1#!/usr/bin/env perl
2#
3# ====================================================================
4# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
5# project. The module is, however, dual licensed under OpenSSL and
6# CRYPTOGAMS licenses depending on where you obtain it. For further
7# details see http://www.openssl.org/~appro/cryptogams/.
8# ====================================================================
9#
10# Version 4.3.
11#
12# You might fail to appreciate this module performance from the first
13# try. If compared to "vanilla" linux-ia32-icc target, i.e. considered
14# to be *the* best Intel C compiler without -KPIC, performance appears
15# to be virtually identical... But try to re-configure with shared
16# library support... Aha! Intel compiler "suddenly" lags behind by 30%
17# [on P4, more on others]:-) And if compared to position-independent
18# code generated by GNU C, this code performs *more* than *twice* as
19# fast! Yes, all this buzz about PIC means that unlike other hand-
20# coded implementations, this one was explicitly designed to be safe
21# to use even in shared library context... This also means that this
22# code isn't necessarily absolutely fastest "ever," because in order
23# to achieve position independence an extra register has to be
24# off-loaded to stack, which affects the benchmark result.
25#
26# Special note about instruction choice. Do you recall RC4_INT code
27# performing poorly on P4? It might be the time to figure out why.
28# RC4_INT code implies effective address calculations in base+offset*4
29# form. Trouble is that it seems that offset scaling turned to be
30# critical path... At least eliminating scaling resulted in 2.8x RC4
31# performance improvement [as you might recall]. As AES code is hungry
32# for scaling too, I [try to] avoid the latter by favoring off-by-2
33# shifts and masking the result with 0xFF<<2 instead of "boring" 0xFF.
34#
35# As was shown by Dean Gaudet <dean@arctic.org>, the above note turned
36# void. Performance improvement with off-by-2 shifts was observed on
37# intermediate implementation, which was spilling yet another register
38# to stack... Final offset*4 code below runs just a tad faster on P4,
39# but exhibits up to 10% improvement on other cores.
40#
41# Second version is "monolithic" replacement for aes_core.c, which in
42# addition to AES_[de|en]crypt implements private_AES_set_[de|en]cryption_key.
43# This made it possible to implement little-endian variant of the
44# algorithm without modifying the base C code. Motivating factor for
45# the undertaken effort was that it appeared that in tight IA-32
46# register window little-endian flavor could achieve slightly higher
47# Instruction Level Parallelism, and it indeed resulted in up to 15%
48# better performance on most recent µ-archs...
49#
50# Third version adds AES_cbc_encrypt implementation, which resulted in
51# up to 40% performance imrovement of CBC benchmark results. 40% was
52# observed on P4 core, where "overall" imrovement coefficient, i.e. if
53# compared to PIC generated by GCC and in CBC mode, was observed to be
54# as large as 4x:-) CBC performance is virtually identical to ECB now
55# and on some platforms even better, e.g. 17.6 "small" cycles/byte on
56# Opteron, because certain function prologues and epilogues are
57# effectively taken out of the loop...
58#
59# Version 3.2 implements compressed tables and prefetch of these tables
60# in CBC[!] mode. Former means that 3/4 of table references are now
61# misaligned, which unfortunately has negative impact on elder IA-32
62# implementations, Pentium suffered 30% penalty, PIII - 10%.
63#
64# Version 3.3 avoids L1 cache aliasing between stack frame and
65# S-boxes, and 3.4 - L1 cache aliasing even between key schedule. The
66# latter is achieved by copying the key schedule to controlled place in
67# stack. This unfortunately has rather strong impact on small block CBC
68# performance, ~2x deterioration on 16-byte block if compared to 3.3.
69#
70# Version 3.5 checks if there is L1 cache aliasing between user-supplied
71# key schedule and S-boxes and abstains from copying the former if
72# there is no. This allows end-user to consciously retain small block
73# performance by aligning key schedule in specific manner.
74#
75# Version 3.6 compresses Td4 to 256 bytes and prefetches it in ECB.
76#
77# Current ECB performance numbers for 128-bit key in CPU cycles per
78# processed byte [measure commonly used by AES benchmarkers] are:
79#
80#		small footprint		fully unrolled
81# P4		24			22
82# AMD K8	20			19
83# PIII		25			23
84# Pentium	81			78
85#
86# Version 3.7 reimplements outer rounds as "compact." Meaning that
87# first and last rounds reference compact 256 bytes S-box. This means
88# that first round consumes a lot more CPU cycles and that encrypt
89# and decrypt performance becomes asymmetric. Encrypt performance
90# drops by 10-12%, while decrypt - by 20-25%:-( 256 bytes S-box is
91# aggressively pre-fetched.
92#
93# Version 4.0 effectively rolls back to 3.6 and instead implements
94# additional set of functions, _[x86|sse]_AES_[en|de]crypt_compact,
95# which use exclusively 256 byte S-box. These functions are to be
96# called in modes not concealing plain text, such as ECB, or when
97# we're asked to process smaller amount of data [or unconditionally
98# on hyper-threading CPU]. Currently it's called unconditionally from
99# AES_[en|de]crypt, which affects all modes, but CBC. CBC routine
100# still needs to be modified to switch between slower and faster
101# mode when appropriate... But in either case benchmark landscape
102# changes dramatically and below numbers are CPU cycles per processed
103# byte for 128-bit key.
104#
105#		ECB encrypt	ECB decrypt	CBC large chunk
106# P4		56[60]		84[100]		23
107# AMD K8	48[44]		70[79]		18
108# PIII		41[50]		61[91]		24
109# Core 2	32[38]		45[70]		18.5
110# Pentium	120		160		77
111#
112# Version 4.1 switches to compact S-box even in key schedule setup.
113#
114# Version 4.2 prefetches compact S-box in every SSE round or in other
115# words every cache-line is *guaranteed* to be accessed within ~50
116# cycles window. Why just SSE? Because it's needed on hyper-threading
117# CPU! Which is also why it's prefetched with 64 byte stride. Best
118# part is that it has no negative effect on performance:-)
119#
120# Version 4.3 implements switch between compact and non-compact block
121# functions in AES_cbc_encrypt depending on how much data was asked
122# to be processed in one stroke.
123#
124######################################################################
125# Timing attacks are classified in two classes: synchronous when
126# attacker consciously initiates cryptographic operation and collects
127# timing data of various character afterwards, and asynchronous when
128# malicious code is executed on same CPU simultaneously with AES,
129# instruments itself and performs statistical analysis of this data.
130#
131# As far as synchronous attacks go the root to the AES timing
132# vulnerability is twofold. Firstly, of 256 S-box elements at most 160
133# are referred to in single 128-bit block operation. Well, in C
134# implementation with 4 distinct tables it's actually as little as 40
135# references per 256 elements table, but anyway... Secondly, even
136# though S-box elements are clustered into smaller amount of cache-
137# lines, smaller than 160 and even 40, it turned out that for certain
138# plain-text pattern[s] or simply put chosen plain-text and given key
139# few cache-lines remain unaccessed during block operation. Now, if
140# attacker can figure out this access pattern, he can deduct the key
141# [or at least part of it]. The natural way to mitigate this kind of
142# attacks is to minimize the amount of cache-lines in S-box and/or
143# prefetch them to ensure that every one is accessed for more uniform
144# timing. But note that *if* plain-text was concealed in such way that
145# input to block function is distributed *uniformly*, then attack
146# wouldn't apply. Now note that some encryption modes, most notably
147# CBC, do mask the plain-text in this exact way [secure cipher output
148# is distributed uniformly]. Yes, one still might find input that
149# would reveal the information about given key, but if amount of
150# candidate inputs to be tried is larger than amount of possible key
151# combinations then attack becomes infeasible. This is why revised
152# AES_cbc_encrypt "dares" to switch to larger S-box when larger chunk
153# of data is to be processed in one stroke. The current size limit of
154# 512 bytes is chosen to provide same [diminishigly low] probability
155# for cache-line to remain untouched in large chunk operation with
156# large S-box as for single block operation with compact S-box and
157# surely needs more careful consideration...
158#
159# As for asynchronous attacks. There are two flavours: attacker code
160# being interleaved with AES on hyper-threading CPU at *instruction*
161# level, and two processes time sharing single core. As for latter.
162# Two vectors. 1. Given that attacker process has higher priority,
163# yield execution to process performing AES just before timer fires
164# off the scheduler, immediately regain control of CPU and analyze the
165# cache state. For this attack to be efficient attacker would have to
166# effectively slow down the operation by several *orders* of magnitute,
167# by ratio of time slice to duration of handful of AES rounds, which
168# unlikely to remain unnoticed. Not to mention that this also means
169# that he would spend correspondigly more time to collect enough
170# statistical data to mount the attack. It's probably appropriate to
171# say that if adeversary reckons that this attack is beneficial and
172# risks to be noticed, you probably have larger problems having him
173# mere opportunity. In other words suggested code design expects you
174# to preclude/mitigate this attack by overall system security design.
175# 2. Attacker manages to make his code interrupt driven. In order for
176# this kind of attack to be feasible, interrupt rate has to be high
177# enough, again comparable to duration of handful of AES rounds. But
178# is there interrupt source of such rate? Hardly, not even 1Gbps NIC
179# generates interrupts at such raging rate...
180#
181# And now back to the former, hyper-threading CPU or more specifically
182# Intel P4. Recall that asynchronous attack implies that malicious
183# code instruments itself. And naturally instrumentation granularity
184# has be noticeably lower than duration of codepath accessing S-box.
185# Given that all cache-lines are accessed during that time that is.
186# Current implementation accesses *all* cache-lines within ~50 cycles
187# window, which is actually *less* than RDTSC latency on Intel P4!
188
189$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
190push(@INC,"${dir}","${dir}../../perlasm");
191require "x86asm.pl";
192
193&asm_init($ARGV[0],"aes-586.pl",$x86only = $ARGV[$#ARGV] eq "386");
194&static_label("AES_Te");
195&static_label("AES_Td");
196
197$s0="eax";
198$s1="ebx";
199$s2="ecx";
200$s3="edx";
201$key="edi";
202$acc="esi";
203$tbl="ebp";
204
205# stack frame layout in _[x86|sse]_AES_* routines, frame is allocated
206# by caller
207$__ra=&DWP(0,"esp");	# return address
208$__s0=&DWP(4,"esp");	# s0 backing store
209$__s1=&DWP(8,"esp");	# s1 backing store
210$__s2=&DWP(12,"esp");	# s2 backing store
211$__s3=&DWP(16,"esp");	# s3 backing store
212$__key=&DWP(20,"esp");	# pointer to key schedule
213$__end=&DWP(24,"esp");	# pointer to end of key schedule
214$__tbl=&DWP(28,"esp");	# %ebp backing store
215
216# stack frame layout in AES_[en|crypt] routines, which differs from
217# above by 4 and overlaps by %ebp backing store
218$_tbl=&DWP(24,"esp");
219$_esp=&DWP(28,"esp");
220
221sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
222
223$speed_limit=512;	# chunks smaller than $speed_limit are
224			# processed with compact routine in CBC mode
225$small_footprint=1;	# $small_footprint=1 code is ~5% slower [on
226			# recent µ-archs], but ~5 times smaller!
227			# I favor compact code to minimize cache
228			# contention and in hope to "collect" 5% back
229			# in real-life applications...
230
231$vertical_spin=0;	# shift "verticaly" defaults to 0, because of
232			# its proof-of-concept status...
233# Note that there is no decvert(), as well as last encryption round is
234# performed with "horizontal" shifts. This is because this "vertical"
235# implementation [one which groups shifts on a given $s[i] to form a
236# "column," unlike "horizontal" one, which groups shifts on different
237# $s[i] to form a "row"] is work in progress. It was observed to run
238# few percents faster on Intel cores, but not AMD. On AMD K8 core it's
239# whole 12% slower:-( So we face a trade-off... Shall it be resolved
240# some day? Till then the code is considered experimental and by
241# default remains dormant...
242
243sub encvert()
244{ my ($te,@s) = @_;
245  my $v0 = $acc, $v1 = $key;
246
247	&mov	($v0,$s[3]);				# copy s3
248	&mov	(&DWP(4,"esp"),$s[2]);			# save s2
249	&mov	($v1,$s[0]);				# copy s0
250	&mov	(&DWP(8,"esp"),$s[1]);			# save s1
251
252	&movz	($s[2],&HB($s[0]));
253	&and	($s[0],0xFF);
254	&mov	($s[0],&DWP(0,$te,$s[0],8));		# s0>>0
255	&shr	($v1,16);
256	&mov	($s[3],&DWP(3,$te,$s[2],8));		# s0>>8
257	&movz	($s[1],&HB($v1));
258	&and	($v1,0xFF);
259	&mov	($s[2],&DWP(2,$te,$v1,8));		# s0>>16
260	 &mov	($v1,$v0);
261	&mov	($s[1],&DWP(1,$te,$s[1],8));		# s0>>24
262
263	&and	($v0,0xFF);
264	&xor	($s[3],&DWP(0,$te,$v0,8));		# s3>>0
265	&movz	($v0,&HB($v1));
266	&shr	($v1,16);
267	&xor	($s[2],&DWP(3,$te,$v0,8));		# s3>>8
268	&movz	($v0,&HB($v1));
269	&and	($v1,0xFF);
270	&xor	($s[1],&DWP(2,$te,$v1,8));		# s3>>16
271	 &mov	($v1,&DWP(4,"esp"));			# restore s2
272	&xor	($s[0],&DWP(1,$te,$v0,8));		# s3>>24
273
274	&mov	($v0,$v1);
275	&and	($v1,0xFF);
276	&xor	($s[2],&DWP(0,$te,$v1,8));		# s2>>0
277	&movz	($v1,&HB($v0));
278	&shr	($v0,16);
279	&xor	($s[1],&DWP(3,$te,$v1,8));		# s2>>8
280	&movz	($v1,&HB($v0));
281	&and	($v0,0xFF);
282	&xor	($s[0],&DWP(2,$te,$v0,8));		# s2>>16
283	 &mov	($v0,&DWP(8,"esp"));			# restore s1
284	&xor	($s[3],&DWP(1,$te,$v1,8));		# s2>>24
285
286	&mov	($v1,$v0);
287	&and	($v0,0xFF);
288	&xor	($s[1],&DWP(0,$te,$v0,8));		# s1>>0
289	&movz	($v0,&HB($v1));
290	&shr	($v1,16);
291	&xor	($s[0],&DWP(3,$te,$v0,8));		# s1>>8
292	&movz	($v0,&HB($v1));
293	&and	($v1,0xFF);
294	&xor	($s[3],&DWP(2,$te,$v1,8));		# s1>>16
295	 &mov	($key,$__key);				# reincarnate v1 as key
296	&xor	($s[2],&DWP(1,$te,$v0,8));		# s1>>24
297}
298
299# Another experimental routine, which features "horizontal spin," but
300# eliminates one reference to stack. Strangely enough runs slower...
301sub enchoriz()
302{ my $v0 = $key, $v1 = $acc;
303
304	&movz	($v0,&LB($s0));			#  3, 2, 1, 0*
305	&rotr	($s2,8);			#  8,11,10, 9
306	&mov	($v1,&DWP(0,$te,$v0,8));	#  0
307	&movz	($v0,&HB($s1));			#  7, 6, 5*, 4
308	&rotr	($s3,16);			# 13,12,15,14
309	&xor	($v1,&DWP(3,$te,$v0,8));	#  5
310	&movz	($v0,&HB($s2));			#  8,11,10*, 9
311	&rotr	($s0,16);			#  1, 0, 3, 2
312	&xor	($v1,&DWP(2,$te,$v0,8));	# 10
313	&movz	($v0,&HB($s3));			# 13,12,15*,14
314	&xor	($v1,&DWP(1,$te,$v0,8));	# 15, t[0] collected
315	&mov	($__s0,$v1);			# t[0] saved
316
317	&movz	($v0,&LB($s1));			#  7, 6, 5, 4*
318	&shr	($s1,16);			#  -, -, 7, 6
319	&mov	($v1,&DWP(0,$te,$v0,8));	#  4
320	&movz	($v0,&LB($s3));			# 13,12,15,14*
321	&xor	($v1,&DWP(2,$te,$v0,8));	# 14
322	&movz	($v0,&HB($s0));			#  1, 0, 3*, 2
323	&and	($s3,0xffff0000);		# 13,12, -, -
324	&xor	($v1,&DWP(1,$te,$v0,8));	#  3
325	&movz	($v0,&LB($s2));			#  8,11,10, 9*
326	&or	($s3,$s1);			# 13,12, 7, 6
327	&xor	($v1,&DWP(3,$te,$v0,8));	#  9, t[1] collected
328	&mov	($s1,$v1);			#  s[1]=t[1]
329
330	&movz	($v0,&LB($s0));			#  1, 0, 3, 2*
331	&shr	($s2,16);			#  -, -, 8,11
332	&mov	($v1,&DWP(2,$te,$v0,8));	#  2
333	&movz	($v0,&HB($s3));			# 13,12, 7*, 6
334	&xor	($v1,&DWP(1,$te,$v0,8));	#  7
335	&movz	($v0,&HB($s2));			#  -, -, 8*,11
336	&xor	($v1,&DWP(0,$te,$v0,8));	#  8
337	&mov	($v0,$s3);
338	&shr	($v0,24);			# 13
339	&xor	($v1,&DWP(3,$te,$v0,8));	# 13, t[2] collected
340
341	&movz	($v0,&LB($s2));			#  -, -, 8,11*
342	&shr	($s0,24);			#  1*
343	&mov	($s2,&DWP(1,$te,$v0,8));	# 11
344	&xor	($s2,&DWP(3,$te,$s0,8));	#  1
345	&mov	($s0,$__s0);			# s[0]=t[0]
346	&movz	($v0,&LB($s3));			# 13,12, 7, 6*
347	&shr	($s3,16);			#   ,  ,13,12
348	&xor	($s2,&DWP(2,$te,$v0,8));	#  6
349	&mov	($key,$__key);			# reincarnate v0 as key
350	&and	($s3,0xff);			#   ,  ,13,12*
351	&mov	($s3,&DWP(0,$te,$s3,8));	# 12
352	&xor	($s3,$s2);			# s[2]=t[3] collected
353	&mov	($s2,$v1);			# s[2]=t[2]
354}
355
356# More experimental code... SSE one... Even though this one eliminates
357# *all* references to stack, it's not faster...
