $t4,0
1220	_ins $t5,0
1221	_ins $t6,0
1222	_ins $t7,0
1223
1224	xor $t0,$t4
1225	xor $t1,$t5
1226	xor $t2,$t6
1227	xor $t3,$t7
1228	#endif
1229
1230	xor $s0,$t0
1231	xor $s1,$t1
1232	xor $s2,$t2
1233	xor $s3,$t3
1234
1235	jr $ra
1236	.end _mips_AES_decrypt
1237
1238	.align 5
1239	.globl AES_decrypt
1240	.ent AES_decrypt
1241	AES_decrypt:
1242	.frame $sp,$FRAMESIZE,$ra
1243	.mask $SAVED_REGS_MASK,-$SZREG
1244	.set noreorder
1245	___
1246	$code.=<<___ if ($flavour =~ /o32/i); # o32 PIC-ification
1247	.cpload $pf
1248	___
1249	$code.=<<___;
1250	$PTR_SUB $sp,$FRAMESIZE
1251	$REG_S $ra,$FRAMESIZE-1*$SZREG($sp)
1252	$REG_S $fp,$FRAMESIZE-2*$SZREG($sp)
1253	$REG_S $s11,$FRAMESIZE-3*$SZREG($sp)
1254	$REG_S $s10,$FRAMESIZE-4*$SZREG($sp)
1255	$REG_S $s9,$FRAMESIZE-5*$SZREG($sp)
1256	$REG_S $s8,$FRAMESIZE-6*$SZREG($sp)
1257	$REG_S $s7,$FRAMESIZE-7*$SZREG($sp)
1258	$REG_S $s6,$FRAMESIZE-8*$SZREG($sp)
1259	$REG_S $s5,$FRAMESIZE-9*$SZREG($sp)
1260	$REG_S $s4,$FRAMESIZE-10*$SZREG($sp)
1261	___
1262	$code.=<<___ if ($flavour =~ /nubi/i); # optimize non-nubi prologue
1263	$REG_S \$15,$FRAMESIZE-11*$SZREG($sp)
1264	$REG_S \$14,$FRAMESIZE-12*$SZREG($sp)
1265	$REG_S \$13,$FRAMESIZE-13*$SZREG($sp)
1266	$REG_S \$12,$FRAMESIZE-14*$SZREG($sp)
1267	$REG_S $gp,$FRAMESIZE-15*$SZREG($sp)
1268	___
1269	$code.=<<___ if ($flavour !~ /o32/i); # non-o32 PIC-ification
1270	.cplocal $Tbl
1271	.cpsetup $pf,$zero,AES_decrypt
1272	___
1273	$code.=<<___;
1274	.set reorder
1275	$PTR_LA $Tbl,AES_Td # PIC-ified 'load address'
1276
1277	#if defined(_MIPS_ARCH_MIPS32R6) \|\| defined(_MIPS_ARCH_MIPS64R6)
1278	lw $s0,0($inp)
1279	lw $s1,4($inp)
1280	lw $s2,8($inp)
1281	lw $s3,12($inp)
1282	#else
1283	lwl $s0,0+$MSB($inp)
1284	lwl $s1,4+$MSB($inp)
1285	lwl $s2,8+$MSB($inp)
1286	lwl $s3,12+$MSB($inp)
1287	lwr $s0,0+$LSB($inp)
1288	lwr $s1,4+$LSB($inp)
1289	lwr $s2,8+$LSB($inp)
1290	lwr $s3,12+$LSB($inp)
1291	#endif
1292
1293	bal _mips_AES_decrypt
1294
1295	#if defined(_MIPS_ARCH_MIPS32R6) \|\| defined(_MIPS_ARCH_MIPS64R6)
1296	sw $s0,0($out)
1297	sw $s1,4($out)
1298	sw $s2,8($out)
1299	sw $s3,12($out)
1300	#else
1301	swr $s0,0+$LSB($out)
1302	swr $s1,4+$LSB($out)
1303	swr $s2,8+$LSB($out)
1304	swr $s3,12+$LSB($out)
1305	swl $s0,0+$MSB($out)
1306	swl $s1,4+$MSB($out)
1307	swl $s2,8+$MSB($out)
1308	swl $s3,12+$MSB($out)
1309	#endif
1310
1311	.set noreorder
1312	$REG_L $ra,$FRAMESIZE-1*$SZREG($sp)
1313	$REG_L $fp,$FRAMESIZE-2*$SZREG($sp)
1314	$REG_L $s11,$FRAMESIZE-3*$SZREG($sp)
1315	$REG_L $s10,$FRAMESIZE-4*$SZREG($sp)
1316	$REG_L $s9,$FRAMESIZE-5*$SZREG($sp)
1317	$REG_L $s8,$FRAMESIZE-6*$SZREG($sp)
1318	$REG_L $s7,$FRAMESIZE-7*$SZREG($sp)
1319	$REG_L $s6,$FRAMESIZE-8*$SZREG($sp)
1320	$REG_L $s5,$FRAMESIZE-9*$SZREG($sp)
1321	$REG_L $s4,$FRAMESIZE-10*$SZREG($sp)
1322	___
1323	$code.=<<___ if ($flavour =~ /nubi/i);
1324	$REG_L \$15,$FRAMESIZE-11*$SZREG($sp)
1325	$REG_L \$14,$FRAMESIZE-12*$SZREG($sp)
1326	$REG_L \$13,$FRAMESIZE-13*$SZREG($sp)
1327	$REG_L \$12,$FRAMESIZE-14*$SZREG($sp)
1328	$REG_L $gp,$FRAMESIZE-15*$SZREG($sp)
1329	___
1330	$code.=<<___;
1331	jr $ra
1332	$PTR_ADD $sp,$FRAMESIZE
1333	.end AES_decrypt
1334	___
1335	}}}
1336
1337
1338	{{{
1339	my $FRAMESIZE=8*$SZREG;
1340	my $SAVED_REGS_MASK = ($flavour =~ /nubi/i) ? "0xc000f008" : "0xc0000000";
1341
1342	my ($inp,$bits,$key,$Tbl)=($a0,$a1,$a2,$a3);
1343	my ($rk0,$rk1,$rk2,$rk3,$rk4,$rk5,$rk6,$rk7)=($a4,$a5,$a6,$a7,$s0,$s1,$s2,$s3);
1344	my ($i0,$i1,$i2,$i3)=($at,$t0,$t1,$t2);
1345	my ($rcon,$cnt)=($gp,$fp);
1346
1347	$code.=<<___;
1348	.align 5
1349	.ent _mips_AES_set_encrypt_key
1350	_mips_AES_set_encrypt_key:
1351	.frame $sp,0,$ra
1352	.set noreorder
1353	beqz $inp,.Lekey_done
1354	li $t0,-1
1355	beqz $key,.Lekey_done
1356	$PTR_ADD $rcon,$Tbl,256
1357
1358	.set reorder
1359	#if defined(_MIPS_ARCH_MIPS32R6) \|\| defined(_MIPS_ARCH_MIPS64R6)
1360	lw $rk0,0($inp) # load 128 bits
1361	lw $rk1,4($inp)
1362	lw $rk2,8($inp)
1363	lw $rk3,12($inp)
1364	#else
1365	lwl $rk0,0+$MSB($inp) # load 128 bits
1366	lwl $rk1,4+$MSB($inp)
1367	lwl $rk2,8+$MSB($inp)
1368	lwl $rk3,12+$MSB($inp)
1369	lwr $rk0,0+$LSB($inp)
1370	lwr $rk1,4+$LSB($inp)
1371	lwr $rk2,8+$LSB($inp)
1372	lwr $rk3,12+$LSB($inp)
1373	#endif
1374	li $at,128
1375	.set noreorder
1376	beq $bits,$at,.L128bits