358sub sse_encbody()
359{
360	&movz	($acc,&LB("eax"));		#  0
361	&mov	("ecx",&DWP(0,$tbl,$acc,8));	#  0
362	&pshufw	("mm2","mm0",0x0d);		#  7, 6, 3, 2
363	&movz	("edx",&HB("eax"));		#  1
364	&mov	("edx",&DWP(3,$tbl,"edx",8));	#  1
365	&shr	("eax",16);			#  5, 4
366
367	&movz	($acc,&LB("ebx"));		# 10
368	&xor	("ecx",&DWP(2,$tbl,$acc,8));	# 10
369	&pshufw	("mm6","mm4",0x08);		# 13,12, 9, 8
370	&movz	($acc,&HB("ebx"));		# 11
371	&xor	("edx",&DWP(1,$tbl,$acc,8));	# 11
372	&shr	("ebx",16);			# 15,14
373
374	&movz	($acc,&HB("eax"));		#  5
375	&xor	("ecx",&DWP(3,$tbl,$acc,8));	#  5
376	&movq	("mm3",QWP(16,$key));
377	&movz	($acc,&HB("ebx"));		# 15
378	&xor	("ecx",&DWP(1,$tbl,$acc,8));	# 15
379	&movd	("mm0","ecx");			# t[0] collected
380
381	&movz	($acc,&LB("eax"));		#  4
382	&mov	("ecx",&DWP(0,$tbl,$acc,8));	#  4
383	&movd	("eax","mm2");			#  7, 6, 3, 2
384	&movz	($acc,&LB("ebx"));		# 14
385	&xor	("ecx",&DWP(2,$tbl,$acc,8));	# 14
386	&movd	("ebx","mm6");			# 13,12, 9, 8
387
388	&movz	($acc,&HB("eax"));		#  3
389	&xor	("ecx",&DWP(1,$tbl,$acc,8));	#  3
390	&movz	($acc,&HB("ebx"));		#  9
391	&xor	("ecx",&DWP(3,$tbl,$acc,8));	#  9
392	&movd	("mm1","ecx");			# t[1] collected
393
394	&movz	($acc,&LB("eax"));		#  2
395	&mov	("ecx",&DWP(2,$tbl,$acc,8));	#  2
396	&shr	("eax",16);			#  7, 6
397	&punpckldq	("mm0","mm1");		# t[0,1] collected
398	&movz	($acc,&LB("ebx"));		#  8
399	&xor	("ecx",&DWP(0,$tbl,$acc,8));	#  8
400	&shr	("ebx",16);			# 13,12
401
402	&movz	($acc,&HB("eax"));		#  7
403	&xor	("ecx",&DWP(1,$tbl,$acc,8));	#  7
404	&pxor	("mm0","mm3");
405	&movz	("eax",&LB("eax"));		#  6
406	&xor	("edx",&DWP(2,$tbl,"eax",8));	#  6
407	&pshufw	("mm1","mm0",0x08);		#  5, 4, 1, 0
408	&movz	($acc,&HB("ebx"));		# 13
409	&xor	("ecx",&DWP(3,$tbl,$acc,8));	# 13
410	&xor	("ecx",&DWP(24,$key));		# t[2]
411	&movd	("mm4","ecx");			# t[2] collected
412	&movz	("ebx",&LB("ebx"));		# 12
413	&xor	("edx",&DWP(0,$tbl,"ebx",8));	# 12
414	&shr	("ecx",16);
415	&movd	("eax","mm1");			#  5, 4, 1, 0
416	&mov	("ebx",&DWP(28,$key));		# t[3]
417	&xor	("ebx","edx");
418	&movd	("mm5","ebx");			# t[3] collected
419	&and	("ebx",0xffff0000);
420	&or	("ebx","ecx");
421
422	&punpckldq	("mm4","mm5");		# t[2,3] collected
423}
424
425######################################################################
426# "Compact" block function
427######################################################################
428
429sub enccompact()
430{ my $Fn = mov;
431  while ($#_>5) { pop(@_); $Fn=sub{}; }
432  my ($i,$te,@s)=@_;
433  my $tmp = $key;
434  my $out = $i==3?$s[0]:$acc;
435
436	# $Fn is used in first compact round and its purpose is to
437	# void restoration of some values from stack, so that after
438	# 4xenccompact with extra argument $key value is left there...
439	if ($i==3)  {	&$Fn	($key,$__key);			}##%edx
440	else        {	&mov	($out,$s[0]);			}
441			&and	($out,0xFF);
442	if ($i==1)  {	&shr	($s[0],16);			}#%ebx[1]
443	if ($i==2)  {	&shr	($s[0],24);			}#%ecx[2]
444			&movz	($out,&BP(-128,$te,$out,1));
445
446	if ($i==3)  {	$tmp=$s[1];				}##%eax
447			&movz	($tmp,&HB($s[1]));
448			&movz	($tmp,&BP(-128,$te,$tmp,1));
449			&shl	($tmp,8);
450			&xor	($out,$tmp);
451
452	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$__s0);		}##%ebx
453	else        {	&mov	($tmp,$s[2]);
454			&shr	($tmp,16);			}
455	if ($i==2)  {	&and	($s[1],0xFF);			}#%edx[2]
456			&and	($tmp,0xFF);
457			&movz	($tmp,&BP(-128,$te,$tmp,1));
458			&shl	($tmp,16);
459			&xor	($out,$tmp);
460
461	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}##%ecx
462	elsif($i==2){	&movz	($tmp,&HB($s[3]));		}#%ebx[2]
463	else        {	&mov	($tmp,$s[3]);
464			&shr	($tmp,24);			}
465			&movz	($tmp,&BP(-128,$te,$tmp,1));
466			&shl	($tmp,24);
467			&xor	($out,$tmp);
468	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
469	if ($i==3)  {	&mov	($s[3],$acc);			}
470	&comment();
471}
472
473sub enctransform()
474{ my @s = ($s0,$s1,$s2,$s3);
475  my $i = shift;
476  my $tmp = $tbl;
477  my $r2  = $key ;
478
479	&mov	($acc,$s[$i]);
480	&and	($acc,0x80808080);
481	&mov	($tmp,$acc);
482	&shr	($tmp,7);
483	&lea	($r2,&DWP(0,$s[$i],$s[$i]));
484	&sub	($acc,$tmp);
485	&and	($r2,0xfefefefe);
486	&and	($acc,0x1b1b1b1b);
487	&mov	($tmp,$s[$i]);
488	&xor	($acc,$r2);	# r2
489
490	&xor	($s[$i],$acc);	# r0 ^ r2
491	&rotl	($s[$i],24);
492	&xor	($s[$i],$acc)	# ROTATE(r2^r0,24) ^ r2
493	&rotr	($tmp,16);
494	&xor	($s[$i],$tmp);
495	&rotr	($tmp,8);
496	&xor	($s[$i],$tmp);
497}
498
499&function_begin_B("_x86_AES_encrypt_compact");
500	# note that caller is expected to allocate stack frame for me!
501	&mov	($__key,$key);			# save key
502
503	&xor	($s0,&DWP(0,$key));		# xor with key
504	&xor	($s1,&DWP(4,$key));
505	&xor	($s2,&DWP(8,$key));
506	&xor	($s3,&DWP(12,$key));
507
508	&mov	($acc,&DWP(240,$key));		# load key->rounds
509	&lea	($acc,&DWP(-2,$acc,$acc));
510	&lea	($acc,&DWP(0,$key,$acc,8));
511	&mov	($__end,$acc);			# end of key schedule
512
513	# prefetch Te4
514	&mov	($key,&DWP(0-128,$tbl));
515	&mov	($acc,&DWP(32-128,$tbl));
516	&mov	($key,&DWP(64-128,$tbl));
517	&mov	($acc,&DWP(96-128,$tbl));
518	&mov	($key,&DWP(128-128,$tbl));
519	&mov	($acc,&DWP(160-128,$tbl));
520	&mov	($key,&DWP(192-128,$tbl));
521	&mov	($acc,&DWP(224-128,$tbl));
522
523	&set_label("loop",16);
524
525		&enccompact(0,$tbl,$s0,$s1,$s2,$s3,1);
526		&enccompact(1,$tbl,$s1,$s2,$s3,$s0,1);
527		&enccompact(2,$tbl,$s2,$s3,$s0,$s1,1);
528		&enccompact(3,$tbl,$s3,$s0,$s1,$s2,1);
529		&enctransform(2);
530		&enctransform(3);
531		&enctransform(0);
532		&enctransform(1);
533		&mov 	($key,$__key);
534		&mov	($tbl,$__tbl);
535		&add	($key,16);		# advance rd_key
536		&xor	($s0,&DWP(0,$key));
537		&xor	($s1,&DWP(4,$key));
538		&xor	($s2,&DWP(8,$key));
539		&xor	($s3,&DWP(12,$key));
540
541	&cmp	($key,$__end);
542	&mov	($__key,$key);
543	&jb	(&label("loop"));
544
545	&enccompact(0,$tbl,$s0,$s1,$s2,$s3);
546	&enccompact(1,$tbl,$s1,$s2,$s3,$s0);
547	&enccompact(2,$tbl,$s2,$s3,$s0,$s1);
548	&enccompact(3,$tbl,$s3,$s0,$s1,$s2);
549
550	&xor	($s0,&DWP(16,$key));
551	&xor	($s1,&DWP(20,$key));
552	&xor	($s2,&DWP(24,$key));
553	&xor	($s3,&DWP(28,$key));
554
555	&ret	();
556&function_end_B("_x86_AES_encrypt_compact");
557
558######################################################################
559# "Compact" SSE block function.
560######################################################################
561#
562# Performance is not actually extraordinary in comparison to pure
563# x86 code. In particular encrypt performance is virtually the same.
564# Decrypt performance on the other hand is 15-20% better on newer
565# µ-archs [but we're thankful for *any* improvement here], and ~50%
566# better on PIII:-) And additionally on the pros side this code
567# eliminates redundant references to stack and thus relieves/
568# minimizes the pressure on the memory bus.
569#
570# MMX register layout                           lsb
571# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
572# |          mm4          |          mm0          |
573# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
574# |     s3    |     s2    |     s1    |     s0    |
575# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
576# |15|14|13|12|11|10| 9| 8| 7| 6| 5| 4| 3| 2| 1| 0|
577# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
578#
579# Indexes translate as s[N/4]>>(8*(N%4)), e.g. 5 means s1>>8.
580# In this terms encryption and decryption "compact" permutation
581# matrices can be depicted as following:
582#
583# encryption              lsb	# decryption              lsb
584# +----++----+----+----+----+	# +----++----+----+----+----+
585# | t0 || 15 | 10 |  5 |  0 |	# | t0 ||  7 | 10 | 13 |  0 |
586# +----++----+----+----+----+	# +----++----+----+----+----+
587# | t1 ||  3 | 14 |  9 |  4 |	# | t1 || 11 | 14 |  1 |  4 |
588# +----++----+----+----+----+	# +----++----+----+----+----+
589# | t2 ||  7 |  2 | 13 |  8 |	# | t2 || 15 |  2 |  5 |  8 |
590# +----++----+----+----+----+	# +----++----+----+----+----+
591# | t3 || 11 |  6 |  1 | 12 |	# | t3 ||  3 |  6 |  9 | 12 |
592# +----++----+----+----+----+	# +----++----+----+----+----+
593#
594######################################################################
595# Why not xmm registers? Short answer. It was actually tested and
596# was not any faster, but *contrary*, most notably on Intel CPUs.
597# Longer answer. Main advantage of using mm registers is that movd
598# latency is lower, especially on Intel P4. While arithmetic
599# instructions are twice as many, they can be scheduled every cycle
600# and not every second one when they are operating on xmm register,
601# so that "arithmetic throughput" remains virtually the same. And
602# finally the code can be executed even on elder SSE-only CPUs:-)
603
604sub sse_enccompact()
605{
606	&pshufw	("mm1","mm0",0x08);		#  5, 4, 1, 0
607	&pshufw	("mm5","mm4",0x0d);		# 15,14,11,10
608	&movd	("eax","mm1");			#  5, 4, 1, 0
609	&movd	("ebx","mm5");			# 15,14,11,10
610
611	&movz	($acc,&LB("eax"));		#  0
612	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  0
613	&pshufw	("mm2","mm0",0x0d);		#  7, 6, 3, 2
614	&movz	("edx",&HB("eax"));		#  1
615	&movz	("edx",&BP(-128,$tbl,"edx",1));	#  1
616	&shl	("edx",8);			#  1
617	&shr	("eax",16);			#  5, 4
618
619	&movz	($acc,&LB("ebx"));		# 10
620	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 10
621	&shl	($acc,16);			# 10
622	&or	("ecx",$acc);			# 10
623	&pshufw	("mm6","mm4",0x08);		# 13,12, 9, 8
624	&movz	($acc,&HB("ebx"));		# 11
625	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 11
626	&shl	($acc,24);			# 11
627	&or	("edx",$acc);			# 11
628	&shr	("ebx",16);			# 15,14
629
630	&movz	($acc,&HB("eax"));		#  5
631	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  5
632	&shl	($acc,8);			#  5
633	&or	("ecx",$acc);			#  5
634	&movz	($acc,&HB("ebx"));		# 15
635	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 15
636	&shl	($acc,24);			# 15
637	&or	("ecx",$acc);			# 15
638	&movd	("mm0","ecx");			# t[0] collected
639
640	&movz	($acc,&LB("eax"));		#  4
641	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  4
642	&movd	("eax","mm2");			#  7, 6, 3, 2
643	&movz	($acc,&LB("ebx"));		# 14
644	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 14
645	&shl	($acc,16);			# 14
646	&or	("ecx",$acc);			# 14
647
648	&movd	("ebx","mm6");			# 13,12, 9, 8
649	&movz	($acc,&HB("eax"));		#  3
650	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  3
651	&shl	($acc,24);			#  3
652	&or	("ecx",$acc);			#  3
653	&movz	($acc,&HB("ebx"));		#  9
654	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  9
655	&shl	($acc,8);			#  9
656	&or	("ecx",$acc);			#  9
657	&movd	("mm1","ecx");			# t[1] collected
658
659	&movz	($acc,&LB("ebx"));		#  8
660	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  8
661	&shr	("ebx",16);			# 13,12
662	&movz	($acc,&LB("eax"));		#  2
663	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  2
664	&shl	($acc,16);			#  2
665	&or	("ecx",$acc);			#  2
666	&shr	("eax",16);			#  7, 6
667
668	&punpckldq	("mm0","mm1");		# t[0,1] collected
669
670	&movz	($acc,&HB("eax"));		#  7
671	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  7
672	&shl	($acc,24);			#  7
673	&or	("ecx",$acc);			#  7
674	&and	("eax",0xff);			#  6
675	&movz	("eax",&BP(-128,$tbl,"eax",1));	#  6
676	&shl	("eax",16);			#  6
677	&or	("edx","eax");			#  6
678	&movz	($acc,&HB("ebx"));		# 13
679	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 13
680	&shl	($acc,8);			# 13
681	&or	("ecx",$acc);			# 13
682	&movd	("mm4","ecx");			# t[2] collected
683	&and	("ebx",0xff);			# 12
684	&movz	("ebx",&BP(-128,$tbl,"ebx",1));	# 12
685	&or	("edx","ebx");			# 12
686	&movd	("mm5","edx");			# t[3] collected
687
688	&punpckldq	("mm4","mm5");		# t[2,3] collected
689}
690
691					if (!$x86only) {
692&function_begin_B("_sse_AES_encrypt_compact");
693	&pxor	("mm0",&QWP(0,$key));	#  7, 6, 5, 4, 3, 2, 1, 0
694	&pxor	("mm4",&QWP(8,$key));	# 15,14,13,12,11,10, 9, 8
695
696	# note that caller is expected to allocate stack frame for me!