1377	li $cnt,10
1378
1379	.set reorder
1380	#if defined(_MIPS_ARCH_MIPS32R6) \|\| defined(_MIPS_ARCH_MIPS64R6)
1381	lw $rk4,16($inp) # load 192 bits
1382	lw $rk5,20($inp)
1383	#else
1384	lwl $rk4,16+$MSB($inp) # load 192 bits
1385	lwl $rk5,20+$MSB($inp)
1386	lwr $rk4,16+$LSB($inp)
1387	lwr $rk5,20+$LSB($inp)
1388	#endif
1389	li $at,192
1390	.set noreorder
1391	beq $bits,$at,.L192bits
1392	li $cnt,8
1393
1394	.set reorder
1395	#if defined(_MIPS_ARCH_MIPS32R6) \|\| defined(_MIPS_ARCH_MIPS64R6)
1396	lw $rk6,24($inp) # load 256 bits
1397	lw $rk7,28($inp)
1398	#else
1399	lwl $rk6,24+$MSB($inp) # load 256 bits
1400	lwl $rk7,28+$MSB($inp)
1401	lwr $rk6,24+$LSB($inp)
1402	lwr $rk7,28+$LSB($inp)
1403	#endif
1404	li $at,256
1405	.set noreorder
1406	beq $bits,$at,.L256bits
1407	li $cnt,7
1408
1409	b .Lekey_done
1410	li $t0,-2
1411
1412	.align 4
1413	.L128bits:
1414	.set reorder
1415	srl $i0,$rk3,16
1416	srl $i1,$rk3,8
1417	and $i0,0xff
1418	and $i1,0xff
1419	and $i2,$rk3,0xff
1420	srl $i3,$rk3,24
1421	$PTR_ADD $i0,$Tbl
1422	$PTR_ADD $i1,$Tbl
1423	$PTR_ADD $i2,$Tbl
1424	$PTR_ADD $i3,$Tbl
1425	lbu $i0,0($i0)
1426	lbu $i1,0($i1)
1427	lbu $i2,0($i2)
1428	lbu $i3,0($i3)
1429
1430	sw $rk0,0($key)
1431	sw $rk1,4($key)
1432	sw $rk2,8($key)
1433	sw $rk3,12($key)
1434	subu $cnt,1
1435	$PTR_ADD $key,16
1436
1437	_bias $i0,24
1438	_bias $i1,16
1439	_bias $i2,8
1440	_bias $i3,0
1441
1442	xor $rk0,$i0
1443	lw $i0,0($rcon)
1444	xor $rk0,$i1
1445	xor $rk0,$i2
1446	xor $rk0,$i3
1447	xor $rk0,$i0
1448
1449	xor $rk1,$rk0
1450	xor $rk2,$rk1
1451	xor $rk3,$rk2
1452
1453	.set noreorder
1454	bnez $cnt,.L128bits
1455	$PTR_ADD $rcon,4
1456
1457	sw $rk0,0($key)
1458	sw $rk1,4($key)
1459	sw $rk2,8($key)
1460	li $cnt,10
1461	sw $rk3,12($key)
1462	li $t0,0
1463	sw $cnt,80($key)
1464	b .Lekey_done
1465	$PTR_SUB $key,10*16
1466
1467	.align 4
1468	.L192bits:
1469	.set reorder
1470	srl $i0,$rk5,16
1471	srl $i1,$rk5,8
1472	and $i0,0xff
1473	and $i1,0xff
1474	and $i2,$rk5,0xff
1475	srl $i3,$rk5,24
1476	$PTR_ADD $i0,$Tbl
1477	$PTR_ADD $i1,$Tbl
1478	$PTR_ADD $i2,$Tbl
1479	$PTR_ADD $i3,$Tbl
1480	lbu $i0,0($i0)
1481	lbu $i1,0($i1)
1482	lbu $i2,0($i2)
1483	lbu $i3,0($i3)
1484
1485	sw $rk0,0($key)
1486	sw $rk1,4($key)
1487	sw $rk2,8($key)
1488	sw $rk3,12($key)
1489	sw $rk4,16($key)
1490	sw $rk5,20($key)
1491	subu $cnt,1
1492	$PTR_ADD $key,24
1493
1494	_bias $i0,24
1495	_bias $i1,16
1496	_bias $i2,8
1497	_bias $i3,0
1498
1499	xor $rk0,$i0
1500	lw $i0,0($rcon)
1501	xor $rk0,$i1
1502	xor $rk0,$i2
1503	xor $rk0,$i3
1504	xor $rk0,$i0
1505
1506	xor $rk1,$rk0
1507	xor $rk2,$rk1
1508	xor $rk3,$rk2
1509	xor $rk4,$rk3
1510	xor $rk5,$rk4
1511
1512	.set noreorder
1513	bnez $cnt,.L192bits
1514	$PTR_ADD $rcon,4
1515
1516	sw $rk0,0($key)
1517	sw $rk1,4($key)
1518	sw $rk2,8($key)
1519	li $cnt,12
1520	sw $rk3,12($key)
1521	li $t0,0
1522	sw $cnt,48($key)
1523	b .Lekey_done
1524	$PTR_SUB $key,12*16
1525
1526	.align 4
1527	.L256bits:
1528	.set reorder
1529	srl $i0,$rk7,16
1530	srl $i1,$rk7,8
1531	and $i0,0xff
1532	and $i1,0xff
1533	and $i2,$rk7,0xff
1534	srl $i3,$rk7,24
1535	$PTR_ADD $i0,$Tbl
1536	$PTR_ADD $i1,$Tbl
1537	$PTR_ADD $i2,$Tbl
1538	$PTR_ADD $i3,$Tbl
1539	lbu $i0,0($i0)
1540	lbu $i1,0($i1)
1541	lbu $i2,0($i2)
1542	lbu $i3,0($i3)
1543
1544	sw $rk0,0($key)
1545	sw $rk1,4($key)
1546	sw $rk2,8($key)
1547	sw $rk3,12($key)
1548	sw $rk4,16($key)
1549	sw $rk5,20($key)
1550	sw $rk6,24($key)
1551	sw $rk7,28($key)
1552	subu $cnt,1
1553
1554	_bias $i0,24
1555	_bias $i1,16
1556	_bias $i2,8
1557	_bias $i3,0
1558
1559	xor $rk0,$i0
1560	lw $i0,0($rcon)
1561	xor $rk0,$i1
1562	xor $rk0,$i2
1563	xor $rk0,$i3
1564	xor $rk0,$i0
1565
1566	xor $rk1,$rk0
1567	xor $rk2,$rk1
1568	xor $rk3,$rk2
1569	beqz $cnt,.L256bits_done
1570
1571	srl $i0,$rk3,24
1572	srl $i1,$rk3,16
1573	srl $i2,$rk3,8
1574	and $i3,$rk3,0xff
1575	and $i1,0xff
1576	and $i2,0xff
1577	$PTR_ADD $i0,$Tbl
1578	$PTR_ADD $i1,$Tbl
1579	$PTR_ADD $i2,$Tbl
1580	$PTR_ADD $i3,$Tbl
1581	lbu $i0,0($i0)
1582	lbu $i1,0($i1)
1583	lbu $i2,0($i2)
1584	lbu $i3,0($i3)
1585	sll $i0,24
1586	sll $i1,16
1587	sll $i2,8
1588
1589	xor $rk4,$i0
1590	xor $rk4,$i1
1591	xor $rk4,$i2
1592	xor $rk4,$i3
1593
1594	xor $rk5,$rk4
1595	xor $rk6,$rk5
1596	xor $rk7,$rk6
1597
1598	$PTR_ADD $key,32
1599	.set noreorder