697	&mov	($acc,&DWP(240,$key));		# load key->rounds
698	&lea	($acc,&DWP(-2,$acc,$acc));
699	&lea	($acc,&DWP(0,$key,$acc,8));
700	&mov	($__end,$acc);			# end of key schedule
701
702	&mov	($s0,0x1b1b1b1b);		# magic constant
703	&mov	(&DWP(8,"esp"),$s0);
704	&mov	(&DWP(12,"esp"),$s0);
705
706	# prefetch Te4
707	&mov	($s0,&DWP(0-128,$tbl));
708	&mov	($s1,&DWP(32-128,$tbl));
709	&mov	($s2,&DWP(64-128,$tbl));
710	&mov	($s3,&DWP(96-128,$tbl));
711	&mov	($s0,&DWP(128-128,$tbl));
712	&mov	($s1,&DWP(160-128,$tbl));
713	&mov	($s2,&DWP(192-128,$tbl));
714	&mov	($s3,&DWP(224-128,$tbl));
715
716	&set_label("loop",16);
717		&sse_enccompact();
718		&add	($key,16);
719		&cmp	($key,$__end);
720		&ja	(&label("out"));
721
722		&movq	("mm2",&QWP(8,"esp"));
723		&pxor	("mm3","mm3");		&pxor	("mm7","mm7");
724		&movq	("mm1","mm0");		&movq	("mm5","mm4");	# r0
725		&pcmpgtb("mm3","mm0");		&pcmpgtb("mm7","mm4");
726		&pand	("mm3","mm2");		&pand	("mm7","mm2");
727		&pshufw	("mm2","mm0",0xb1);	&pshufw	("mm6","mm4",0xb1);# ROTATE(r0,16)
728		&paddb	("mm0","mm0");		&paddb	("mm4","mm4");
729		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# = r2
730		&pshufw	("mm3","mm2",0xb1);	&pshufw	("mm7","mm6",0xb1);# r0
731		&pxor	("mm1","mm0");		&pxor	("mm5","mm4");	# r0^r2
732		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= ROTATE(r0,16)
733
734		&movq	("mm2","mm3");		&movq	("mm6","mm7");
735		&pslld	("mm3",8);		&pslld	("mm7",8);
736		&psrld	("mm2",24);		&psrld	("mm6",24);
737		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= r0<<8
738		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= r0>>24
739
740		&movq	("mm3","mm1");		&movq	("mm7","mm5");
741		&movq	("mm2",&QWP(0,$key));	&movq	("mm6",&QWP(8,$key));
742		&psrld	("mm1",8);		&psrld	("mm5",8);
743		&mov	($s0,&DWP(0-128,$tbl));
744		&pslld	("mm3",24);		&pslld	("mm7",24);
745		&mov	($s1,&DWP(64-128,$tbl));
746		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= (r2^r0)<<8
747		&mov	($s2,&DWP(128-128,$tbl));
748		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= (r2^r0)>>24
749		&mov	($s3,&DWP(192-128,$tbl));
750
751		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");
752	&jmp	(&label("loop"));
753
754	&set_label("out",16);
755	&pxor	("mm0",&QWP(0,$key));
756	&pxor	("mm4",&QWP(8,$key));
757
758	&ret	();
759&function_end_B("_sse_AES_encrypt_compact");
760					}
761
762######################################################################
763# Vanilla block function.
764######################################################################
765
766sub encstep()
767{ my ($i,$te,@s) = @_;
768  my $tmp = $key;
769  my $out = $i==3?$s[0]:$acc;
770
771	# lines marked with #%e?x[i] denote "reordered" instructions...
772	if ($i==3)  {	&mov	($key,$__key);			}##%edx
773	else        {	&mov	($out,$s[0]);
774			&and	($out,0xFF);			}
775	if ($i==1)  {	&shr	($s[0],16);			}#%ebx[1]
776	if ($i==2)  {	&shr	($s[0],24);			}#%ecx[2]
777			&mov	($out,&DWP(0,$te,$out,8));
778
779	if ($i==3)  {	$tmp=$s[1];				}##%eax
780			&movz	($tmp,&HB($s[1]));
781			&xor	($out,&DWP(3,$te,$tmp,8));
782
783	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$__s0);		}##%ebx
784	else        {	&mov	($tmp,$s[2]);
785			&shr	($tmp,16);			}
786	if ($i==2)  {	&and	($s[1],0xFF);			}#%edx[2]
787			&and	($tmp,0xFF);
788			&xor	($out,&DWP(2,$te,$tmp,8));
789
790	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}##%ecx
791	elsif($i==2){	&movz	($tmp,&HB($s[3]));		}#%ebx[2]
792	else        {	&mov	($tmp,$s[3]);
793			&shr	($tmp,24)			}
794			&xor	($out,&DWP(1,$te,$tmp,8));
795	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
796	if ($i==3)  {	&mov	($s[3],$acc);			}
797			&comment();
798}
799
800sub enclast()
801{ my ($i,$te,@s)=@_;
802  my $tmp = $key;
803  my $out = $i==3?$s[0]:$acc;
804
805	if ($i==3)  {	&mov	($key,$__key);			}##%edx
806	else        {	&mov	($out,$s[0]);			}
807			&and	($out,0xFF);
808	if ($i==1)  {	&shr	($s[0],16);			}#%ebx[1]
809	if ($i==2)  {	&shr	($s[0],24);			}#%ecx[2]
810			&mov	($out,&DWP(2,$te,$out,8));
811			&and	($out,0x000000ff);
812
813	if ($i==3)  {	$tmp=$s[1];				}##%eax
814			&movz	($tmp,&HB($s[1]));
815			&mov	($tmp,&DWP(0,$te,$tmp,8));
816			&and	($tmp,0x0000ff00);
817			&xor	($out,$tmp);
818
819	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$__s0);		}##%ebx
820	else        {	&mov	($tmp,$s[2]);
821			&shr	($tmp,16);			}
822	if ($i==2)  {	&and	($s[1],0xFF);			}#%edx[2]
823			&and	($tmp,0xFF);
824			&mov	($tmp,&DWP(0,$te,$tmp,8));
825			&and	($tmp,0x00ff0000);
826			&xor	($out,$tmp);
827
828	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}##%ecx
829	elsif($i==2){	&movz	($tmp,&HB($s[3]));		}#%ebx[2]
830	else        {	&mov	($tmp,$s[3]);
831			&shr	($tmp,24);			}
832			&mov	($tmp,&DWP(2,$te,$tmp,8));
833			&and	($tmp,0xff000000);
834			&xor	($out,$tmp);
835	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
836	if ($i==3)  {	&mov	($s[3],$acc);			}
837}
838
839&function_begin_B("_x86_AES_encrypt");
840	if ($vertical_spin) {
841		# I need high parts of volatile registers to be accessible...
842		&exch	($s1="edi",$key="ebx");
843		&mov	($s2="esi",$acc="ecx");
844	}
845
846	# note that caller is expected to allocate stack frame for me!
847	&mov	($__key,$key);			# save key
848
849	&xor	($s0,&DWP(0,$key));		# xor with key
850	&xor	($s1,&DWP(4,$key));
851	&xor	($s2,&DWP(8,$key));
852	&xor	($s3,&DWP(12,$key));
853
854	&mov	($acc,&DWP(240,$key));		# load key->rounds
855
856	if ($small_footprint) {
857	    &lea	($acc,&DWP(-2,$acc,$acc));
858	    &lea	($acc,&DWP(0,$key,$acc,8));
859	    &mov	($__end,$acc);		# end of key schedule
860
861	    &set_label("loop",16);
862		if ($vertical_spin) {
863		    &encvert($tbl,$s0,$s1,$s2,$s3);
864		} else {
865		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
866		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
867		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
868		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
869		}
870		&add	($key,16);		# advance rd_key
871		&xor	($s0,&DWP(0,$key));
872		&xor	($s1,&DWP(4,$key));
873		&xor	($s2,&DWP(8,$key));
874		&xor	($s3,&DWP(12,$key));
875	    &cmp	($key,$__end);
876	    &mov	($__key,$key);
877	    &jb		(&label("loop"));
878	}
879	else {
880	    &cmp	($acc,10);
881	    &jle	(&label("10rounds"));
882	    &cmp	($acc,12);
883	    &jle	(&label("12rounds"));
884
885	&set_label("14rounds",4);
886	    for ($i=1;$i<3;$i++) {
887		if ($vertical_spin) {
888		    &encvert($tbl,$s0,$s1,$s2,$s3);
889		} else {
890		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
891		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
892		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
893		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
894		}
895		&xor	($s0,&DWP(16*$i+0,$key));
896		&xor	($s1,&DWP(16*$i+4,$key));
897		&xor	($s2,&DWP(16*$i+8,$key));
898		&xor	($s3,&DWP(16*$i+12,$key));
899	    }
900	    &add	($key,32);
901	    &mov	($__key,$key);		# advance rd_key
902	&set_label("12rounds",4);
903	    for ($i=1;$i<3;$i++) {
904		if ($vertical_spin) {
905		    &encvert($tbl,$s0,$s1,$s2,$s3);
906		} else {
907		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
908		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
909		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
910		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
911		}
912		&xor	($s0,&DWP(16*$i+0,$key));
913		&xor	($s1,&DWP(16*$i+4,$key));
914		&xor	($s2,&DWP(16*$i+8,$key));
915		&xor	($s3,&DWP(16*$i+12,$key));
916	    }
917	    &add	($key,32);
918	    &mov	($__key,$key);		# advance rd_key
919	&set_label("10rounds",4);
920	    for ($i=1;$i<10;$i++) {
921		if ($vertical_spin) {
922		    &encvert($tbl,$s0,$s1,$s2,$s3);
923		} else {
924		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
925		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
926		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
927		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
928		}
929		&xor	($s0,&DWP(16*$i+0,$key));
930		&xor	($s1,&DWP(16*$i+4,$key));
931		&xor	($s2,&DWP(16*$i+8,$key));
932		&xor	($s3,&DWP(16*$i+12,$key));
933	    }
934	}
935
936	if ($vertical_spin) {
937	    # "reincarnate" some registers for "horizontal" spin...
938	    &mov	($s1="ebx",$key="edi");
939	    &mov	($s2="ecx",$acc="esi");
940	}
941	&enclast(0,$tbl,$s0,$s1,$s2,$s3);
942	&enclast(1,$tbl,$s1,$s2,$s3,$s0);
943	&enclast(2,$tbl,$s2,$s3,$s0,$s1);
944	&enclast(3,$tbl,$s3,$s0,$s1,$s2);
945
946	&add	($key,$small_footprint?16:160);
947	&xor	($s0,&DWP(0,$key));
948	&xor	($s1,&DWP(4,$key));
949	&xor	($s2,&DWP(8,$key));
950	&xor	($s3,&DWP(12,$key));
951
952	&ret	();
953
954&set_label("AES_Te",64);	# Yes! I keep it in the code segment!
955	&_data_word(0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6);
956	&_data_word(0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591);
957	&_data_word(0x50303060, 0x03010102, 0xa96767ce, 0x7d2b2b56);
958	&_data_word(0x19fefee7, 0x62d7d7b5, 0xe6abab4d, 0x9a7676ec);
959	&_data_word(0x45caca8f, 0x9d82821f, 0x40c9c989, 0x877d7dfa);
960	&_data_word(0x15fafaef, 0xeb5959b2, 0xc947478e, 0x0bf0f0fb);
961	&_data_word(0xecadad41, 0x67d4d4b3, 0xfda2a25f, 0xeaafaf45);
962	&_data_word(0xbf9c9c23, 0xf7a4a453, 0x967272e4, 0x5bc0c09b);
963	&_data_word(0xc2b7b775, 0x1cfdfde1, 0xae93933d, 0x6a26264c);
964	&_data_word(0x5a36366c, 0x413f3f7e, 0x02f7f7f5, 0x4fcccc83);
965	&_data_word(0x5c343468, 0xf4a5a551, 0x34e5e5d1, 0x08f1f1f9);
966	&_data_word(0x937171e2, 0x73d8d8ab, 0x53313162, 0x3f15152a);
967	&_data_word(0x0c040408, 0x52c7c795, 0x65232346, 0x5ec3c39d);
968	&_data_word(0x28181830, 0xa1969637, 0x0f05050a, 0xb59a9a2f);
969	&_data_word(0x0907070e, 0x36121224, 0x9b80801b, 0x3de2e2df);
970	&_data_word(0x26ebebcd, 0x6927274e, 0xcdb2b27f, 0x9f7575ea);
971	&_data_word(0x1b090912, 0x9e83831d, 0x742c2c58, 0x2e1a1a34);
972	&_data_word(0x2d1b1b36, 0xb26e6edc, 0xee5a5ab4, 0xfba0a05b);
973	&_data_word(0xf65252a4, 0x4d3b3b76, 0x61d6d6b7, 0xceb3b37d);
974	&_data_word(0x7b292952, 0x3ee3e3dd, 0x712f2f5e, 0x97848413);
975	&_data_word(0xf55353a6, 0x68d1d1b9, 0x00000000, 0x2cededc1);
976	&_data_word(0x60202040, 0x1ffcfce3, 0xc8b1b179, 0xed5b5bb6);
977	&_data_word(0xbe6a6ad4, 0x46cbcb8d, 0xd9bebe67, 0x4b393972);
978	&_data_word(0xde4a4a94, 0xd44c4c98, 0xe85858b0, 0x4acfcf85);
979	&_data_word(0x6bd0d0bb, 0x2aefefc5, 0xe5aaaa4f, 0x16fbfbed);
980	&_data_word(0xc5434386, 0xd74d4d9a, 0x55333366, 0x94858511);
981	&_data_word(0xcf45458a, 0x10f9f9e9, 0x06020204, 0x817f7ffe);
982	&_data_word(0xf05050a0, 0x443c3c78, 0xba9f9f25, 0xe3a8a84b);
983	&_data_word(0xf35151a2, 0xfea3a35d, 0xc0404080, 0x8a8f8f05);
984	&_data_word(0xad92923f, 0xbc9d9d21, 0x48383870, 0x04f5f5f1);
985	&_data_word(0xdfbcbc63, 0xc1b6b677, 0x75dadaaf, 0x63212142);
986	&_data_word(0x30101020, 0x1affffe5, 0x0ef3f3fd, 0x6dd2d2bf);
987	&_data_word(0x4ccdcd81, 0x140c0c18, 0x35131326, 0x2fececc3);
988	&_data_word(0xe15f5fbe, 0xa2979735, 0xcc444488, 0x3917172e);
989	&_data_word(0x57c4c493, 0xf2a7a755, 0x827e7efc, 0x473d3d7a);
990	&_data_word(0xac6464c8, 0xe75d5dba, 0x2b191932, 0x957373e6);
991	&_data_word(0xa06060c0, 0x98818119, 0xd14f4f9e, 0x7fdcdca3);
992	&_data_word(0x66222244, 0x7e2a2a54, 0xab90903b, 0x8388880b);
993	&_data_word(0xca46468c, 0x29eeeec7, 0xd3b8b86b, 0x3c141428);
994	&_data_word(0x79dedea7, 0xe25e5ebc, 0x1d0b0b16, 0x76dbdbad);
995	&_data_word(0x3be0e0db, 0x56323264, 0x4e3a3a74, 0x1e0a0a14);
996	&_data_word(0xdb494992, 0x0a06060c, 0x6c242448, 0xe45c5cb8);
997	&_data_word(0x5dc2c29f, 0x6ed3d3bd, 0xefacac43, 0xa66262c4);
998	&_data_word(0xa8919139, 0xa4959531, 0x37e4e4d3, 0x8b7979f2);
999	&_data_word(0x32e7e7d5, 0x43c8c88b, 0x5937376e, 0xb76d6dda);
1000	&_data_word(0x8c8d8d01, 0x64d5d5b1, 0xd24e4e9c, 0xe0a9a949);
1001	&_data_word(0xb46c6cd8, 0xfa5656ac, 0x07f4f4f3, 0x25eaeacf);
1002	&_data_word(0xaf6565ca, 0x8e7a7af4, 0xe9aeae47, 0x18080810);
1003	&_data_word(0xd5baba6f, 0x887878f0, 0x6f25254a, 0x722e2e5c);
1004	&_data_word(0x241c1c38, 0xf1a6a657, 0xc7b4b473, 0x51c6c697);
1005	&_data_word(0x23e8e8cb, 0x7cdddda1, 0x9c7474e8, 0x211f1f3e);
1006	&_data_word(0xdd4b4b96, 0xdcbdbd61, 0x868b8b0d, 0x858a8a0f);
1007	&_data_word(0x907070e0, 0x423e3e7c, 0xc4b5b571, 0xaa6666cc);
1008	&_data_word(0xd8484890, 0x05030306, 0x01f6f6f7, 0x120e0e1c);
1009	&_data_word(0xa36161c2, 0x5f35356a, 0xf95757ae, 0xd0b9b969);
1010	&_data_word(0x91868617, 0x58c1c199, 0x271d1d3a, 0xb99e9e27);
1011	&_data_word(0x38e1e1d9, 0x13f8f8eb, 0xb398982b, 0x33111122);
1012	&_data_word(0xbb6969d2, 0x70d9d9a9, 0x898e8e07, 0xa7949433);
1013	&_data_word(0xb69b9b2d, 0x221e1e3c, 0x92878715, 0x20e9e9c9);
1014	&_data_word(0x49cece87, 0xff5555aa, 0x78282850, 0x7adfdfa5);
1015	&_data_word(0x8f8c8c03, 0xf8a1a159, 0x80898909, 0x170d0d1a);
1016	&_data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0);
1017	&_data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e);
1018	&_data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c);
1019
1020#Te4	# four copies of Te4 to choose from to avoid L1 aliasing
1021	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
1022	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
1023	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
1024	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
1025	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
1026	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
1027	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
1028	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
1029	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
1030	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
1031	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
1032	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
1033	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
1034	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
1035	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
1036	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
1037	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
1038	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
1039	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
1040	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
1041	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
1042	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
1043	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
1044	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
1045	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
1046	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
1047	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
1048	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
1049	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
1050	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
1051	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
1052	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
1053
1054	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
1055	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
1056	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
1057	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
1058	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
1059	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
1060	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
1061	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
1062	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
1063	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
1064	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
1065	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
1066	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
1067	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
1068	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
1069	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
1070	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
1071	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
1072	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
1073	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
1074	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
1075	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
1076	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
1077	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
1078	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
1079	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
1080	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
1081	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
1082	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
1083	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
1084	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
1085	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
1086
1087	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
1088	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
1089	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
1090	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
1091	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
1092	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
1093	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
1094	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
1095	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
1096	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
1097	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
1098	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
1099	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
1100	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
1101	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
1102	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
1103	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
1104	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
1105	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
1106	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
1107	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
1108	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
1109	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
1110	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
1111	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
1112	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
1113	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
1114	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
1115	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
1116	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
1117	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
1118	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
1119
1120	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
1121	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
1122	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
1123	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
1124	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
1125	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
1126	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
1127	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
1128	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
1129	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
1130	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
1131	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
1132	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
1133	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
1134	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
1135	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
1136	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
1137	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
1138	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
1139	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
1140	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
1141	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
1142	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
1143	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
1144	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
1145	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
1146	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
1147	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
1148	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
1149	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
1150	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
1151	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
1152#rcon:
1153	&data_word(0x00000001, 0x00000002, 0x00000004, 0x00000008);
1154	&data_word(0x00000010, 0x00000020, 0x00000040, 0x00000080);
1155	&data_word(0x0000001b, 0x00000036, 0x00000000, 0x00000000);
1156	&data_word(0x00000000, 0x00000000, 0x00000000, 0x00000000);
1157&function_end_B("_x86_AES_encrypt");
1158
1159# void AES_encrypt (const void *inp,void *out,const AES_KEY *key);
1160&function_begin("AES_encrypt");
1161	&mov	($acc,&wparam(0));		# load inp
1162	&mov	($key,&wparam(2));		# load key
1163
1164	&mov	($s0,"esp");
1165	&sub	("esp",36);
1166	&and	("esp",-64);			# align to cache-line
1167
1168	# place stack frame just "above" the key schedule
1169	&lea	($s1,&DWP(-64-63,$key));
1170	&sub	($s1,"esp");
1171	&neg	($s1);
1172	&and	($s1,0x3C0);	# modulo 1024, but aligned to cache-line
1173	&sub	("esp",$s1);
1174	&add	("esp",4);	# 4 is reserved for caller's return address
1175	&mov	($_esp,$s0);			# save stack pointer
1176
1177	&call   (&label("pic_point"));          # make it PIC!