1600	b .L256bits
1601	$PTR_ADD $rcon,4
1602
1603	.L256bits_done:
1604	sw $rk0,32($key)
1605	sw $rk1,36($key)
1606	sw $rk2,40($key)
1607	li $cnt,14
1608	sw $rk3,44($key)
1609	li $t0,0
1610	sw $cnt,48($key)
1611	$PTR_SUB $key,12*16
1612
1613	.Lekey_done:
1614	jr $ra
1615	nop
1616	.end _mips_AES_set_encrypt_key
1617
1618	.globl AES_set_encrypt_key
1619	.ent AES_set_encrypt_key
1620	AES_set_encrypt_key:
1621	.frame $sp,$FRAMESIZE,$ra
1622	.mask $SAVED_REGS_MASK,-$SZREG
1623	.set noreorder
1624	___
1625	$code.=<<___ if ($flavour =~ /o32/i); # o32 PIC-ification
1626	.cpload $pf
1627	___
1628	$code.=<<___;
1629	$PTR_SUB $sp,$FRAMESIZE
1630	$REG_S $ra,$FRAMESIZE-1*$SZREG($sp)
1631	$REG_S $fp,$FRAMESIZE-2*$SZREG($sp)
1632	___
1633	$code.=<<___ if ($flavour =~ /nubi/i); # optimize non-nubi prologue
1634	$REG_S $s3,$FRAMESIZE-3*$SZREG($sp)
1635	$REG_S $s2,$FRAMESIZE-4*$SZREG($sp)
1636	$REG_S $s1,$FRAMESIZE-5*$SZREG($sp)
1637	$REG_S $s0,$FRAMESIZE-6*$SZREG($sp)
1638	$REG_S $gp,$FRAMESIZE-7*$SZREG($sp)
1639	___
1640	$code.=<<___ if ($flavour !~ /o32/i); # non-o32 PIC-ification
1641	.cplocal $Tbl
1642	.cpsetup $pf,$zero,AES_set_encrypt_key
1643	___
1644	$code.=<<___;
1645	.set reorder
1646	$PTR_LA $Tbl,AES_Te4 # PIC-ified 'load address'
1647
1648	bal _mips_AES_set_encrypt_key
1649
1650	.set noreorder
1651	move $a0,$t0
1652	$REG_L $ra,$FRAMESIZE-1*$SZREG($sp)
1653	$REG_L $fp,$FRAMESIZE-2*$SZREG($sp)
1654	___
1655	$code.=<<___ if ($flavour =~ /nubi/i);
1656	$REG_L $s3,$FRAMESIZE-11*$SZREG($sp)
1657	$REG_L $s2,$FRAMESIZE-12*$SZREG($sp)
1658	$REG_L $s1,$FRAMESIZE-13*$SZREG($sp)
1659	$REG_L $s0,$FRAMESIZE-14*$SZREG($sp)
1660	$REG_L $gp,$FRAMESIZE-15*$SZREG($sp)
1661	___
1662	$code.=<<___;
1663	jr $ra
1664	$PTR_ADD $sp,$FRAMESIZE
1665	.end AES_set_encrypt_key
1666	___
1667
1668
1669	my ($head,$tail)=($inp,$bits);
1670	my ($tp1,$tp2,$tp4,$tp8,$tp9,$tpb,$tpd,$tpe)=($a4,$a5,$a6,$a7,$s0,$s1,$s2,$s3);
1671	my ($m,$x80808080,$x7f7f7f7f,$x1b1b1b1b)=($at,$t0,$t1,$t2);
1672	$code.=<<___;
1673	.align 5
1674	.globl AES_set_decrypt_key
1675	.ent AES_set_decrypt_key
1676	AES_set_decrypt_key:
1677	.frame $sp,$FRAMESIZE,$ra
1678	.mask $SAVED_REGS_MASK,-$SZREG
1679	.set noreorder
1680	___
1681	$code.=<<___ if ($flavour =~ /o32/i); # o32 PIC-ification
1682	.cpload $pf
1683	___
1684	$code.=<<___;
1685	$PTR_SUB $sp,$FRAMESIZE
1686	$REG_S $ra,$FRAMESIZE-1*$SZREG($sp)
1687	$REG_S $fp,$FRAMESIZE-2*$SZREG($sp)
1688	___
1689	$code.=<<___ if ($flavour =~ /nubi/i); # optimize non-nubi prologue
1690	$REG_S $s3,$FRAMESIZE-3*$SZREG($sp)
1691	$REG_S $s2,$FRAMESIZE-4*$SZREG($sp)
1692	$REG_S $s1,$FRAMESIZE-5*$SZREG($sp)
1693	$REG_S $s0,$FRAMESIZE-6*$SZREG($sp)
1694	$REG_S $gp,$FRAMESIZE-7*$SZREG($sp)
1695	___
1696	$code.=<<___ if ($flavour !~ /o32/i); # non-o32 PIC-ification
1697	.cplocal $Tbl
1698	.cpsetup $pf,$zero,AES_set_decrypt_key
1699	___
1700	$code.=<<___;
1701	.set reorder
1702	$PTR_LA $Tbl,AES_Te4 # PIC-ified 'load address'
1703
1704	bal _mips_AES_set_encrypt_key
1705
1706	bltz $t0,.Ldkey_done
1707
1708	sll $at,$cnt,4
1709	$PTR_ADD $head,$key,0
1710	$PTR_ADD $tail,$key,$at
1711	.align 4
1712	.Lswap:
1713	lw $rk0,0($head)
1714	lw $rk1,4($head)
1715	lw $rk2,8($head)
1716	lw $rk3,12($head)
1717	lw $rk4,0($tail)
1718	lw $rk5,4($tail)
1719	lw $rk6,8($tail)
1720	lw $rk7,12($tail)
1721	sw $rk0,0($tail)
1722	sw $rk1,4($tail)
1723	sw $rk2,8($tail)
1724	sw $rk3,12($tail)
1725	$PTR_ADD $head,16
1726	$PTR_SUB $tail,16
1727	sw $rk4,-16($head)
1728	sw $rk5,-12($head)
1729	sw $rk6,-8($head)
1730	sw $rk7,-4($head)
1731	bne $head,$tail,.Lswap
1732
1733	lw $tp1,16($key) # modulo-scheduled
1734	lui $x80808080,0x8080
1735	subu $cnt,1
1736	or $x80808080,0x8080
1737	sll $cnt,2
1738	$PTR_ADD $key,16
1739	lui $x1b1b1b1b,0x1b1b
1740	nor $x7f7f7f7f,$zero,$x80808080
1741	or $x1b1b1b1b,0x1b1b
1742	.align 4
1743	.Lmix:
1744	and $m,$tp1,$x80808080
1745	and $tp2,$tp1,$x7f7f7f7f
1746	srl $tp4,$m,7
1747	addu $tp2,$tp2 # tp2<<1
1748	subu $m,$tp4
1749	and $m,$x1b1b1b1b
1750	xor $tp2,$m
1751
1752	and $m,$tp2,$x80808080
1753	and $tp4,$tp2,$x7f7f7f7f
1754	srl $tp8,$m,7
1755	addu $tp4,$tp4 # tp4<<1
1756	subu $m,$tp8
1757	and $m,$x1b1b1b1b
1758	xor $tp4,$m
1759
1760	and $m,$tp4,$x80808080
1761	and $tp8,$tp4,$x7f7f7f7f
1762	srl $tp9,$m,7
1763	addu $tp8,$tp8 # tp8<<1
1764	subu $m,$tp9
1765	and $m,$x1b1b1b1b
1766	xor $tp8,$m
1767
1768	xor $tp9,$tp8,$tp1
1769	xor $tpe,$tp8,$tp4
1770	xor $tpb,$tp9,$tp2
1771	xor $tpd,$tp9,$tp4
1772
1773	#if defined(_MIPS_ARCH_MIPS32R2) \|\| defined(_MIPS_ARCH_MIPS64R2)
1774	rotr $tp1,$tpd,16
1775	xor $tpe,$tp2
1776	rotr $tp2,$tp9,8
1777	xor $tpe,$tp1
1778	rotr $tp4,$tpb,24
1779	xor $tpe,$tp2
1780	lw $tp1,4($key) # modulo-scheduled
1781	xor $tpe,$tp4
1782	#else
1783	_ror $tp1,$tpd,16
1784	xor $tpe,$tp2
1785	_ror $tp2,$tpd,-16
1786	xor $tpe,$tp1
1787	_ror $tp1,$tp9,8
1788	xor $tpe,$tp2
1789	_ror $tp2,$tp9,-24
1790	xor $tpe,$tp1
1791	_ror $tp1,$tpb,24
1792	xor $tpe,$tp2
1793	_ror $tp2,$tpb,-8
1794	xor $tpe,$tp1
1795	lw $tp1,4($key) # modulo-scheduled
1796	xor $tpe,$tp2
1797	#endif
1798	subu $cnt,1
1799	sw $tpe,0($key)
1800	$PTR_ADD $key,4
1801	bnez $cnt,.Lmix
1802
1803	li $t0,0
1804	.Ldkey_done:
1805	.set noreorder
1806	move $a0,$t0
1807	$REG_L $ra,$FRAMESIZE-1*$SZREG($sp)
1808	$REG_L $fp,$FRAMESIZE-2*$SZREG($sp)
1809	___
1810	$code.=<<___ if ($flavour =~ /nubi/i);
1811	$REG_L $s3,$FRAMESIZE-11*$SZREG($sp)
1812	$REG_L $s2,$FRAMESIZE-12*$SZREG($sp)
1813	$REG_L $s1,$FRAMESIZE-13*$SZREG($sp)
1814	$REG_L $s0,$FRAMESIZE-14*$SZREG($sp)
1815	$REG_L $gp,$FRAMESIZE-15*$SZREG($sp)
1816	___
1817	$code.=<<___;
1818	jr $ra
1819	$PTR_ADD $sp,$FRAMESIZE
1820	.end AES_set_decrypt_key
1821	___
1822	}}}
1823
1824	######################################################################
1825	# Tables are kept in endian-neutral manner
1826	$code.=<<___;
1827	.rdata
1828	.align 10
1829	AES_Te:
1830	.byte 0xc6,0x63,0x63,0xa5, 0xf8,0x7c,0x7c,0x84 # Te0
1831	.byte 0xee,0x77,0x77,0x99, 0xf6,0x7b,0x7b,0x8d
1832	.byte 0xff,0xf2,0xf2,0x0d, 0xd6,0x6b,0x6b,0xbd
1833	.byte 0xde,0x6f,0x6f,0xb1, 0x91,0xc5,0xc5,0x54
1834	.byte 0x60,0x30,0x30,0x50, 0x02,0x01,0x01,0x03
1835	.byte 0xce,0x67,0x67,0xa9, 0x56,0x2b,0x2b,0x7d
1836	.byte 0xe7,0xfe,0xfe,0x19, 0xb5,0xd7,0xd7,0x62
1837	.byte 0x4d,0xab,0xab,0xe6, 0xec,0x76,0x76,0x9a
1838	.byte 0x8f,0xca,0xca,0x45, 0x1f,0x82,0x82,0x9d
1839	.byte 0x89,0xc9,0xc9,0x40, 0xfa,0x7d,0x7d,0x87
1840	.byte 0xef,0xfa,0xfa,0x15, 0xb2,0x59,0x59,0xeb
1841	.byte 0x8e,0x47,0x47,0xc9, 0xfb,0xf0,0xf0,0x0b
1842	.byte 0x41,0xad,0xad,0xec, 0xb3,0xd4,0xd4,0x67
1843	.byte 0x5f,0xa2,0xa2,0xfd, 0x45,0xaf,0xaf,0xea
1844	.byte 0x23,0x9c,0x9c,0xbf, 0x53,0xa4,0xa4,0xf7
1845	.byte 0xe4,0x72,0x72,0x96, 0x9b,0xc0,0xc0,0x5b
1846	.byte 0x75,0xb7,0xb7,0xc2, 0xe1,0xfd,0xfd,0x1c
1847	.byte 0x3d,0x93,0x93,0xae, 0x4c,0x26,0x26,0x6a
1848	.byte 0x6c,0x36,0x36,0x5a, 0x7e,0x3f,0x3f,0x41
1849	.byte 0xf5,0xf7,0xf7,0x02, 0x83,0xcc,0xcc,0x4f
1850	.byte 0x68,0x34,0x34,0x5c, 0x51,0xa5,0xa5,0xf4
1851	.byte 0xd1,0xe5,0xe5,0x34, 0xf9,0xf1,0xf1,0x08
1852	.byte 0xe2,0x71,0x71,0x93, 0xab,0xd8,0xd8,0x73
1853	.byte 0x62,0x31,0x31,0x53, 0x2a,0x15,0x15,0x3f
1854	.byte 0x08,0x04,0x04,0x0c, 0x95,0xc7,0xc7,0x52
1855	.byte 0x46,0x23,0x23,0x65, 0x9d,0xc3,0xc3,0x5e
1856	.byte 0x30,0x18,0x18,0x28, 0x37,0x96,0x96,0xa1
1857	.byte 0x0a,0x05,0x05,0x0f, 0x2f,0x9a,0x9a,0xb5
1858	.byte 0x0e,0x07,0x07,0x09, 0x24,0x12,0x12,0x36
1859	.byte 0x1b,0x80,0x80,0x9b, 0xdf,0xe2,0xe2,0x3d
1860	.byte 0xcd,0xeb,0xeb,0x26, 0x4e,0x27,0x27,0x69
1861	.byte 0x7f,0xb2,0xb2,0xcd, 0xea,0x75,0x75,0x9f
1862	.byte 0x12,0x09,0x09,0x1b, 0x1d,0x83,0x83,0x9e
1863	.byte 0x58,0x2c,0x2c,0x74, 0x34,0x1a,0x1a,0x2e
1864	.byte 0x36,0x1b,0x1b,0x2d, 0xdc,0x6e,0x6e,0xb2
1865	.byte 0xb4,0x5a,0x5a,0xee, 0x5b,0xa0,0xa0,0xfb
1866	.byte 0xa4,0x52,0x52,0xf6, 0x76,0x3b,0x3b,0x4d
1867	.byte 0xb7,0xd6,0xd6,0x61, 0x7d,0xb3,0xb3,0xce
1868	.byte 0x52,0x29,0x29,0x7b, 0xdd,0xe3,0xe3,0x3e
1869	.byte 0x5e,0x2f,0x2f,0x71, 0x13,0x84,0x84,0x97
1870	.byte 0xa6,0x53,0x53,0xf5, 0xb9,0xd1,0xd1,0x68
1871	.byte 0x00,0x00,0x00,0x00, 0xc1,0xed,0xed,0x2c
1872	.byte 0x40,0x20,0x20,0x60, 0xe3,0xfc,0xfc,0x1f
1873	.byte 0x79,0xb1,0xb1,0xc8, 0xb6,0x5b,0x5b,0xed
1874	.byte 0xd4,0x6a,0x6a,0xbe, 0x8d,0xcb,0xcb,0x46
1875	.byte 0x67,0xbe,0xbe,0xd9, 0x72,0x39,0x39,0x4b
1876	.byte 0x94,0x4a,0x4a,0xde, 0x98,0x4c,0x4c,0xd4
1877	.byte 0xb0,0x58,0x58,0xe8, 0x85,0xcf,0xcf,0x4a
1878	.byte 0xbb,0xd0,0xd0,0x6b, 0xc5,0xef,0xef,0x2a
1879	.byte 0x4f,0xaa,0xaa,0xe5, 0xed,0xfb,0xfb,0x16
1880	.byte 0x86,0x43,0x43,0xc5, 0x9a,0x4d,0x4d,0xd7