1178	&set_label("pic_point");
1179	&blindpop($tbl);
1180	&picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if (!$x86only);
1181	&lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
1182
1183	# pick Te4 copy which can't "overlap" with stack frame or key schedule
1184	&lea	($s1,&DWP(768-4,"esp"));
1185	&sub	($s1,$tbl);
1186	&and	($s1,0x300);
1187	&lea	($tbl,&DWP(2048+128,$tbl,$s1));
1188
1189					if (!$x86only) {
1190	&bt	(&DWP(0,$s0),25);	# check for SSE bit
1191	&jnc	(&label("x86"));
1192
1193	&movq	("mm0",&QWP(0,$acc));
1194	&movq	("mm4",&QWP(8,$acc));
1195	&call	("_sse_AES_encrypt_compact");
1196	&mov	("esp",$_esp);			# restore stack pointer
1197	&mov	($acc,&wparam(1));		# load out
1198	&movq	(&QWP(0,$acc),"mm0");		# write output data
1199	&movq	(&QWP(8,$acc),"mm4");
1200	&emms	();
1201	&function_end_A();
1202					}
1203	&set_label("x86",16);
1204	&mov	($_tbl,$tbl);
1205	&mov	($s0,&DWP(0,$acc));		# load input data
1206	&mov	($s1,&DWP(4,$acc));
1207	&mov	($s2,&DWP(8,$acc));
1208	&mov	($s3,&DWP(12,$acc));
1209	&call	("_x86_AES_encrypt_compact");
1210	&mov	("esp",$_esp);			# restore stack pointer
1211	&mov	($acc,&wparam(1));		# load out
1212	&mov	(&DWP(0,$acc),$s0);		# write output data
1213	&mov	(&DWP(4,$acc),$s1);
1214	&mov	(&DWP(8,$acc),$s2);
1215	&mov	(&DWP(12,$acc),$s3);
1216&function_end("AES_encrypt");
1217
1218#--------------------------------------------------------------------#
1219
1220######################################################################
1221# "Compact" block function
1222######################################################################
1223
1224sub deccompact()
1225{ my $Fn = mov;
1226  while ($#_>5) { pop(@_); $Fn=sub{}; }
1227  my ($i,$td,@s)=@_;
1228  my $tmp = $key;
1229  my $out = $i==3?$s[0]:$acc;
1230
1231	# $Fn is used in first compact round and its purpose is to
1232	# void restoration of some values from stack, so that after
1233	# 4xdeccompact with extra argument $key, $s0 and $s1 values
1234	# are left there...
1235	if($i==3)   {	&$Fn	($key,$__key);			}
1236	else        {	&mov	($out,$s[0]);			}
1237			&and	($out,0xFF);
1238			&movz	($out,&BP(-128,$td,$out,1));
1239
1240	if ($i==3)  {	$tmp=$s[1];				}
1241			&movz	($tmp,&HB($s[1]));
1242			&movz	($tmp,&BP(-128,$td,$tmp,1));
1243			&shl	($tmp,8);
1244			&xor	($out,$tmp);
1245
1246	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$acc);		}
1247	else        {	mov	($tmp,$s[2]);			}
1248			&shr	($tmp,16);
1249			&and	($tmp,0xFF);
1250			&movz	($tmp,&BP(-128,$td,$tmp,1));
1251			&shl	($tmp,16);
1252			&xor	($out,$tmp);
1253
1254	if ($i==3)  {	$tmp=$s[3]; &$Fn ($s[2],$__s1);		}
1255	else        {	&mov	($tmp,$s[3]);			}
1256			&shr	($tmp,24);
1257			&movz	($tmp,&BP(-128,$td,$tmp,1));
1258			&shl	($tmp,24);
1259			&xor	($out,$tmp);
1260	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
1261	if ($i==3)  {	&$Fn	($s[3],$__s0);			}
1262}
1263
1264# must be called with 2,3,0,1 as argument sequence!!!
1265sub dectransform()
1266{ my @s = ($s0,$s1,$s2,$s3);
1267  my $i = shift;
1268  my $tmp = $key;
1269  my $tp2 = @s[($i+2)%4]; $tp2 = @s[2] if ($i==1);
1270  my $tp4 = @s[($i+3)%4]; $tp4 = @s[3] if ($i==1);
1271  my $tp8 = $tbl;
1272
1273	&mov	($acc,$s[$i]);
1274	&and	($acc,0x80808080);
1275	&mov	($tmp,$acc);
1276	&shr	($tmp,7);
1277	&lea	($tp2,&DWP(0,$s[$i],$s[$i]));
1278	&sub	($acc,$tmp);
1279	&and	($tp2,0xfefefefe);
1280	&and	($acc,0x1b1b1b1b);
1281	&xor	($acc,$tp2);
1282	&mov	($tp2,$acc);
1283
1284	&and	($acc,0x80808080);
1285	&mov	($tmp,$acc);
1286	&shr	($tmp,7);
1287	&lea	($tp4,&DWP(0,$tp2,$tp2));
1288	&sub	($acc,$tmp);
1289	&and	($tp4,0xfefefefe);
1290	&and	($acc,0x1b1b1b1b);
1291	 &xor	($tp2,$s[$i]);	# tp2^tp1
1292	&xor	($acc,$tp4);
1293	&mov	($tp4,$acc);
1294
1295	&and	($acc,0x80808080);
1296	&mov	($tmp,$acc);
1297	&shr	($tmp,7);
1298	&lea	($tp8,&DWP(0,$tp4,$tp4));
1299	&sub	($acc,$tmp);
1300	&and	($tp8,0xfefefefe);
1301	&and	($acc,0x1b1b1b1b);
1302	 &xor	($tp4,$s[$i]);	# tp4^tp1
1303	 &rotl	($s[$i],8);	# = ROTATE(tp1,8)
1304	&xor	($tp8,$acc);
1305
1306	&xor	($s[$i],$tp2);
1307	&xor	($tp2,$tp8);
1308	&rotl	($tp2,24);
1309	&xor	($s[$i],$tp4);
1310	&xor	($tp4,$tp8);
1311	&rotl	($tp4,16);
1312	&xor	($s[$i],$tp8);	# ^= tp8^(tp4^tp1)^(tp2^tp1)
1313	&rotl	($tp8,8);
1314	&xor	($s[$i],$tp2);	# ^= ROTATE(tp8^tp2^tp1,24)
1315	&xor	($s[$i],$tp4);	# ^= ROTATE(tp8^tp4^tp1,16)
1316	 &mov	($s[0],$__s0)			if($i==2); #prefetch $s0
1317	 &mov	($s[1],$__s1)			if($i==3); #prefetch $s1
1318	 &mov	($s[2],$__s2)			if($i==1);
1319	&xor	($s[$i],$tp8);	# ^= ROTATE(tp8,8)
1320
1321	&mov	($s[3],$__s3)			if($i==1);
1322	&mov	(&DWP(4+4*$i,"esp"),$s[$i])	if($i>=2);
1323}
1324
1325&function_begin_B("_x86_AES_decrypt_compact");
1326	# note that caller is expected to allocate stack frame for me!
1327	&mov	($__key,$key);			# save key
1328
1329	&xor	($s0,&DWP(0,$key));		# xor with key
1330	&xor	($s1,&DWP(4,$key));
1331	&xor	($s2,&DWP(8,$key));
1332	&xor	($s3,&DWP(12,$key));
1333
1334	&mov	($acc,&DWP(240,$key));		# load key->rounds
1335
1336	&lea	($acc,&DWP(-2,$acc,$acc));
1337	&lea	($acc,&DWP(0,$key,$acc,8));
1338	&mov	($__end,$acc);			# end of key schedule
1339
1340	# prefetch Td4
1341	&mov	($key,&DWP(0-128,$tbl));
1342	&mov	($acc,&DWP(32-128,$tbl));
1343	&mov	($key,&DWP(64-128,$tbl));
1344	&mov	($acc,&DWP(96-128,$tbl));
1345	&mov	($key,&DWP(128-128,$tbl));
1346	&mov	($acc,&DWP(160-128,$tbl));
1347	&mov	($key,&DWP(192-128,$tbl));
1348	&mov	($acc,&DWP(224-128,$tbl));
1349
1350	&set_label("loop",16);
1351
1352		&deccompact(0,$tbl,$s0,$s3,$s2,$s1,1);
1353		&deccompact(1,$tbl,$s1,$s0,$s3,$s2,1);
1354		&deccompact(2,$tbl,$s2,$s1,$s0,$s3,1);
1355		&deccompact(3,$tbl,$s3,$s2,$s1,$s0,1);
1356		&dectransform(2);
1357		&dectransform(3);
1358		&dectransform(0);
1359		&dectransform(1);
1360		&mov 	($key,$__key);
1361		&mov	($tbl,$__tbl);
1362		&add	($key,16);		# advance rd_key
1363		&xor	($s0,&DWP(0,$key));
1364		&xor	($s1,&DWP(4,$key));
1365		&xor	($s2,&DWP(8,$key));
1366		&xor	($s3,&DWP(12,$key));
1367
1368	&cmp	($key,$__end);
1369	&mov	($__key,$key);
1370	&jb	(&label("loop"));
1371
1372	&deccompact(0,$tbl,$s0,$s3,$s2,$s1);
1373	&deccompact(1,$tbl,$s1,$s0,$s3,$s2);
1374	&deccompact(2,$tbl,$s2,$s1,$s0,$s3);
1375	&deccompact(3,$tbl,$s3,$s2,$s1,$s0);
1376
1377	&xor	($s0,&DWP(16,$key));
1378	&xor	($s1,&DWP(20,$key));
1379	&xor	($s2,&DWP(24,$key));
1380	&xor	($s3,&DWP(28,$key));
1381
1382	&ret	();
1383&function_end_B("_x86_AES_decrypt_compact");
1384
1385######################################################################
1386# "Compact" SSE block function.
1387######################################################################
1388
1389sub sse_deccompact()
1390{
1391	&pshufw	("mm1","mm0",0x0c);		#  7, 6, 1, 0
1392	&movd	("eax","mm1");			#  7, 6, 1, 0
1393
1394	&pshufw	("mm5","mm4",0x09);		# 13,12,11,10
1395	&movz	($acc,&LB("eax"));		#  0
1396	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  0
1397	&movd	("ebx","mm5");			# 13,12,11,10
1398	&movz	("edx",&HB("eax"));		#  1
1399	&movz	("edx",&BP(-128,$tbl,"edx",1));	#  1
1400	&shl	("edx",8);			#  1
1401
1402	&pshufw	("mm2","mm0",0x06);		#  3, 2, 5, 4
1403	&movz	($acc,&LB("ebx"));		# 10
1404	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 10
1405	&shl	($acc,16);			# 10
1406	&or	("ecx",$acc);			# 10
1407	&shr	("eax",16);			#  7, 6
1408	&movz	($acc,&HB("ebx"));		# 11
1409	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 11
1410	&shl	($acc,24);			# 11
1411	&or	("edx",$acc);			# 11
1412	&shr	("ebx",16);			# 13,12
1413
1414	&pshufw	("mm6","mm4",0x03);		# 9, 8,15,14
1415	&movz	($acc,&HB("eax"));		#  7
1416	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  7
1417	&shl	($acc,24);			#  7
1418	&or	("ecx",$acc);			#  7
1419	&movz	($acc,&HB("ebx"));		# 13
1420	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 13
1421	&shl	($acc,8);			# 13
1422	&or	("ecx",$acc);			# 13
1423	&movd	("mm0","ecx");			# t[0] collected
1424
1425	&movz	($acc,&LB("eax"));		#  6
1426	&movd	("eax","mm2");			#  3, 2, 5, 4
1427	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  6
1428	&shl	("ecx",16);			#  6
1429	&movz	($acc,&LB("ebx"));		# 12
1430	&movd	("ebx","mm6");			#  9, 8,15,14
1431	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 12
1432	&or	("ecx",$acc);			# 12
1433
1434	&movz	($acc,&LB("eax"));		#  4
1435	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  4
1436	&or	("edx",$acc);			#  4
1437	&movz	($acc,&LB("ebx"));		# 14
1438	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 14
1439	&shl	($acc,16);			# 14
1440	&or	("edx",$acc);			# 14
1441	&movd	("mm1","edx");			# t[1] collected
1442
1443	&movz	($acc,&HB("eax"));		#  5
1444	&movz	("edx",&BP(-128,$tbl,$acc,1));	#  5
1445	&shl	("edx",8);			#  5
1446	&movz	($acc,&HB("ebx"));		# 15
1447	&shr	("eax",16);			#  3, 2
1448	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 15
1449	&shl	($acc,24);			# 15
1450	&or	("edx",$acc);			# 15
1451	&shr	("ebx",16);			#  9, 8
1452
1453	&punpckldq	("mm0","mm1");		# t[0,1] collected
1454
1455	&movz	($acc,&HB("ebx"));		#  9
1456	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  9
1457	&shl	($acc,8);			#  9
1458	&or	("ecx",$acc);			#  9
1459	&and	("ebx",0xff);			#  8
1460	&movz	("ebx",&BP(-128,$tbl,"ebx",1));	#  8
1461	&or	("edx","ebx");			#  8
1462	&movz	($acc,&LB("eax"));		#  2
1463	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  2
1464	&shl	($acc,16);			#  2
1465	&or	("edx",$acc);			#  2
1466	&movd	("mm4","edx");			# t[2] collected
1467	&movz	("eax",&HB("eax"));		#  3
1468	&movz	("eax",&BP(-128,$tbl,"eax",1));	#  3
1469	&shl	("eax",24);			#  3
1470	&or	("ecx","eax");			#  3
1471	&movd	("mm5","ecx");			# t[3] collected
1472
1473	&punpckldq	("mm4","mm5");		# t[2,3] collected
1474}
1475
1476					if (!$x86only) {
1477&function_begin_B("_sse_AES_decrypt_compact");
1478	&pxor	("mm0",&QWP(0,$key));	#  7, 6, 5, 4, 3, 2, 1, 0
1479	&pxor	("mm4",&QWP(8,$key));	# 15,14,13,12,11,10, 9, 8
1480
1481	# note that caller is expected to allocate stack frame for me!