1881	.byte 0x66,0x33,0x33,0x55, 0x11,0x85,0x85,0x94
1882	.byte 0x8a,0x45,0x45,0xcf, 0xe9,0xf9,0xf9,0x10
1883	.byte 0x04,0x02,0x02,0x06, 0xfe,0x7f,0x7f,0x81
1884	.byte 0xa0,0x50,0x50,0xf0, 0x78,0x3c,0x3c,0x44
1885	.byte 0x25,0x9f,0x9f,0xba, 0x4b,0xa8,0xa8,0xe3
1886	.byte 0xa2,0x51,0x51,0xf3, 0x5d,0xa3,0xa3,0xfe
1887	.byte 0x80,0x40,0x40,0xc0, 0x05,0x8f,0x8f,0x8a
1888	.byte 0x3f,0x92,0x92,0xad, 0x21,0x9d,0x9d,0xbc
1889	.byte 0x70,0x38,0x38,0x48, 0xf1,0xf5,0xf5,0x04
1890	.byte 0x63,0xbc,0xbc,0xdf, 0x77,0xb6,0xb6,0xc1
1891	.byte 0xaf,0xda,0xda,0x75, 0x42,0x21,0x21,0x63
1892	.byte 0x20,0x10,0x10,0x30, 0xe5,0xff,0xff,0x1a
1893	.byte 0xfd,0xf3,0xf3,0x0e, 0xbf,0xd2,0xd2,0x6d
1894	.byte 0x81,0xcd,0xcd,0x4c, 0x18,0x0c,0x0c,0x14
1895	.byte 0x26,0x13,0x13,0x35, 0xc3,0xec,0xec,0x2f
1896	.byte 0xbe,0x5f,0x5f,0xe1, 0x35,0x97,0x97,0xa2
1897	.byte 0x88,0x44,0x44,0xcc, 0x2e,0x17,0x17,0x39
1898	.byte 0x93,0xc4,0xc4,0x57, 0x55,0xa7,0xa7,0xf2
1899	.byte 0xfc,0x7e,0x7e,0x82, 0x7a,0x3d,0x3d,0x47
1900	.byte 0xc8,0x64,0x64,0xac, 0xba,0x5d,0x5d,0xe7
1901	.byte 0x32,0x19,0x19,0x2b, 0xe6,0x73,0x73,0x95
1902	.byte 0xc0,0x60,0x60,0xa0, 0x19,0x81,0x81,0x98
1903	.byte 0x9e,0x4f,0x4f,0xd1, 0xa3,0xdc,0xdc,0x7f
1904	.byte 0x44,0x22,0x22,0x66, 0x54,0x2a,0x2a,0x7e
1905	.byte 0x3b,0x90,0x90,0xab, 0x0b,0x88,0x88,0x83
1906	.byte 0x8c,0x46,0x46,0xca, 0xc7,0xee,0xee,0x29
1907	.byte 0x6b,0xb8,0xb8,0xd3, 0x28,0x14,0x14,0x3c
1908	.byte 0xa7,0xde,0xde,0x79, 0xbc,0x5e,0x5e,0xe2
1909	.byte 0x16,0x0b,0x0b,0x1d, 0xad,0xdb,0xdb,0x76
1910	.byte 0xdb,0xe0,0xe0,0x3b, 0x64,0x32,0x32,0x56
1911	.byte 0x74,0x3a,0x3a,0x4e, 0x14,0x0a,0x0a,0x1e
1912	.byte 0x92,0x49,0x49,0xdb, 0x0c,0x06,0x06,0x0a
1913	.byte 0x48,0x24,0x24,0x6c, 0xb8,0x5c,0x5c,0xe4
1914	.byte 0x9f,0xc2,0xc2,0x5d, 0xbd,0xd3,0xd3,0x6e
1915	.byte 0x43,0xac,0xac,0xef, 0xc4,0x62,0x62,0xa6
1916	.byte 0x39,0x91,0x91,0xa8, 0x31,0x95,0x95,0xa4
1917	.byte 0xd3,0xe4,0xe4,0x37, 0xf2,0x79,0x79,0x8b
1918	.byte 0xd5,0xe7,0xe7,0x32, 0x8b,0xc8,0xc8,0x43
1919	.byte 0x6e,0x37,0x37,0x59, 0xda,0x6d,0x6d,0xb7
1920	.byte 0x01,0x8d,0x8d,0x8c, 0xb1,0xd5,0xd5,0x64
1921	.byte 0x9c,0x4e,0x4e,0xd2, 0x49,0xa9,0xa9,0xe0
1922	.byte 0xd8,0x6c,0x6c,0xb4, 0xac,0x56,0x56,0xfa
1923	.byte 0xf3,0xf4,0xf4,0x07, 0xcf,0xea,0xea,0x25
1924	.byte 0xca,0x65,0x65,0xaf, 0xf4,0x7a,0x7a,0x8e
1925	.byte 0x47,0xae,0xae,0xe9, 0x10,0x08,0x08,0x18
1926	.byte 0x6f,0xba,0xba,0xd5, 0xf0,0x78,0x78,0x88
1927	.byte 0x4a,0x25,0x25,0x6f, 0x5c,0x2e,0x2e,0x72
1928	.byte 0x38,0x1c,0x1c,0x24, 0x57,0xa6,0xa6,0xf1
1929	.byte 0x73,0xb4,0xb4,0xc7, 0x97,0xc6,0xc6,0x51
1930	.byte 0xcb,0xe8,0xe8,0x23, 0xa1,0xdd,0xdd,0x7c
1931	.byte 0xe8,0x74,0x74,0x9c, 0x3e,0x1f,0x1f,0x21
1932	.byte 0x96,0x4b,0x4b,0xdd, 0x61,0xbd,0xbd,0xdc
1933	.byte 0x0d,0x8b,0x8b,0x86, 0x0f,0x8a,0x8a,0x85
1934	.byte 0xe0,0x70,0x70,0x90, 0x7c,0x3e,0x3e,0x42
1935	.byte 0x71,0xb5,0xb5,0xc4, 0xcc,0x66,0x66,0xaa
1936	.byte 0x90,0x48,0x48,0xd8, 0x06,0x03,0x03,0x05
1937	.byte 0xf7,0xf6,0xf6,0x01, 0x1c,0x0e,0x0e,0x12
1938	.byte 0xc2,0x61,0x61,0xa3, 0x6a,0x35,0x35,0x5f
1939	.byte 0xae,0x57,0x57,0xf9, 0x69,0xb9,0xb9,0xd0
1940	.byte 0x17,0x86,0x86,0x91, 0x99,0xc1,0xc1,0x58
1941	.byte 0x3a,0x1d,0x1d,0x27, 0x27,0x9e,0x9e,0xb9
1942	.byte 0xd9,0xe1,0xe1,0x38, 0xeb,0xf8,0xf8,0x13
1943	.byte 0x2b,0x98,0x98,0xb3, 0x22,0x11,0x11,0x33
1944	.byte 0xd2,0x69,0x69,0xbb, 0xa9,0xd9,0xd9,0x70
1945	.byte 0x07,0x8e,0x8e,0x89, 0x33,0x94,0x94,0xa7
1946	.byte 0x2d,0x9b,0x9b,0xb6, 0x3c,0x1e,0x1e,0x22
1947	.byte 0x15,0x87,0x87,0x92, 0xc9,0xe9,0xe9,0x20
1948	.byte 0x87,0xce,0xce,0x49, 0xaa,0x55,0x55,0xff
1949	.byte 0x50,0x28,0x28,0x78, 0xa5,0xdf,0xdf,0x7a
1950	.byte 0x03,0x8c,0x8c,0x8f, 0x59,0xa1,0xa1,0xf8
1951	.byte 0x09,0x89,0x89,0x80, 0x1a,0x0d,0x0d,0x17
1952	.byte 0x65,0xbf,0xbf,0xda, 0xd7,0xe6,0xe6,0x31
1953	.byte 0x84,0x42,0x42,0xc6, 0xd0,0x68,0x68,0xb8
1954	.byte 0x82,0x41,0x41,0xc3, 0x29,0x99,0x99,0xb0
1955	.byte 0x5a,0x2d,0x2d,0x77, 0x1e,0x0f,0x0f,0x11
1956	.byte 0x7b,0xb0,0xb0,0xcb, 0xa8,0x54,0x54,0xfc
1957	.byte 0x6d,0xbb,0xbb,0xd6, 0x2c,0x16,0x16,0x3a
1958
1959	AES_Td:
1960	.byte 0x51,0xf4,0xa7,0x50, 0x7e,0x41,0x65,0x53 # Td0
1961	.byte 0x1a,0x17,0xa4,0xc3, 0x3a,0x27,0x5e,0x96
1962	.byte 0x3b,0xab,0x6b,0xcb, 0x1f,0x9d,0x45,0xf1
1963	.byte 0xac,0xfa,0x58,0xab, 0x4b,0xe3,0x03,0x93