1482	&mov	($acc,&DWP(240,$key));		# load key->rounds
1483	&lea	($acc,&DWP(-2,$acc,$acc));
1484	&lea	($acc,&DWP(0,$key,$acc,8));
1485	&mov	($__end,$acc);			# end of key schedule
1486
1487	&mov	($s0,0x1b1b1b1b);		# magic constant
1488	&mov	(&DWP(8,"esp"),$s0);
1489	&mov	(&DWP(12,"esp"),$s0);
1490
1491	# prefetch Td4
1492	&mov	($s0,&DWP(0-128,$tbl));
1493	&mov	($s1,&DWP(32-128,$tbl));
1494	&mov	($s2,&DWP(64-128,$tbl));
1495	&mov	($s3,&DWP(96-128,$tbl));
1496	&mov	($s0,&DWP(128-128,$tbl));
1497	&mov	($s1,&DWP(160-128,$tbl));
1498	&mov	($s2,&DWP(192-128,$tbl));
1499	&mov	($s3,&DWP(224-128,$tbl));
1500
1501	&set_label("loop",16);
1502		&sse_deccompact();
1503		&add	($key,16);
1504		&cmp	($key,$__end);
1505		&ja	(&label("out"));
1506
1507		# ROTATE(x^y,N) == ROTATE(x,N)^ROTATE(y,N)
1508		&movq	("mm3","mm0");		&movq	("mm7","mm4");
1509		&movq	("mm2","mm0",1);	&movq	("mm6","mm4",1);
1510		&movq	("mm1","mm0");		&movq	("mm5","mm4");
1511		&pshufw	("mm0","mm0",0xb1);	&pshufw	("mm4","mm4",0xb1);# = ROTATE(tp0,16)
1512		&pslld	("mm2",8);		&pslld	("mm6",8);
1513		&psrld	("mm3",8);		&psrld	("mm7",8);
1514		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= tp0<<8
1515		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp0>>8
1516		&pslld	("mm2",16);		&pslld	("mm6",16);
1517		&psrld	("mm3",16);		&psrld	("mm7",16);
1518		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= tp0<<24
1519		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp0>>24
1520
1521		&movq	("mm3",&QWP(8,"esp"));
1522		&pxor	("mm2","mm2");		&pxor	("mm6","mm6");
1523		&pcmpgtb("mm2","mm1");		&pcmpgtb("mm6","mm5");
1524		&pand	("mm2","mm3");		&pand	("mm6","mm3");
1525		&paddb	("mm1","mm1");		&paddb	("mm5","mm5");
1526		&pxor	("mm1","mm2");		&pxor	("mm5","mm6");	# tp2
1527		&movq	("mm3","mm1");		&movq	("mm7","mm5");
1528		&movq	("mm2","mm1");		&movq	("mm6","mm5");
1529		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp2
1530		&pslld	("mm3",24);		&pslld	("mm7",24);
1531		&psrld	("mm2",8);		&psrld	("mm6",8);
1532		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp2<<24
1533		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= tp2>>8
1534
1535		&movq	("mm2",&QWP(8,"esp"));
1536		&pxor	("mm3","mm3");		&pxor	("mm7","mm7");
1537		&pcmpgtb("mm3","mm1");		&pcmpgtb("mm7","mm5");
1538		&pand	("mm3","mm2");		&pand	("mm7","mm2");
1539		&paddb	("mm1","mm1");		&paddb	("mm5","mm5");
1540		&pxor	("mm1","mm3");		&pxor	("mm5","mm7");	# tp4
1541		&pshufw	("mm3","mm1",0xb1);	&pshufw	("mm7","mm5",0xb1);
1542		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp4
1543		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= ROTATE(tp4,16)
1544
1545		&pxor	("mm3","mm3");		&pxor	("mm7","mm7");
1546		&pcmpgtb("mm3","mm1");		&pcmpgtb("mm7","mm5");
1547		&pand	("mm3","mm2");		&pand	("mm7","mm2");
1548		&paddb	("mm1","mm1");		&paddb	("mm5","mm5");
1549		&pxor	("mm1","mm3");		&pxor	("mm5","mm7");	# tp8
1550		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp8
1551		&movq	("mm3","mm1");		&movq	("mm7","mm5");
1552		&pshufw	("mm2","mm1",0xb1);	&pshufw	("mm6","mm5",0xb1);
1553		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= ROTATE(tp8,16)
1554		&pslld	("mm1",8);		&pslld	("mm5",8);
1555		&psrld	("mm3",8);		&psrld	("mm7",8);
1556		&movq	("mm2",&QWP(0,$key));	&movq	("mm6",&QWP(8,$key));
1557		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp8<<8
1558		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp8>>8
1559		&mov	($s0,&DWP(0-128,$tbl));
1560		&pslld	("mm1",16);		&pslld	("mm5",16);
1561		&mov	($s1,&DWP(64-128,$tbl));
1562		&psrld	("mm3",16);		&psrld	("mm7",16);
1563		&mov	($s2,&DWP(128-128,$tbl));
1564		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp8<<24
1565		&mov	($s3,&DWP(192-128,$tbl));
1566		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp8>>24
1567
1568		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");
1569	&jmp	(&label("loop"));
1570
1571	&set_label("out",16);
1572	&pxor	("mm0",&QWP(0,$key));
1573	&pxor	("mm4",&QWP(8,$key));
1574
1575	&ret	();
1576&function_end_B("_sse_AES_decrypt_compact");
1577					}
1578
1579######################################################################
1580# Vanilla block function.
1581######################################################################
1582
1583sub decstep()
1584{ my ($i,$td,@s) = @_;
1585  my $tmp = $key;
1586  my $out = $i==3?$s[0]:$acc;
1587
1588	# no instructions are reordered, as performance appears
1589	# optimal... or rather that all attempts to reorder didn't
1590	# result in better performance [which by the way is not a
1591	# bit lower than ecryption].
1592	if($i==3)   {	&mov	($key,$__key);			}
1593	else        {	&mov	($out,$s[0]);			}
1594			&and	($out,0xFF);
1595			&mov	($out,&DWP(0,$td,$out,8));
1596
1597	if ($i==3)  {	$tmp=$s[1];				}
1598			&movz	($tmp,&HB($s[1]));
1599			&xor	($out,&DWP(3,$td,$tmp,8));
1600
1601	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$acc);		}
1602	else        {	&mov	($tmp,$s[2]);			}
1603			&shr	($tmp,16);
1604			&and	($tmp,0xFF);
1605			&xor	($out,&DWP(2,$td,$tmp,8));
1606
1607	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}
1608	else        {	&mov	($tmp,$s[3]);			}
1609			&shr	($tmp,24);
1610			&xor	($out,&DWP(1,$td,$tmp,8));
1611	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
1612	if ($i==3)  {	&mov	($s[3],$__s0);			}
1613			&comment();
1614}
1615
1616sub declast()
1617{ my ($i,$td,@s)=@_;
1618  my $tmp = $key;
1619  my $out = $i==3?$s[0]:$acc;
1620
1621	if($i==0)   {	&lea	($td,&DWP(2048+128,$td));
1622			&mov	($tmp,&DWP(0-128,$td));
1623			&mov	($acc,&DWP(32-128,$td));
1624			&mov	($tmp,&DWP(64-128,$td));
1625			&mov	($acc,&DWP(96-128,$td));
1626			&mov	($tmp,&DWP(128-128,$td));
1627			&mov	($acc,&DWP(160-128,$td));
1628			&mov	($tmp,&DWP(192-128,$td));
1629			&mov	($acc,&DWP(224-128,$td));
1630			&lea	($td,&DWP(-128,$td));		}
1631	if($i==3)   {	&mov	($key,$__key);			}
1632	else        {	&mov	($out,$s[0]);			}
1633			&and	($out,0xFF);
1634			&movz	($out,&BP(0,$td,$out,1));
1635
1636	if ($i==3)  {	$tmp=$s[1];				}
1637			&movz	($tmp,&HB($s[1]));
1638			&movz	($tmp,&BP(0,$td,$tmp,1));
1639			&shl	($tmp,8);
1640			&xor	($out,$tmp);
1641
1642	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$acc);		}
1643	else        {	mov	($tmp,$s[2]);			}
1644			&shr	($tmp,16);
1645			&and	($tmp,0xFF);
1646			&movz	($tmp,&BP(0,$td,$tmp,1));
1647			&shl	($tmp,16);
1648			&xor	($out,$tmp);
1649
1650	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}
1651	else        {	&mov	($tmp,$s[3]);			}
1652			&shr	($tmp,24);
1653			&movz	($tmp,&BP(0,$td,$tmp,1));
1654			&shl	($tmp,24);
1655			&xor	($out,$tmp);
1656	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
1657	if ($i==3)  {	&mov	($s[3],$__s0);
1658			&lea	($td,&DWP(-2048,$td));		}
1659}
1660
1661&function_begin_B("_x86_AES_decrypt");
1662	# note that caller is expected to allocate stack frame for me!
1663	&mov	($__key,$key);			# save key
1664
1665	&xor	($s0,&DWP(0,$key));		# xor with key
1666	&xor	($s1,&DWP(4,$key));
1667	&xor	($s2,&DWP(8,$key));
1668	&xor	($s3,&DWP(12,$key));
1669
1670	&mov	($acc,&DWP(240,$key));		# load key->rounds
1671
1672	if ($small_footprint) {
1673	    &lea	($acc,&DWP(-2,$acc,$acc));
1674	    &lea	($acc,&DWP(0,$key,$acc,8));
1675	    &mov	($__end,$acc);		# end of key schedule
1676	    &set_label("loop",16);
1677		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
1678		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
1679		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
1680		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
1681		&add	($key,16);		# advance rd_key
1682		&xor	($s0,&DWP(0,$key));
1683		&xor	($s1,&DWP(4,$key));
1684		&xor	($s2,&DWP(8,$key));
1685		&xor	($s3,&DWP(12,$key));
1686	    &cmp	($key,$__end);
1687	    &mov	($__key,$key);
1688	    &jb		(&label("loop"));
1689	}
1690	else {
1691	    &cmp	($acc,10);
1692	    &jle	(&label("10rounds"));
1693	    &cmp	($acc,12);
1694	    &jle	(&label("12rounds"));
1695
1696	&set_label("14rounds",4);
1697	    for ($i=1;$i<3;$i++) {
1698		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
1699		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
1700		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
1701		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
1702		&xor	($s0,&DWP(16*$i+0,$key));
1703		&xor	($s1,&DWP(16*$i+4,$key));
1704		&xor	($s2,&DWP(16*$i+8,$key));
1705		&xor	($s3,&DWP(16*$i+12,$key));
1706	    }
1707	    &add	($key,32);
1708	    &mov	($__key,$key);		# advance rd_key
1709	&set_label("12rounds",4);
1710	    for ($i=1;$i<3;$i++) {
1711		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
1712		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
1713		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
1714		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
1715		&xor	($s0,&DWP(16*$i+0,$key));
1716		&xor	($s1,&DWP(16*$i+4,$key));
1717		&xor	($s2,&DWP(16*$i+8,$key));
1718		&xor	($s3,&DWP(16*$i+12,$key));
1719	    }
1720	    &add	($key,32);
1721	    &mov	($__key,$key);		# advance rd_key
1722	&set_label("10rounds",4);
1723	    for ($i=1;$i<10;$i++) {
1724		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
1725		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
1726		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
1727		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
1728		&xor	($s0,&DWP(16*$i+0,$key));
1729		&xor	($s1,&DWP(16*$i+4,$key));
1730		&xor	($s2,&DWP(16*$i+8,$key));
1731		&xor	($s3,&DWP(16*$i+12,$key));
1732	    }
1733	}
1734
1735	&declast(0,$tbl,$s0,$s3,$s2,$s1);
1736	&declast(1,$tbl,$s1,$s0,$s3,$s2);
1737	&declast(2,$tbl,$s2,$s1,$s0,$s3);
1738	&declast(3,$tbl,$s3,$s2,$s1,$s0);
1739
1740	&add	($key,$small_footprint?16:160);
1741	&xor	($s0,&DWP(0,$key));
1742	&xor	($s1,&DWP(4,$key));
1743	&xor	($s2,&DWP(8,$key));
1744	&xor	($s3,&DWP(12,$key));
1745
1746	&ret	();
1747
1748&set_label("AES_Td",64);	# Yes! I keep it in the code segment!