1964	.byte 0x20,0x30,0xfa,0x55, 0xad,0x76,0x6d,0xf6
1965	.byte 0x88,0xcc,0x76,0x91, 0xf5,0x02,0x4c,0x25
1966	.byte 0x4f,0xe5,0xd7,0xfc, 0xc5,0x2a,0xcb,0xd7
1967	.byte 0x26,0x35,0x44,0x80, 0xb5,0x62,0xa3,0x8f
1968	.byte 0xde,0xb1,0x5a,0x49, 0x25,0xba,0x1b,0x67
1969	.byte 0x45,0xea,0x0e,0x98, 0x5d,0xfe,0xc0,0xe1
1970	.byte 0xc3,0x2f,0x75,0x02, 0x81,0x4c,0xf0,0x12
1971	.byte 0x8d,0x46,0x97,0xa3, 0x6b,0xd3,0xf9,0xc6
1972	.byte 0x03,0x8f,0x5f,0xe7, 0x15,0x92,0x9c,0x95
1973	.byte 0xbf,0x6d,0x7a,0xeb, 0x95,0x52,0x59,0xda
1974	.byte 0xd4,0xbe,0x83,0x2d, 0x58,0x74,0x21,0xd3
1975	.byte 0x49,0xe0,0x69,0x29, 0x8e,0xc9,0xc8,0x44
1976	.byte 0x75,0xc2,0x89,0x6a, 0xf4,0x8e,0x79,0x78
1977	.byte 0x99,0x58,0x3e,0x6b, 0x27,0xb9,0x71,0xdd
1978	.byte 0xbe,0xe1,0x4f,0xb6, 0xf0,0x88,0xad,0x17
1979	.byte 0xc9,0x20,0xac,0x66, 0x7d,0xce,0x3a,0xb4
1980	.byte 0x63,0xdf,0x4a,0x18, 0xe5,0x1a,0x31,0x82
1981	.byte 0x97,0x51,0x33,0x60, 0x62,0x53,0x7f,0x45
1982	.byte 0xb1,0x64,0x77,0xe0, 0xbb,0x6b,0xae,0x84
1983	.byte 0xfe,0x81,0xa0,0x1c, 0xf9,0x08,0x2b,0x94
1984	.byte 0x70,0x48,0x68,0x58, 0x8f,0x45,0xfd,0x19
1985	.byte 0x94,0xde,0x6c,0x87, 0x52,0x7b,0xf8,0xb7
1986	.byte 0xab,0x73,0xd3,0x23, 0x72,0x4b,0x02,0xe2
1987	.byte 0xe3,0x1f,0x8f,0x57, 0x66,0x55,0xab,0x2a
1988	.byte 0xb2,0xeb,0x28,0x07, 0x2f,0xb5,0xc2,0x03
1989	.byte 0x86,0xc5,0x7b,0x9a, 0xd3,0x37,0x08,0xa5
1990	.byte 0x30,0x28,0x87,0xf2, 0x23,0xbf,0xa5,0xb2
1991	.byte 0x02,0x03,0x6a,0xba, 0xed,0x16,0x82,0x5c
1992	.byte 0x8a,0xcf,0x1c,0x2b, 0xa7,0x79,0xb4,0x92
1993	.byte 0xf3,0x07,0xf2,0xf0, 0x4e,0x69,0xe2,0xa1
1994	.byte 0x65,0xda,0xf4,0xcd, 0x06,0x05,0xbe,0xd5
1995	.byte 0xd1,0x34,0x62,0x1f, 0xc4,0xa6,0xfe,0x8a
1996	.byte 0x34,0x2e,0x53,0x9d, 0xa2,0xf3,0x55,0xa0
1997	.byte 0x05,0x8a,0xe1,0x32, 0xa4,0xf6,0xeb,0x75
1998	.byte 0x0b,0x83,0xec,0x39, 0x40,0x60,0xef,0xaa
1999	.byte 0x5e,0x71,0x9f,0x06, 0xbd,0x6e,0x10,0x51
2000	.byte 0x3e,0x21,0x8a,0xf9, 0x96,0xdd,0x06,0x3d
2001	.byte 0xdd,0x3e,0x05,0xae, 0x4d,0xe6,0xbd,0x46
2002	.byte 0x91,0x54,0x8d,0xb5, 0x71,0xc4,0x5d,0x05
2003	.byte 0x04,0x06,0xd4,0x6f, 0x60,0x50,0x15,0xff
2004	.byte 0x19,0x98,0xfb,0x24, 0xd6,0xbd,0xe9,0x97
2005	.byte 0x89,0x40,0x43,0xcc, 0x67,0xd9,0x9e,0x77
2006	.byte 0xb0,0xe8,0x42,0xbd, 0x07,0x89,0x8b,0x88
2007	.byte 0xe7,0x19,0x5b,0x38, 0x79,0xc8,0xee,0xdb
2008	.byte 0xa1,0x7c,0x0a,0x47, 0x7c,0x42,0x0f,0xe9
2009	.byte 0xf8,0x84,0x1e,0xc9, 0x00,0x00,0x00,0x00
2010	.byte 0x09,0x80,0x86,0x83, 0x32,0x2b,0xed,0x48
2011	.byte 0x1e,0x11,0x70,0xac, 0x6c,0x5a,0x72,0x4e
2012	.byte 0xfd,0x0e,0xff,0xfb, 0x0f,0x85,0x38,0x56
2013	.byte 0x3d,0xae,0xd5,0x1e, 0x36,0x2d,0x39,0x27
2014	.byte 0x0a,0x0f,0xd9,0x64, 0x68,0x5c,0xa6,0x21
2015	.byte 0x9b,0x5b,0x54,0xd1, 0x24,0x36,0x2e,0x3a
2016	.byte 0x0c,0x0a,0x67,0xb1, 0x93,0x57,0xe7,0x0f
2017	.byte 0xb4,0xee,0x96,0xd2, 0x1b,0x9b,0x91,0x9e
2018	.byte 0x80,0xc0,0xc5,0x4f, 0x61,0xdc,0x20,0xa2
2019	.byte 0x5a,0x77,0x4b,0x69, 0x1c,0x12,0x1a,0x16
2020	.byte 0xe2,0x93,0xba,0x0a, 0xc0,0xa0,0x2a,0xe5
2021	.byte 0x3c,0x22,0xe0,0x43, 0x12,0x1b,0x17,0x1d
2022	.byte 0x0e,0x09,0x0d,0x0b, 0xf2,0x8b,0xc7,0xad
2023	.byte 0x2d,0xb6,0xa8,0xb9, 0x14,0x1e,0xa9,0xc8
2024	.byte 0x57,0xf1,0x19,0x85, 0xaf,0x75,0x07,0x4c
2025	.byte 0xee,0x99,0xdd,0xbb, 0xa3,0x7f,0x60,0xfd
2026	.byte 0xf7,0x01,0x26,0x9f, 0x5c,0x72,0xf5,0xbc
2027	.byte 0x44,0x66,0x3b,0xc5, 0x5b,0xfb,0x7e,0x34
2028	.byte 0x8b,0x43,0x29,0x76, 0xcb,0x23,0xc6,0xdc
2029	.byte 0xb6,0xed,0xfc,0x68, 0xb8,0xe4,0xf1,0x63
2030	.byte 0xd7,0x31,0xdc,0xca, 0x42,0x63,0x85,0x10
2031	.byte 0x13,0x97,0x22,0x40, 0x84,0xc6,0x11,0x20
2032	.byte 0x85,0x4a,0x24,0x7d, 0xd2,0xbb,0x3d,0xf8
2033	.byte 0xae,0xf9,0x32,0x11, 0xc7,0x29,0xa1,0x6d
2034	.byte 0x1d,0x9e,0x2f,0x4b, 0xdc,0xb2,0x30,0xf3
2035	.byte 0x0d,0x86,0x52,0xec, 0x77,0xc1,0xe3,0xd0
2036	.byte 0x2b,0xb3,0x16,0x6c, 0xa9,0x70,0xb9,0x99
2037	.byte 0x11,0x94,0x48,0xfa, 0x47,0xe9,0x64,0x22
2038	.byte 0xa8,0xfc,0x8c,0xc4, 0xa0,0xf0,0x3f,0x1a
2039	.byte 0x56,0x7d,0x2c,0xd8, 0x22,0x33,0x90,0xef
2040	.byte 0x87,0x49,0x4e,0xc7, 0xd9,0x38,0xd1,0xc1
2041	.byte 0x8c,0xca,0xa2,0xfe, 0x98,0xd4,0x0b,0x36
2042	.byte 0xa6,0xf5,0x81,0xcf, 0xa5,0x7a,0xde,0x28
2043	.byte 0xda,0xb7,0x8e,0x26, 0x3f,0xad,0xbf,0xa4
2044	.byte 0x2c,0x3a,0x9d,0xe4, 0x50,0x78,0x92,0x0d
2045	.byte 0x6a,0x5f,0xcc,0x9b, 0x54,0x7e,0x46,0x62