1749	&_data_word(0x50a7f451, 0x5365417e, 0xc3a4171a, 0x965e273a);
1750	&_data_word(0xcb6bab3b, 0xf1459d1f, 0xab58faac, 0x9303e34b);
1751	&_data_word(0x55fa3020, 0xf66d76ad, 0x9176cc88, 0x254c02f5);
1752	&_data_word(0xfcd7e54f, 0xd7cb2ac5, 0x80443526, 0x8fa362b5);
1753	&_data_word(0x495ab1de, 0x671bba25, 0x980eea45, 0xe1c0fe5d);
1754	&_data_word(0x02752fc3, 0x12f04c81, 0xa397468d, 0xc6f9d36b);
1755	&_data_word(0xe75f8f03, 0x959c9215, 0xeb7a6dbf, 0xda595295);
1756	&_data_word(0x2d83bed4, 0xd3217458, 0x2969e049, 0x44c8c98e);
1757	&_data_word(0x6a89c275, 0x78798ef4, 0x6b3e5899, 0xdd71b927);
1758	&_data_word(0xb64fe1be, 0x17ad88f0, 0x66ac20c9, 0xb43ace7d);
1759	&_data_word(0x184adf63, 0x82311ae5, 0x60335197, 0x457f5362);
1760	&_data_word(0xe07764b1, 0x84ae6bbb, 0x1ca081fe, 0x942b08f9);
1761	&_data_word(0x58684870, 0x19fd458f, 0x876cde94, 0xb7f87b52);
1762	&_data_word(0x23d373ab, 0xe2024b72, 0x578f1fe3, 0x2aab5566);
1763	&_data_word(0x0728ebb2, 0x03c2b52f, 0x9a7bc586, 0xa50837d3);
1764	&_data_word(0xf2872830, 0xb2a5bf23, 0xba6a0302, 0x5c8216ed);
1765	&_data_word(0x2b1ccf8a, 0x92b479a7, 0xf0f207f3, 0xa1e2694e);
1766	&_data_word(0xcdf4da65, 0xd5be0506, 0x1f6234d1, 0x8afea6c4);
1767	&_data_word(0x9d532e34, 0xa055f3a2, 0x32e18a05, 0x75ebf6a4);
1768	&_data_word(0x39ec830b, 0xaaef6040, 0x069f715e, 0x51106ebd);
1769	&_data_word(0xf98a213e, 0x3d06dd96, 0xae053edd, 0x46bde64d);
1770	&_data_word(0xb58d5491, 0x055dc471, 0x6fd40604, 0xff155060);
1771	&_data_word(0x24fb9819, 0x97e9bdd6, 0xcc434089, 0x779ed967);
1772	&_data_word(0xbd42e8b0, 0x888b8907, 0x385b19e7, 0xdbeec879);
1773	&_data_word(0x470a7ca1, 0xe90f427c, 0xc91e84f8, 0x00000000);
1774	&_data_word(0x83868009, 0x48ed2b32, 0xac70111e, 0x4e725a6c);
1775	&_data_word(0xfbff0efd, 0x5638850f, 0x1ed5ae3d, 0x27392d36);
1776	&_data_word(0x64d90f0a, 0x21a65c68, 0xd1545b9b, 0x3a2e3624);
1777	&_data_word(0xb1670a0c, 0x0fe75793, 0xd296eeb4, 0x9e919b1b);
1778	&_data_word(0x4fc5c080, 0xa220dc61, 0x694b775a, 0x161a121c);
1779	&_data_word(0x0aba93e2, 0xe52aa0c0, 0x43e0223c, 0x1d171b12);
1780	&_data_word(0x0b0d090e, 0xadc78bf2, 0xb9a8b62d, 0xc8a91e14);
1781	&_data_word(0x8519f157, 0x4c0775af, 0xbbdd99ee, 0xfd607fa3);
1782	&_data_word(0x9f2601f7, 0xbcf5725c, 0xc53b6644, 0x347efb5b);
1783	&_data_word(0x7629438b, 0xdcc623cb, 0x68fcedb6, 0x63f1e4b8);
1784	&_data_word(0xcadc31d7, 0x10856342, 0x40229713, 0x2011c684);
1785	&_data_word(0x7d244a85, 0xf83dbbd2, 0x1132f9ae, 0x6da129c7);
1786	&_data_word(0x4b2f9e1d, 0xf330b2dc, 0xec52860d, 0xd0e3c177);
1787	&_data_word(0x6c16b32b, 0x99b970a9, 0xfa489411, 0x2264e947);
1788	&_data_word(0xc48cfca8, 0x1a3ff0a0, 0xd82c7d56, 0xef903322);
1789	&_data_word(0xc74e4987, 0xc1d138d9, 0xfea2ca8c, 0x360bd498);
1790	&_data_word(0xcf81f5a6, 0x28de7aa5, 0x268eb7da, 0xa4bfad3f);
1791	&_data_word(0xe49d3a2c, 0x0d927850, 0x9bcc5f6a, 0x62467e54);
1792	&_data_word(0xc2138df6, 0xe8b8d890, 0x5ef7392e, 0xf5afc382);
1793	&_data_word(0xbe805d9f, 0x7c93d069, 0xa92dd56f, 0xb31225cf);
1794	&_data_word(0x3b99acc8, 0xa77d1810, 0x6e639ce8, 0x7bbb3bdb);
1795	&_data_word(0x097826cd, 0xf418596e, 0x01b79aec, 0xa89a4f83);
1796	&_data_word(0x656e95e6, 0x7ee6ffaa, 0x08cfbc21, 0xe6e815ef);
1797	&_data_word(0xd99be7ba, 0xce366f4a, 0xd4099fea, 0xd67cb029);
1798	&_data_word(0xafb2a431, 0x31233f2a, 0x3094a5c6, 0xc066a235);
1799	&_data_word(0x37bc4e74, 0xa6ca82fc, 0xb0d090e0, 0x15d8a733);
1800	&_data_word(0x4a9804f1, 0xf7daec41, 0x0e50cd7f, 0x2ff69117);
1801	&_data_word(0x8dd64d76, 0x4db0ef43, 0x544daacc, 0xdf0496e4);
1802	&_data_word(0xe3b5d19e, 0x1b886a4c, 0xb81f2cc1, 0x7f516546);
1803	&_data_word(0x04ea5e9d, 0x5d358c01, 0x737487fa, 0x2e410bfb);
1804	&_data_word(0x5a1d67b3, 0x52d2db92, 0x335610e9, 0x1347d66d);
1805	&_data_word(0x8c61d79a, 0x7a0ca137, 0x8e14f859, 0x893c13eb);
1806	&_data_word(0xee27a9ce, 0x35c961b7, 0xede51ce1, 0x3cb1477a);
1807	&_data_word(0x59dfd29c, 0x3f73f255, 0x79ce1418, 0xbf37c773);
1808	&_data_word(0xeacdf753, 0x5baafd5f, 0x146f3ddf, 0x86db4478);
1809	&_data_word(0x81f3afca, 0x3ec468b9, 0x2c342438, 0x5f40a3c2);
1810	&_data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff);
1811	&_data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664);
1812	&_data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0);
1813
1814#Td4:	# four copies of Td4 to choose from to avoid L1 aliasing
1815	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1816	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1817	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1818	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1819	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1820	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1821	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1822	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1823	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1824	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1825	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1826	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1827	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1828	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1829	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1830	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1831	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1832	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1833	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1834	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1835	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1836	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1837	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1838	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1839	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1840	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1841	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1842	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1843	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1844	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1845	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1846	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1847
1848	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1849	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1850	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1851	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1852	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1853	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1854	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1855	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1856	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1857	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1858	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1859	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1860	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1861	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1862	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1863	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1864	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1865	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1866	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1867	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1868	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1869	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1870	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1871	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1872	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1873	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1874	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1875	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1876	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1877	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1878	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1879	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1880
1881	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1882	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1883	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1884	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1885	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1886	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1887	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1888	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1889	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1890	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1891	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1892	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1893	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1894	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1895	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1896	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1897	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1898	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1899	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1900	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1901	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1902	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1903	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1904	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1905	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1906	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1907	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1908	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1909	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1910	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1911	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1912	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1913
1914	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1915	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1916	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1917	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1918	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1919	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1920	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1921	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1922	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1923	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1924	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1925	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1926	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1927	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1928	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1929	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1930	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1931	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1932	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1933	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1934	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1935	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1936	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1937	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1938	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1939	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1940	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1941	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1942	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1943	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1944	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1945	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1946&function_end_B("_x86_AES_decrypt");
1947
1948# void AES_decrypt (const void *inp,void *out,const AES_KEY *key);
1949&function_begin("AES_decrypt");
1950	&mov	($acc,&wparam(0));		# load inp
1951	&mov	($key,&wparam(2));		# load key
1952
1953	&mov	($s0,"esp");
1954	&sub	("esp",36);
1955	&and	("esp",-64);			# align to cache-line
1956
1957	# place stack frame just "above" the key schedule
1958	&lea	($s1,&DWP(-64-63,$key));
1959	&sub	($s1,"esp");
1960	&neg	($s1);
1961	&and	($s1,0x3C0);	# modulo 1024, but aligned to cache-line
1962	&sub	("esp",$s1);
1963	&add	("esp",4);	# 4 is reserved for caller's return address
1964	&mov	($_esp,$s0);	# save stack pointer
1965
1966	&call   (&label("pic_point"));          # make it PIC!
1967	&set_label("pic_point");
1968	&blindpop($tbl);
1969	&picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if(!$x86only);
1970	&lea    ($tbl,&DWP(&label("AES_Td")."-".&label("pic_point"),$tbl));
1971
1972	# pick Td4 copy which can't "overlap" with stack frame or key schedule
1973	&lea	($s1,&DWP(768-4,"esp"));
1974	&sub	($s1,$tbl);
1975	&and	($s1,0x300);
1976	&lea	($tbl,&DWP(2048+128,$tbl,$s1));
1977
1978					if (!$x86only) {
1979	&bt	(&DWP(0,$s0),25);	# check for SSE bit
1980	&jnc	(&label("x86"));
1981
1982	&movq	("mm0",&QWP(0,$acc));
1983	&movq	("mm4",&QWP(8,$acc));
1984	&call	("_sse_AES_decrypt_compact");
1985	&mov	("esp",$_esp);			# restore stack pointer
1986	&mov	($acc,&wparam(1));		# load out
1987	&movq	(&QWP(0,$acc),"mm0");		# write output data
1988	&movq	(&QWP(8,$acc),"mm4");
1989	&emms	();
1990	&function_end_A();
1991					}
1992	&set_label("x86",16);
1993	&mov	($_tbl,$tbl);
1994	&mov	($s0,&DWP(0,$acc));		# load input data
1995	&mov	($s1,&DWP(4,$acc));
1996	&mov	($s2,&DWP(8,$acc));
1997	&mov	($s3,&DWP(12,$acc));
1998	&call	("_x86_AES_decrypt_compact");
1999	&mov	("esp",$_esp);			# restore stack pointer
2000	&mov	($acc,&wparam(1));		# load out
2001	&mov	(&DWP(0,$acc),$s0);		# write output data
2002	&mov	(&DWP(4,$acc),$s1);
2003	&mov	(&DWP(8,$acc),$s2);
2004	&mov	(&DWP(12,$acc),$s3);
2005&function_end("AES_decrypt");
2006
2007# void AES_cbc_encrypt (const void char *inp, unsigned char *out,
2008#			size_t length, const AES_KEY *key,
2009#			unsigned char *ivp,const int enc);
2010{
2011# stack frame layout
2012#             -4(%esp)		# return address	 0(%esp)
2013#              0(%esp)		# s0 backing store	 4(%esp)
2014#              4(%esp)		# s1 backing store	 8(%esp)
2015#              8(%esp)		# s2 backing store	12(%esp)
2016#             12(%esp)		# s3 backing store	16(%esp)
2017#             16(%esp)		# key backup		20(%esp)
2018#             20(%esp)		# end of key schedule	24(%esp)
2019#             24(%esp)		# %ebp backup		28(%esp)
2020#             28(%esp)		# %esp backup
2021my $_inp=&DWP(32,"esp");	# copy of wparam(0)
2022my $_out=&DWP(36,"esp");	# copy of wparam(1)
2023my $_len=&DWP(40,"esp");	# copy of wparam(2)
2024my $_key=&DWP(44,"esp");	# copy of wparam(3)
2025my $_ivp=&DWP(48,"esp");	# copy of wparam(4)
2026my $_tmp=&DWP(52,"esp");	# volatile variable
2027#
2028my $ivec=&DWP(60,"esp");	# ivec[16]
2029my $aes_key=&DWP(76,"esp");	# copy of aes_key
2030my $mark=&DWP(76+240,"esp");	# copy of aes_key->rounds
2031
2032&function_begin("AES_cbc_encrypt");
2033	&mov	($s2 eq "ecx"? $s2 : "",&wparam(2));	# load len
2034	&cmp	($s2,0);
2035	&je	(&label("drop_out"));
2036
2037	&call   (&label("pic_point"));		# make it PIC!
2038	&set_label("pic_point");
2039	&blindpop($tbl);
2040	&picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if(!$x86only);
2041
2042	&cmp	(&wparam(5),0);
2043	&lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
2044	&jne	(&label("picked_te"));
2045	&lea	($tbl,&DWP(&label("AES_Td")."-".&label("AES_Te"),$tbl));
2046	&set_label("picked_te");
2047
2048	# one can argue if this is required
2049	&pushf	();
2050	&cld	();
2051
2052	&cmp	($s2,$speed_limit);
2053	&jb	(&label("slow_way"));
2054	&test	($s2,15);
2055	&jnz	(&label("slow_way"));
2056					if (!$x86only) {
2057	&bt	(&DWP(0,$s0),28);	# check for hyper-threading bit
2058	&jc	(&label("slow_way"));
2059					}
2060	# pre-allocate aligned stack frame...
2061	&lea	($acc,&DWP(-80-244,"esp"));
2062	&and	($acc,-64);
2063
2064	# ... and make sure it doesn't alias with $tbl modulo 4096
2065	&mov	($s0,$tbl);
2066	&lea	($s1,&DWP(2048+256,$tbl));
2067	&mov	($s3,$acc);
2068	&and	($s0,0xfff);		# s = %ebp&0xfff
2069	&and	($s1,0xfff);		# e = (%ebp+2048+256)&0xfff
2070	&and	($s3,0xfff);		# p = %esp&0xfff
2071
2072	&cmp	($s3,$s1);		# if (p>=e) %esp =- (p-e);
2073	&jb	(&label("tbl_break_out"));
2074	&sub	($s3,$s1);
2075	&sub	($acc,$s3);
2076	&jmp	(&label("tbl_ok"));
2077	&set_label("tbl_break_out",4);	# else %esp -= (p-s)&0xfff + framesz;
2078	&sub	($s3,$s0);
2079	&and	($s3,0xfff);
2080	&add	($s3,384);
2081	&sub	($acc,$s3);
2082	&set_label("tbl_ok",4);
2083
2084	&lea	($s3,&wparam(0));	# obtain pointer to parameter block
2085	&exch	("esp",$acc);		# allocate stack frame
2086	&add	("esp",4);		# reserve for return address!
2087	&mov	($_tbl,$tbl);		# save %ebp
2088	&mov	($_esp,$acc);		# save %esp
2089
2090	&mov	($s0,&DWP(0,$s3));	# load inp
2091	&mov	($s1,&DWP(4,$s3));	# load out
2092	#&mov	($s2,&DWP(8,$s3));	# load len
2093	&mov	($key,&DWP(12,$s3));	# load key
2094	&mov	($acc,&DWP(16,$s3));	# load ivp
2095	&mov	($s3,&DWP(20,$s3));	# load enc flag
2096
2097	&mov	($_inp,$s0);		# save copy of inp
2098	&mov	($_out,$s1);		# save copy of out
2099	&mov	($_len,$s2);		# save copy of len
2100	&mov	($_key,$key);		# save copy of key
2101	&mov	($_ivp,$acc);		# save copy of ivp
2102
2103	&mov	($mark,0);		# copy of aes_key->rounds = 0;
2104	# do we copy key schedule to stack?
2105	&mov	($s1 eq "ebx" ? $s1 : "",$key);
2106	&mov	($s2 eq "ecx" ? $s2 : "",244/4);
2107	&sub	($s1,$tbl);
2108	&mov	("esi",$key);
2109	&and	($s1,0xfff);
2110	&lea	("edi",$aes_key);
2111	&cmp	($s1,2048+256);
2112	&jb	(&label("do_copy"));
2113	&cmp	($s1,4096-244);
2114	&jb	(&label("skip_copy"));
2115	&set_label("do_copy",4);
2116		&mov	($_key,"edi");
2117		&data_word(0xA5F3F689);	# rep movsd
2118	&set_label("skip_copy");
2119
2120	&mov	($key,16);
2121	&set_label("prefetch_tbl",4);
2122		&mov	($s0,&DWP(0,$tbl));
2123		&mov	($s1,&DWP(32,$tbl));
2124		&mov	($s2,&DWP(64,$tbl));
2125		&mov	($acc,&DWP(96,$tbl));
2126		&lea	($tbl,&DWP(128,$tbl));
2127		&sub	($key,1);
2128	&jnz	(&label("prefetch_tbl"));
2129	&sub	($tbl,2048);
2130
2131	&mov	($acc,$_inp);
2132	&mov	($key,$_ivp);
2133
2134	&cmp	($s3,0);
2135	&je	(&label("fast_decrypt"));
2136
2137#----------------------------- ENCRYPT -----------------------------#
2138	&mov	($s0,&DWP(0,$key));		# load iv
2139	&mov	($s1,&DWP(4,$key));
2140
2141	&set_label("fast_enc_loop",16);
2142		&mov	($s2,&DWP(8,$key));
2143		&mov	($s3,&DWP(12,$key));
2144
2145		&xor	($s0,&DWP(0,$acc));	# xor input data
2146		&xor	($s1,&DWP(4,$acc));
2147		&xor	($s2,&DWP(8,$acc));
2148		&xor	($s3,&DWP(12,$acc));
2149
2150		&mov	($key,$_key);		# load key
2151		&call	("_x86_AES_encrypt");
2152
2153		&mov	($acc,$_inp);		# load inp
2154		&mov	($key,$_out);		# load out
2155
2156		&mov	(&DWP(0,$key),$s0);	# save output data
2157		&mov	(&DWP(4,$key),$s1);
2158		&mov	(&DWP(8,$key),$s2);
2159		&mov	(&DWP(12,$key),$s3);
2160
2161		&lea	($acc,&DWP(16,$acc));	# advance inp
2162		&mov	($s2,$_len);		# load len
2163		&mov	($_inp,$acc);		# save inp
2164		&lea	($s3,&DWP(16,$key));	# advance out
2165		&mov	($_out,$s3);		# save out
2166		&sub	($s2,16);		# decrease len
2167		&mov	($_len,$s2);		# save len
2168	&jnz	(&label("fast_enc_loop"));
2169	&mov	($acc,$_ivp);		# load ivp
2170	&mov	($s2,&DWP(8,$key));	# restore last 2 dwords
2171	&mov	($s3,&DWP(12,$key));
2172	&mov	(&DWP(0,$acc),$s0);	# save ivec
2173	&mov	(&DWP(4,$acc),$s1);
2174	&mov	(&DWP(8,$acc),$s2);
2175	&mov	(&DWP(12,$acc),$s3);
2176
2177	&cmp	($mark,0);		# was the key schedule copied?
2178	&mov	("edi",$_key);
2179	&je	(&label("skip_ezero"));
2180	# zero copy of key schedule
2181	&mov	("ecx",240/4);
2182	&xor	("eax","eax");
2183	&align	(4);
2184	&data_word(0xABF3F689);	# rep stosd
2185	&set_label("skip_ezero")
2186	&mov	("esp",$_esp);
2187	&popf	();
2188    &set_label("drop_out");
2189	&function_end_A();
2190	&pushf	();			# kludge, never executed
2191
2192#----------------------------- DECRYPT -----------------------------#
2193&set_label("fast_decrypt",16);
2194
2195	&cmp	($acc,$_out);
2196	&je	(&label("fast_dec_in_place"));	# in-place processing...
2197
2198	&mov	($_tmp,$key);
2199
2200	&align	(4);
2201	&set_label("fast_dec_loop",16);
2202		&mov	($s0,&DWP(0,$acc));	# read input
2203		&mov	($s1,&DWP(4,$acc));
2204		&mov	($s2,&DWP(8,$acc));
2205		&mov	($s3,&DWP(12,$acc));
2206
2207		&mov	($key,$_key);		# load key
2208		&call	("_x86_AES_decrypt");
2209
2210		&mov	($key,$_tmp);		# load ivp
2211		&mov	($acc,$_len);		# load len
2212		&xor	($s0,&DWP(0,$key));	# xor iv
2213		&xor	($s1,&DWP(4,$key));
2214		&xor	($s2,&DWP(8,$key));
2215		&xor	($s3,&DWP(12,$key));
2216
2217		&mov	($key,$_out);		# load out
2218		&mov	($acc,$_inp);		# load inp
2219
2220		&mov	(&DWP(0,$key),$s0);	# write output
2221		&mov	(&DWP(4,$key),$s1);
2222		&mov	(&DWP(8,$key),$s2);
2223		&mov	(&DWP(12,$key),$s3);
2224
2225		&mov	($s2,$_len);		# load len
2226		&mov	($_tmp,$acc);		# save ivp
2227		&lea	($acc,&DWP(16,$acc));	# advance inp
2228		&mov	($_inp,$acc);		# save inp
2229		&lea	($key,&DWP(16,$key));	# advance out
2230		&mov	($_out,$key);		# save out
2231		&sub	($s2,16);		# decrease len
2232		&mov	($_len,$s2);		# save len
2233	&jnz	(&label("fast_dec_loop"));
2234	&mov	($key,$_tmp);		# load temp ivp
2235	&mov	($acc,$_ivp);		# load user ivp
2236	&mov	($s0,&DWP(0,$key));	# load iv
2237	&mov	($s1,&DWP(4,$key));
2238	&mov	($s2,&DWP(8,$key));
2239	&mov	($s3,&DWP(12,$key));
2240	&mov	(&DWP(0,$acc),$s0);	# copy back to user
2241	&mov	(&DWP(4,$acc),$s1);
2242	&mov	(&DWP(8,$acc),$s2);
2243	&mov	(&DWP(12,$acc),$s3);
2244	&jmp	(&label("fast_dec_out"));
2245
2246    &set_label("fast_dec_in_place",16);
2247	&set_label("fast_dec_in_place_loop");
2248		&mov	($s0,&DWP(0,$acc));	# read input
2249		&mov	($s1,&DWP(4,$acc));
2250		&mov	($s2,&DWP(8,$acc));
2251		&mov	($s3,&DWP(12,$acc));
2252
2253		&lea	($key,$ivec);
2254		&mov	(&DWP(0,$key),$s0);	# copy to temp
2255		&mov	(&DWP(4,$key),$s1);
2256		&mov	(&DWP(8,$key),$s2);
2257		&mov	(&DWP(12,$key),$s3);
2258
2259		&mov	($key,$_key);		# load key
2260		&call	("_x86_AES_decrypt");
2261
2262		&mov	($key,$_ivp);		# load ivp
2263		&mov	($acc,$_out);		# load out
2264		&xor	($s0,&DWP(0,$key));	# xor iv
2265		&xor	($s1,&DWP(4,$key));
2266		&xor	($s2,&DWP(8,$key));
2267		&xor	($s3,&DWP(12,$key));
2268
2269		&mov	(&DWP(0,$acc),$s0);	# write output
2270		&mov	(&DWP(4,$acc),$s1);
2271		&mov	(&DWP(8,$acc),$s2);
2272		&mov	(&DWP(12,$acc),$s3);
2273
2274		&lea	($acc,&DWP(16,$acc));	# advance out
2275		&mov	($_out,$acc);		# save out
2276
2277		&lea	($acc,$ivec);
2278		&mov	($s0,&DWP(0,$acc));	# read temp
2279		&mov	($s1,&DWP(4,$acc));
2280		&mov	($s2,&DWP(8,$acc));
2281		&mov	($s3,&DWP(12,$acc));
2282
2283		&mov	(&DWP(0,$key),$s0);	# copy iv
2284		&mov	(&DWP(4,$key),$s1);
2285		&mov	(&DWP(8,$key),$s2);
2286		&mov	(&DWP(12,$key),$s3);
2287
2288		&mov	($acc,$_inp);		# load inp
2289		&mov	($s2,$_len);		# load len
2290		&lea	($acc,&DWP(16,$acc));	# advance inp
2291		&mov	($_inp,$acc);		# save inp
2292		&sub	($s2,16);		# decrease len
2293		&mov	($_len,$s2);		# save len
2294	&jnz	(&label("fast_dec_in_place_loop"));
2295
2296    &set_label("fast_dec_out",4);
2297	&cmp	($mark,0);		# was the key schedule copied?
2298	&mov	("edi",$_key);
2299	&je	(&label("skip_dzero"));
2300	# zero copy of key schedule
2301	&mov	("ecx",240/4);
2302	&xor	("eax","eax");
2303	&align	(4);
2304	&data_word(0xABF3F689);	# rep stosd
2305	&set_label("skip_dzero")
2306	&mov	("esp",$_esp);
2307	&popf	();
2308	&function_end_A();
2309	&pushf	();			# kludge, never executed
2310
2311#--------------------------- SLOW ROUTINE ---------------------------#
2312&set_label("slow_way",16);
2313
2314	&mov	($s0,&DWP(0,$s0)) if (!$x86only);# load OPENSSL_ia32cap
2315	&mov	($key,&wparam(3));	# load key
2316
2317	# pre-allocate aligned stack frame...
2318	&lea	($acc,&DWP(-80,"esp"));
2319	&and	($acc,-64);
2320
2321	# ... and make sure it doesn't alias with $key modulo 1024
2322	&lea	($s1,&DWP(-80-63,$key));
2323	&sub	($s1,$acc);
2324	&neg	($s1);
2325	&and	($s1,0x3C0);	# modulo 1024, but aligned to cache-line
2326	&sub	($acc,$s1);
2327
2328	# pick S-box copy which can't overlap with stack frame or $key
2329	&lea	($s1,&DWP(768,$acc));
2330	&sub	($s1,$tbl);
2331	&and	($s1,0x300);
2332	&lea	($tbl,&DWP(2048+128,$tbl,$s1));
2333
2334	&lea	($s3,&wparam(0));	# pointer to parameter block
2335
2336	&exch	("esp",$acc);
2337	&add	("esp",4);		# reserve for return address!
2338	&mov	($_tbl,$tbl);		# save %ebp
2339	&mov	($_esp,$acc);		# save %esp
2340	&mov	($_tmp,$s0);		# save OPENSSL_ia32cap
2341
2342	&mov	($s0,&DWP(0,$s3));	# load inp
2343	&mov	($s1,&DWP(4,$s3));	# load out
2344	#&mov	($s2,&DWP(8,$s3));	# load len
2345	#&mov	($key,&DWP(12,$s3));	# load key
2346	&mov	($acc,&DWP(16,$s3));	# load ivp
2347	&mov	($s3,&DWP(20,$s3));	# load enc flag
2348
2349	&mov	($_inp,$s0);		# save copy of inp
2350	&mov	($_out,$s1);		# save copy of out
2351	&mov	($_len,$s2);		# save copy of len
2352	&mov	($_key,$key);		# save copy of key
2353	&mov	($_ivp,$acc);		# save copy of ivp
2354
2355	&mov	($key,$acc);
2356	&mov	($acc,$s0);
2357
2358	&cmp	($s3,0);
2359	&je	(&label("slow_decrypt"));
2360
2361#--------------------------- SLOW ENCRYPT ---------------------------#
2362	&cmp	($s2,16);
2363	&mov	($s3,$s1);
2364	&jb	(&label("slow_enc_tail"));
2365
2366					if (!$x86only) {
2367	&bt	($_tmp,25);		# check for SSE bit
2368	&jnc	(&label("slow_enc_x86"));
2369
2370	&movq	("mm0",&QWP(0,$key));	# load iv
2371	&movq	("mm4",&QWP(8,$key));
2372
2373	&set_label("slow_enc_loop_sse",16);
2374		&pxor	("mm0",&QWP(0,$acc));	# xor input data
2375		&pxor	("mm4",&QWP(8,$acc));
2376
2377		&mov	($key,$_key);
2378		&call	("_sse_AES_encrypt_compact");
2379
2380		&mov	($acc,$_inp);		# load inp
2381		&mov	($key,$_out);		# load out
2382		&mov	($s2,$_len);		# load len
2383
2384		&movq	(&QWP(0,$key),"mm0");	# save output data
2385		&movq	(&QWP(8,$key),"mm4");
2386
2387		&lea	($acc,&DWP(16,$acc));	# advance inp
2388		&mov	($_inp,$acc);		# save inp
2389		&lea	($s3,&DWP(16,$key));	# advance out
2390		&mov	($_out,$s3);		# save out
2391		&sub	($s2,16);		# decrease len
2392		&cmp	($s2,16);
2393		&mov	($_len,$s2);		# save len
2394	&jae	(&label("slow_enc_loop_sse"));
2395	&test	($s2,15);
2396	&jnz	(&label("slow_enc_tail"));
2397	&mov	($acc,$_ivp);		# load ivp
2398	&movq	(&QWP(0,$acc),"mm0");	# save ivec
2399	&movq	(&QWP(8,$acc),"mm4");
2400	&emms	();
2401	&mov	("esp",$_esp);
2402	&popf	();
2403	&function_end_A();
2404	&pushf	();			# kludge, never executed
2405					}
2406    &set_label("slow_enc_x86",16);
2407	&mov	($s0,&DWP(0,$key));	# load iv
2408	&mov	($s1,&DWP(4,$key));
2409
2410	&set_label("slow_enc_loop_x86",4);
2411		&mov	($s2,&DWP(8,$key));
2412		&mov	($s3,&DWP(12,$key));
2413
2414		&xor	($s0,&DWP(0,$acc));	# xor input data
2415		&xor	($s1,&DWP(4,$acc));
2416		&xor	($s2,&DWP(8,$acc));
2417		&xor	($s3,&DWP(12,$acc));
2418
2419		&mov	($key,$_key);		# load key
2420		&call	("_x86_AES_encrypt_compact");
2421
2422		&mov	($acc,$_inp);		# load inp
2423		&mov	($key,$_out);		# load out
2424
2425		&mov	(&DWP(0,$key),$s0);	# save output data
2426		&mov	(&DWP(4,$key),$s1);
2427		&mov	(&DWP(8,$key),$s2);
2428		&mov	(&DWP(12,$key),$s3);
2429
2430		&mov	($s2,$_len);		# load len
2431		&lea	($acc,&DWP(16,$acc));	# advance inp
2432		&mov	($_inp,$acc);		# save inp
2433		&lea	($s3,&DWP(16,$key));	# advance out
2434		&mov	($_out,$s3);		# save out
2435		&sub	($s2,16);		# decrease len
2436		&cmp	($s2,16);
2437		&mov	($_len,$s2);		# save len
2438	&jae	(&label("slow_enc_loop_x86"));
2439	&test	($s2,15);
2440	&jnz	(&label("slow_enc_tail"));
2441	&mov	($acc,$_ivp);		# load ivp
2442	&mov	($s2,&DWP(8,$key));	# restore last dwords
2443	&mov	($s3,&DWP(12,$key));
2444	&mov	(&DWP(0,$acc),$s0);	# save ivec
2445	&mov	(&DWP(4,$acc),$s1);
2446	&mov	(&DWP(8,$acc),$s2);
2447	&mov	(&DWP(12,$acc),$s3);
2448
2449	&mov	("esp",$_esp);
2450	&popf	();
2451	&function_end_A();
2452	&pushf	();			# kludge, never executed
2453
2454    &set_label("slow_enc_tail",16);
2455	&emms	()	if (!$x86only);
2456	&mov	($key eq "edi"? $key:"",$s3);	# load out to edi
2457	&mov	($s1,16);
2458	&sub	($s1,$s2);
2459	&cmp	($key,$acc eq "esi"? $acc:"");	# compare with inp
2460	&je	(&label("enc_in_place"));
2461	&align	(4);
2462	&data_word(0xA4F3F689);	# rep movsb	# copy input
2463	&jmp	(&label("enc_skip_in_place"));
2464    &set_label("enc_in_place");
2465	&lea	($key,&DWP(0,$key,$s2));
2466    &set_label("enc_skip_in_place");
2467	&mov	($s2,$s1);
2468	&xor	($s0,$s0);
2469	&align	(4);
2470	&data_word(0xAAF3F689);	# rep stosb	# zero tail
2471
2472	&mov	($key,$_ivp);			# restore ivp
2473	&mov	($acc,$s3);			# output as input
2474	&mov	($s0,&DWP(0,$key));
2475	&mov	($s1,&DWP(4,$key));
2476	&mov	($_len,16);			# len=16
2477	&jmp	(&label("slow_enc_loop_x86"));	# one more spin...