2046	.byte 0xf6,0x8d,0x13,0xc2, 0x90,0xd8,0xb8,0xe8
2047	.byte 0x2e,0x39,0xf7,0x5e, 0x82,0xc3,0xaf,0xf5
2048	.byte 0x9f,0x5d,0x80,0xbe, 0x69,0xd0,0x93,0x7c
2049	.byte 0x6f,0xd5,0x2d,0xa9, 0xcf,0x25,0x12,0xb3
2050	.byte 0xc8,0xac,0x99,0x3b, 0x10,0x18,0x7d,0xa7
2051	.byte 0xe8,0x9c,0x63,0x6e, 0xdb,0x3b,0xbb,0x7b
2052	.byte 0xcd,0x26,0x78,0x09, 0x6e,0x59,0x18,0xf4
2053	.byte 0xec,0x9a,0xb7,0x01, 0x83,0x4f,0x9a,0xa8
2054	.byte 0xe6,0x95,0x6e,0x65, 0xaa,0xff,0xe6,0x7e
2055	.byte 0x21,0xbc,0xcf,0x08, 0xef,0x15,0xe8,0xe6
2056	.byte 0xba,0xe7,0x9b,0xd9, 0x4a,0x6f,0x36,0xce
2057	.byte 0xea,0x9f,0x09,0xd4, 0x29,0xb0,0x7c,0xd6
2058	.byte 0x31,0xa4,0xb2,0xaf, 0x2a,0x3f,0x23,0x31
2059	.byte 0xc6,0xa5,0x94,0x30, 0x35,0xa2,0x66,0xc0
2060	.byte 0x74,0x4e,0xbc,0x37, 0xfc,0x82,0xca,0xa6
2061	.byte 0xe0,0x90,0xd0,0xb0, 0x33,0xa7,0xd8,0x15
2062	.byte 0xf1,0x04,0x98,0x4a, 0x41,0xec,0xda,0xf7
2063	.byte 0x7f,0xcd,0x50,0x0e, 0x17,0x91,0xf6,0x2f
2064	.byte 0x76,0x4d,0xd6,0x8d, 0x43,0xef,0xb0,0x4d
2065	.byte 0xcc,0xaa,0x4d,0x54, 0xe4,0x96,0x04,0xdf
2066	.byte 0x9e,0xd1,0xb5,0xe3, 0x4c,0x6a,0x88,0x1b
2067	.byte 0xc1,0x2c,0x1f,0xb8, 0x46,0x65,0x51,0x7f
2068	.byte 0x9d,0x5e,0xea,0x04, 0x01,0x8c,0x35,0x5d
2069	.byte 0xfa,0x87,0x74,0x73, 0xfb,0x0b,0x41,0x2e
2070	.byte 0xb3,0x67,0x1d,0x5a, 0x92,0xdb,0xd2,0x52
2071	.byte 0xe9,0x10,0x56,0x33, 0x6d,0xd6,0x47,0x13
2072	.byte 0x9a,0xd7,0x61,0x8c, 0x37,0xa1,0x0c,0x7a
2073	.byte 0x59,0xf8,0x14,0x8e, 0xeb,0x13,0x3c,0x89
2074	.byte 0xce,0xa9,0x27,0xee, 0xb7,0x61,0xc9,0x35
2075	.byte 0xe1,0x1c,0xe5,0xed, 0x7a,0x47,0xb1,0x3c
2076	.byte 0x9c,0xd2,0xdf,0x59, 0x55,0xf2,0x73,0x3f
2077	.byte 0x18,0x14,0xce,0x79, 0x73,0xc7,0x37,0xbf
2078	.byte 0x53,0xf7,0xcd,0xea, 0x5f,0xfd,0xaa,0x5b
2079	.byte 0xdf,0x3d,0x6f,0x14, 0x78,0x44,0xdb,0x86
2080	.byte 0xca,0xaf,0xf3,0x81, 0xb9,0x68,0xc4,0x3e
2081	.byte 0x38,0x24,0x34,0x2c, 0xc2,0xa3,0x40,0x5f
2082	.byte 0x16,0x1d,0xc3,0x72, 0xbc,0xe2,0x25,0x0c
2083	.byte 0x28,0x3c,0x49,0x8b, 0xff,0x0d,0x95,0x41
2084	.byte 0x39,0xa8,0x01,0x71, 0x08,0x0c,0xb3,0xde
2085	.byte 0xd8,0xb4,0xe4,0x9c, 0x64,0x56,0xc1,0x90
2086	.byte 0x7b,0xcb,0x84,0x61, 0xd5,0x32,0xb6,0x70
2087	.byte 0x48,0x6c,0x5c,0x74, 0xd0,0xb8,0x57,0x42
2088
2089	.byte 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38 # Td4
2090	.byte 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
2091	.byte 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
2092	.byte 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
2093	.byte 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
2094	.byte 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
2095	.byte 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
2096	.byte 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
2097	.byte 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
2098	.byte 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
2099	.byte 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
2100	.byte 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
2101	.byte 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
2102	.byte 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
2103	.byte 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
2104	.byte 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
2105	.byte 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
2106	.byte 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
2107	.byte 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
2108	.byte 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
2109	.byte 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
2110	.byte 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
2111	.byte 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
2112	.byte 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
2113	.byte 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
2114	.byte 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
2115	.byte 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
2116	.byte 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
2117	.byte 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
2118	.byte 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
2119	.byte 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