2478
2479#--------------------------- SLOW DECRYPT ---------------------------#
2480&set_label("slow_decrypt",16);
2481					if (!$x86only) {
2482	&bt	($_tmp,25);		# check for SSE bit
2483	&jnc	(&label("slow_dec_loop_x86"));
2484
2485	&set_label("slow_dec_loop_sse",4);
2486		&movq	("mm0",&QWP(0,$acc));	# read input
2487		&movq	("mm4",&QWP(8,$acc));
2488
2489		&mov	($key,$_key);
2490		&call	("_sse_AES_decrypt_compact");
2491
2492		&mov	($acc,$_inp);		# load inp
2493		&lea	($s0,$ivec);
2494		&mov	($s1,$_out);		# load out
2495		&mov	($s2,$_len);		# load len
2496		&mov	($key,$_ivp);		# load ivp
2497
2498		&movq	("mm1",&QWP(0,$acc));	# re-read input
2499		&movq	("mm5",&QWP(8,$acc));
2500
2501		&pxor	("mm0",&QWP(0,$key));	# xor iv
2502		&pxor	("mm4",&QWP(8,$key));
2503
2504		&movq	(&QWP(0,$key),"mm1");	# copy input to iv
2505		&movq	(&QWP(8,$key),"mm5");
2506
2507		&sub	($s2,16);		# decrease len
2508		&jc	(&label("slow_dec_partial_sse"));
2509
2510		&movq	(&QWP(0,$s1),"mm0");	# write output
2511		&movq	(&QWP(8,$s1),"mm4");
2512
2513		&lea	($s1,&DWP(16,$s1));	# advance out
2514		&mov	($_out,$s1);		# save out
2515		&lea	($acc,&DWP(16,$acc));	# advance inp
2516		&mov	($_inp,$acc);		# save inp
2517		&mov	($_len,$s2);		# save len
2518	&jnz	(&label("slow_dec_loop_sse"));
2519	&emms	();
2520	&mov	("esp",$_esp);
2521	&popf	();
2522	&function_end_A();
2523	&pushf	();			# kludge, never executed
2524
2525    &set_label("slow_dec_partial_sse",16);
2526	&movq	(&QWP(0,$s0),"mm0");	# save output to temp
2527	&movq	(&QWP(8,$s0),"mm4");
2528	&emms	();
2529
2530	&add	($s2 eq "ecx" ? "ecx":"",16);
2531	&mov	("edi",$s1);		# out
2532	&mov	("esi",$s0);		# temp
2533	&align	(4);
2534	&data_word(0xA4F3F689);		# rep movsb # copy partial output
2535
2536	&mov	("esp",$_esp);
2537	&popf	();
2538	&function_end_A();
2539	&pushf	();			# kludge, never executed
2540					}
2541	&set_label("slow_dec_loop_x86",16);
2542		&mov	($s0,&DWP(0,$acc));	# read input
2543		&mov	($s1,&DWP(4,$acc));
2544		&mov	($s2,&DWP(8,$acc));
2545		&mov	($s3,&DWP(12,$acc));
2546
2547		&lea	($key,$ivec);
2548		&mov	(&DWP(0,$key),$s0);	# copy to temp
2549		&mov	(&DWP(4,$key),$s1);
2550		&mov	(&DWP(8,$key),$s2);
2551		&mov	(&DWP(12,$key),$s3);
2552
2553		&mov	($key,$_key);		# load key
2554		&call	("_x86_AES_decrypt_compact");
2555
2556		&mov	($key,$_ivp);		# load ivp
2557		&mov	($acc,$_len);		# load len
2558		&xor	($s0,&DWP(0,$key));	# xor iv
2559		&xor	($s1,&DWP(4,$key));
2560		&xor	($s2,&DWP(8,$key));
2561		&xor	($s3,&DWP(12,$key));
2562
2563		&sub	($acc,16);
2564		&jc	(&label("slow_dec_partial_x86"));
2565
2566		&mov	($_len,$acc);		# save len
2567		&mov	($acc,$_out);		# load out
2568
2569		&mov	(&DWP(0,$acc),$s0);	# write output
2570		&mov	(&DWP(4,$acc),$s1);
2571		&mov	(&DWP(8,$acc),$s2);
2572		&mov	(&DWP(12,$acc),$s3);
2573
2574		&lea	($acc,&DWP(16,$acc));	# advance out
2575		&mov	($_out,$acc);		# save out
2576
2577		&lea	($acc,$ivec);
2578		&mov	($s0,&DWP(0,$acc));	# read temp
2579		&mov	($s1,&DWP(4,$acc));
2580		&mov	($s2,&DWP(8,$acc));
2581		&mov	($s3,&DWP(12,$acc));
2582
2583		&mov	(&DWP(0,$key),$s0);	# copy it to iv
2584		&mov	(&DWP(4,$key),$s1);
2585		&mov	(&DWP(8,$key),$s2);
2586		&mov	(&DWP(12,$key),$s3);
2587
2588		&mov	($acc,$_inp);		# load inp
2589		&lea	($acc,&DWP(16,$acc));	# advance inp
2590		&mov	($_inp,$acc);		# save inp
2591	&jnz	(&label("slow_dec_loop_x86"));
2592	&mov	("esp",$_esp);
2593	&popf	();
2594	&function_end_A();
2595	&pushf	();			# kludge, never executed
2596
2597    &set_label("slow_dec_partial_x86",16);
2598	&lea	($acc,$ivec);
2599	&mov	(&DWP(0,$acc),$s0);	# save output to temp
2600	&mov	(&DWP(4,$acc),$s1);
2601	&mov	(&DWP(8,$acc),$s2);
2602	&mov	(&DWP(12,$acc),$s3);
2603
2604	&mov	($acc,$_inp);
2605	&mov	($s0,&DWP(0,$acc));	# re-read input
2606	&mov	($s1,&DWP(4,$acc));
2607	&mov	($s2,&DWP(8,$acc));
2608	&mov	($s3,&DWP(12,$acc));
2609
2610	&mov	(&DWP(0,$key),$s0);	# copy it to iv
2611	&mov	(&DWP(4,$key),$s1);
2612	&mov	(&DWP(8,$key),$s2);
2613	&mov	(&DWP(12,$key),$s3);
2614
2615	&mov	("ecx",$_len);
2616	&mov	("edi",$_out);
2617	&lea	("esi",$ivec);
2618	&align	(4);
2619	&data_word(0xA4F3F689);		# rep movsb # copy partial output
2620
2621	&mov	("esp",$_esp);
2622	&popf	();
2623&function_end("AES_cbc_encrypt");
2624}
2625
2626#------------------------------------------------------------------#
2627
2628sub enckey()
2629{
2630	&movz	("esi",&LB("edx"));		# rk[i]>>0
2631	&movz	("ebx",&BP(-128,$tbl,"esi",1));
2632	&movz	("esi",&HB("edx"));		# rk[i]>>8
2633	&shl	("ebx",24);
2634	&xor	("eax","ebx");
2635
2636	&movz	("ebx",&BP(-128,$tbl,"esi",1));
2637	&shr	("edx",16);
2638	&movz	("esi",&LB("edx"));		# rk[i]>>16
2639	&xor	("eax","ebx");
2640
2641	&movz	("ebx",&BP(-128,$tbl,"esi",1));
2642	&movz	("esi",&HB("edx"));		# rk[i]>>24
2643	&shl	("ebx",8);
2644	&xor	("eax","ebx");
2645
2646	&movz	("ebx",&BP(-128,$tbl,"esi",1));
2647	&shl	("ebx",16);
2648	&xor	("eax","ebx");
2649
2650	&xor	("eax",&DWP(1024-128,$tbl,"ecx",4));	# rcon
2651}
2652
2653&function_begin("_x86_AES_set_encrypt_key");
2654	&mov	("esi",&wparam(1));		# user supplied key
2655	&mov	("edi",&wparam(3));		# private key schedule
2656
2657	&test	("esi",-1);
2658	&jz	(&label("badpointer"));
2659	&test	("edi",-1);
2660	&jz	(&label("badpointer"));
2661
2662	&call	(&label("pic_point"));
2663	&set_label("pic_point");
2664	&blindpop($tbl);
2665	&lea	($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
2666	&lea	($tbl,&DWP(2048+128,$tbl));
2667
2668	# prefetch Te4
2669	&mov	("eax",&DWP(0-128,$tbl));
2670	&mov	("ebx",&DWP(32-128,$tbl));
2671	&mov	("ecx",&DWP(64-128,$tbl));
2672	&mov	("edx",&DWP(96-128,$tbl));
2673	&mov	("eax",&DWP(128-128,$tbl));
2674	&mov	("ebx",&DWP(160-128,$tbl));
2675	&mov	("ecx",&DWP(192-128,$tbl));
2676	&mov	("edx",&DWP(224-128,$tbl));
2677
2678	&mov	("ecx",&wparam(2));		# number of bits in key
2679	&cmp	("ecx",128);
2680	&je	(&label("10rounds"));
2681	&cmp	("ecx",192);
2682	&je	(&label("12rounds"));
2683	&cmp	("ecx",256);
2684	&je	(&label("14rounds"));
2685	&mov	("eax",-2);			# invalid number of bits
2686	&jmp	(&label("exit"));
2687
2688    &set_label("10rounds");
2689	&mov	("eax",&DWP(0,"esi"));		# copy first 4 dwords
2690	&mov	("ebx",&DWP(4,"esi"));
2691	&mov	("ecx",&DWP(8,"esi"));
2692	&mov	("edx",&DWP(12,"esi"));
2693	&mov	(&DWP(0,"edi"),"eax");
2694	&mov	(&DWP(4,"edi"),"ebx");
2695	&mov	(&DWP(8,"edi"),"ecx");
2696	&mov	(&DWP(12,"edi"),"edx");
2697
2698	&xor	("ecx","ecx");
2699	&jmp	(&label("10shortcut"));
2700
2701	&align	(4);
2702	&set_label("10loop");
2703		&mov	("eax",&DWP(0,"edi"));		# rk[0]
2704		&mov	("edx",&DWP(12,"edi"));		# rk[3]
2705	&set_label("10shortcut");
2706		&enckey	();
2707
2708		&mov	(&DWP(16,"edi"),"eax");		# rk[4]
2709		&xor	("eax",&DWP(4,"edi"));
2710		&mov	(&DWP(20,"edi"),"eax");		# rk[5]
2711		&xor	("eax",&DWP(8,"edi"));
2712		&mov	(&DWP(24,"edi"),"eax");		# rk[6]
2713		&xor	("eax",&DWP(12,"edi"));
2714		&mov	(&DWP(28,"edi"),"eax");		# rk[7]
2715		&inc	("ecx");
2716		&add	("edi",16);
2717		&cmp	("ecx",10);
2718	&jl	(&label("10loop"));
2719
2720	&mov	(&DWP(80,"edi"),10);		# setup number of rounds
2721	&xor	("eax","eax");
2722	&jmp	(&label("exit"));
2723
2724    &set_label("12rounds");
2725	&mov	("eax",&DWP(0,"esi"));		# copy first 6 dwords
2726	&mov	("ebx",&DWP(4,"esi"));
2727	&mov	("ecx",&DWP(8,"esi"));
2728	&mov	("edx",&DWP(12,"esi"));
2729	&mov	(&DWP(0,"edi"),"eax");
2730	&mov	(&DWP(4,"edi"),"ebx");
2731	&mov	(&DWP(8,"edi"),"ecx");
2732	&mov	(&DWP(12,"edi"),"edx");
2733	&mov	("ecx",&DWP(16,"esi"));
2734	&mov	("edx",&DWP(20,"esi"));
2735	&mov	(&DWP(16,"edi"),"ecx");
2736	&mov	(&DWP(20,"edi"),"edx");
2737
2738	&xor	("ecx","ecx");
2739	&jmp	(&label("12shortcut"));
2740
2741	&align	(4);
2742	&set_label("12loop");
2743		&mov	("eax",&DWP(0,"edi"));		# rk[0]
2744		&mov	("edx",&DWP(20,"edi"));		# rk[5]
2745	&set_label("12shortcut");
2746		&enckey	();
2747
2748		&mov	(&DWP(24,"edi"),"eax");		# rk[6]
2749		&xor	("eax",&DWP(4,"edi"));
2750		&mov	(&DWP(28,"edi"),"eax");		# rk[7]
2751		&xor	("eax",&DWP(8,"edi"));
2752		&mov	(&DWP(32,"edi"),"eax");		# rk[8]
2753		&xor	("eax",&DWP(12,"edi"));
2754		&mov	(&DWP(36,"edi"),"eax");		# rk[9]
2755
2756		&cmp	("ecx",7);
2757		&je	(&label("12break"));
2758		&inc	("ecx");
2759
2760		&xor	("eax",&DWP(16,"edi"));
2761		&mov	(&DWP(40,"edi"),"eax");		# rk[10]
2762		&xor	("eax",&DWP(20,"edi"));
2763		&mov	(&DWP(44,"edi"),"eax");		# rk[11]
2764
2765		&add	("edi",24);
2766	&jmp	(&label("12loop"));
2767
2768	&set_label("12break");
2769	&mov	(&DWP(72,"edi"),12);		# setup number of rounds
2770	&xor	("eax","eax");
2771	&jmp	(&label("exit"));
2772
2773    &set_label("14rounds");
2774	&mov	("eax",&DWP(0,"esi"));		# copy first 8 dwords
2775	&mov	("ebx",&DWP(4,"esi"));
2776	&mov	("ecx",&DWP(8,"esi"));
2777	&mov	("edx",&DWP(12,"esi"));
2778	&mov	(&DWP(0,"edi"),"eax");
2779	&mov	(&DWP(4,"edi"),"ebx");
2780	&mov	(&DWP(8,"edi"),"ecx");
2781	&mov	(&DWP(12,"edi"),"edx");
2782	&mov	("eax",&DWP(16,"esi"));
2783	&mov	("ebx",&DWP(20,"esi"));
2784	&mov	("ecx",&DWP(24,"esi"));
2785	&mov	("edx",&DWP(28,"esi"));
2786	&mov	(&DWP(16,"edi"),"eax");
2787	&mov	(&DWP(20,"edi"),"ebx");
2788	&mov	(&DWP(24,"edi"),"ecx");
2789	&mov	(&DWP(28,"edi"),"edx");
2790
2791	&xor	("ecx","ecx");
2792	&jmp	(&label("14shortcut"));
2793
2794	&align	(4);
2795	&set_label("14loop");
2796		&mov	("edx",&DWP(28,"edi"));		# rk[7]
2797	&set_label("14shortcut");
2798		&mov	("eax",&DWP(0,"edi"));		# rk[0]
2799
2800		&enckey	();
2801
2802		&mov	(&DWP(32,"edi"),"eax");		# rk[8]
2803		&xor	("eax",&DWP(4,"edi"));
2804		&mov	(&DWP(36,"edi"),"eax");		# rk[9]
2805		&xor	("eax",&DWP(8,"edi"));
2806		&mov	(&DWP(40,"edi"),"eax");		# rk[10]
2807		&xor	("eax",&DWP(12,"edi"));
2808		&mov	(&DWP(44,"edi"),"eax");		# rk[11]
2809
2810		&cmp	("ecx",6);
2811		&je	(&label("14break"));
2812		&inc	("ecx");
2813
2814		&mov	("edx","eax");
2815		&mov	("eax",&DWP(16,"edi"));		# rk[4]
2816		&movz	("esi",&LB("edx"));		# rk[11]>>0
2817		&movz	("ebx",&BP(-128,$tbl,"esi",1));
2818		&movz	("esi",&HB("edx"));		# rk[11]>>8
2819		&xor	("eax","ebx");
2820
2821		&movz	("ebx",&BP(-128,$tbl,"esi",1));
2822		&shr	("edx",16);
2823		&shl	("ebx",8);
2824		&movz	("esi",&LB("edx"));		# rk[11]>>16
2825		&xor	("eax","ebx");
2826
2827		&movz	("ebx",&BP(-128,$tbl,"esi",1));
2828		&movz	("esi",&HB("edx"));		# rk[11]>>24
2829		&shl	("ebx",16);
2830		&xor	("eax","ebx");
2831
2832		&movz	("ebx",&BP(-128,$tbl,"esi",1));
2833		&shl	("ebx",24);
2834		&xor	("eax","ebx");
2835
2836		&mov	(&DWP(48,"edi"),"eax");		# rk[12]
2837		&xor	("eax",&DWP(20,"edi"));
2838		&mov	(&DWP(52,"edi"),"eax");		# rk[13]
2839		&xor	("eax",&DWP(24,"edi"));
2840		&mov	(&DWP(56,"edi"),"eax");		# rk[14]
2841		&xor	("eax",&DWP(28,"edi"));
2842		&mov	(&DWP(60,"edi"),"eax");		# rk[15]
2843
2844		&add	("edi",32);
2845	&jmp	(&label("14loop"));
2846
2847	&set_label("14break");
2848	&mov	(&DWP(48,"edi"),14);		# setup number of rounds
2849	&xor	("eax","eax");
2850	&jmp	(&label("exit"));
2851
2852    &set_label("badpointer");
2853	&mov	("eax",-1);
2854    &set_label("exit");
2855&function_end("_x86_AES_set_encrypt_key");
2856
2857# int private_AES_set_encrypt_key(const unsigned char *userKey, const int bits,
2858#                        AES_KEY *key)
2859&function_begin_B("private_AES_set_encrypt_key");
2860	&call	("_x86_AES_set_encrypt_key");
2861	&ret	();
2862&function_end_B("private_AES_set_encrypt_key");
2863
2864sub deckey()
2865{ my ($i,$key,$tp1,$tp2,$tp4,$tp8) = @_;
2866  my $tmp = $tbl;
2867
2868	&mov	($acc,$tp1);
2869	&and	($acc,0x80808080);
2870	&mov	($tmp,$acc);
2871	&shr	($tmp,7);
2872	&lea	($tp2,&DWP(0,$tp1,$tp1));
2873	&sub	($acc,$tmp);
2874	&and	($tp2,0xfefefefe);
2875	&and	($acc,0x1b1b1b1b);
2876	&xor	($acc,$tp2);
2877	&mov	($tp2,$acc);
2878
2879	&and	($acc,0x80808080);
2880	&mov	($tmp,$acc);
2881	&shr	($tmp,7);
2882	&lea	($tp4,&DWP(0,$tp2,$tp2));
2883	&sub	($acc,$tmp);
2884	&and	($tp4,0xfefefefe);
2885	&and	($acc,0x1b1b1b1b);
2886	 &xor	($tp2,$tp1);	# tp2^tp1
2887	&xor	($acc,$tp4);
2888	&mov	($tp4,$acc);
2889
2890	&and	($acc,0x80808080);
2891	&mov	($tmp,$acc);
2892	&shr	($tmp,7);
2893	&lea	($tp8,&DWP(0,$tp4,$tp4));
2894	 &xor	($tp4,$tp1);	# tp4^tp1
2895	&sub	($acc,$tmp);
2896	&and	($tp8,0xfefefefe);
2897	&and	($acc,0x1b1b1b1b);
2898	 &rotl	($tp1,8);	# = ROTATE(tp1,8)
2899	&xor	($tp8,$acc);
2900
2901	&mov	($tmp,&DWP(4*($i+1),$key));	# modulo-scheduled load
2902
2903	&xor	($tp1,$tp2);
2904	&xor	($tp2,$tp8);
2905	&xor	($tp1,$tp4);
2906	&rotl	($tp2,24);
2907	&xor	($tp4,$tp8);
2908	&xor	($tp1,$tp8);	# ^= tp8^(tp4^tp1)^(tp2^tp1)
2909	&rotl	($tp4,16);
2910	&xor	($tp1,$tp2);	# ^= ROTATE(tp8^tp2^tp1,24)
2911	&rotl	($tp8,8);
2912	&xor	($tp1,$tp4);	# ^= ROTATE(tp8^tp4^tp1,16)
2913	&mov	($tp2,$tmp);
2914	&xor	($tp1,$tp8);	# ^= ROTATE(tp8,8)
2915
2916	&mov	(&DWP(4*$i,$key),$tp1);
2917}
2918
2919# int private_AES_set_decrypt_key(const unsigned char *userKey, const int bits,
2920#                        AES_KEY *key)
2921&function_begin_B("private_AES_set_decrypt_key");
2922	&call	("_x86_AES_set_encrypt_key");
2923	&cmp	("eax",0);
2924	&je	(&label("proceed"));
2925	&ret	();
2926
2927    &set_label("proceed");
2928	&push	("ebp");
2929	&push	("ebx");
2930	&push	("esi");
2931	&push	("edi");
2932
2933	&mov	("esi",&wparam(2));
2934	&mov	("ecx",&DWP(240,"esi"));	# pull number of rounds
2935	&lea	("ecx",&DWP(0,"","ecx",4));
2936	&lea	("edi",&DWP(0,"esi","ecx",4));	# pointer to last chunk
2937
2938	&set_label("invert",4);			# invert order of chunks
2939		&mov	("eax",&DWP(0,"esi"));
2940		&mov	("ebx",&DWP(4,"esi"));
2941		&mov	("ecx",&DWP(0,"edi"));
2942		&mov	("edx",&DWP(4,"edi"));
2943		&mov	(&DWP(0,"edi"),"eax");
2944		&mov	(&DWP(4,"edi"),"ebx");
2945		