2120	.byte 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
2121
2122	AES_Te4:
2123	.byte 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5 # Te4
2124	.byte 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
2125	.byte 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
2126	.byte 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
2127	.byte 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
2128	.byte 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
2129	.byte 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
2130	.byte 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
2131	.byte 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
2132	.byte 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
2133	.byte 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
2134	.byte 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
2135	.byte 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
2136	.byte 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
2137	.byte 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
2138	.byte 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
2139	.byte 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
2140	.byte 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
2141	.byte 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
2142	.byte 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
2143	.byte 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
2144	.byte 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
2145	.byte 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
2146	.byte 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
2147	.byte 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
2148	.byte 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
2149	.byte 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
2150	.byte 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
2151	.byte 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
2152	.byte 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
2153	.byte 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
2154	.byte 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
2155
2156	.byte 0x01,0x00,0x00,0x00, 0x02,0x00,0x00,0x00 # rcon
2157	.byte 0x04,0x00,0x00,0x00, 0x08,0x00,0x00,0x00
2158	.byte 0x10,0x00,0x00,0x00, 0x20,0x00,0x00,0x00
2159	.byte 0x40,0x00,0x00,0x00, 0x80,0x00,0x00,0x00
2160	.byte 0x1B,0x00,0x00,0x00, 0x36,0x00,0x00,0x00
2161	___
2162
2163
2164	foreach (split("\n",$code)) {
2165	s/\`([^\`]*)\`/eval $1/ge;
2166
2167	# made-up _instructions, _xtr, _ins, _ror and _bias, cope
2168	# with byte order dependencies...
2169	if (/^\s+_/) {
2170	s/(_[a-z]+\s+)(\$[0-9]+),([^,]+)(#.)$/$1$2,$2,$3/;
2171
2172	s/_xtr\s+(\$[0-9]+),(\$[0-9]+),([0-9]+(\-2)*)/
2173	sprintf("srl\t$1,$2,%d",$big_endian ? eval($3)
2174	: eval("24-$3"))/e or
2175	s/_ins\s+(\$[0-9]+),(\$[0-9]+),([0-9]+)/
2176	sprintf("sll\t$1,$2,%d",$big_endian ? eval($3)
2177	: eval("24-$3"))/e or
2178	s/_ins2\s+(\$[0-9]+),(\$[0-9]+),([0-9]+)/
2179	sprintf("ins\t$1,$2,%d,8",$big_endian ? eval($3)
2180	: eval("24-$3"))/e or
2181	s/_ror\s+(\$[0-9]+),(\$[0-9]+),(\-?[0-9]+)/
2182	sprintf("srl\t$1,$2,%d",$big_endian ? eval($3)
2183	: eval("$3*-1"))/e or
2184	s/_bias\s+(\$[0-9]+),(\$[0-9]+),([0-9]+)/
2185	sprintf("sll\t$1,$2,%d",$big_endian ? eval($3)
2186	: eval("($3-16)&31"))/e;
2187
2188	s/srl\s+(\$[0-9]+),(\$[0-9]+),\-([0-9]+)/
2189	sprintf("sll\t$1,$2,$3")/e or
2190	s/srl\s+(\$[0-9]+),(\$[0-9]+),0/
2191	sprintf("and\t$1,$2,0xff")/e or
2192	s/(sll\s+\$[0-9]+,\$[0-9]+,0)/#$1/;
2193	}
2194
2195	# convert lwl/lwr and swr/swl to little-endian order
2196	if (!$big_endian && /^\s+[sl]w[lr]\s+/) {
2197	s/([sl]wl.*)([0-9]+)$(\$[0-9]+)$/
2198	sprintf("$1%d($3)",eval("$2-$2%4+($2%4-1)&3"))/e or
2199	s/([sl]wr.*)([0-9]+)$(\$[0-9]+)$/
2200	sprintf("$1%d($3)",eval("$2-$2%4+($2%4+1)&3"))/e;
2201	}
2202
2203	if (!$big_endian) {
2204	s/(rotr\s+\$[0-9]+,\$[0-9]+),([0-9]+)/sprintf("$1,%d",32-$2)/e;
2205	s/(ext\s+\$[0-9]+,\$[0-9]+),([0-9]+),8/sprintf("$1,%d,8",24-$2)/e;
2206	}
2207
2208	print $_,"\n";
2209	}
2210
2211	close STDOUT or die "error closing STDOUT: $!";

注意: 瀏覽 TracBrowser 來幫助您使用儲存庫瀏覽器

source: vbox/trunk/src/libs/openssl-3.1.0/crypto/aes/asm/aes-mips.pl@ 99371

以其他格式下載: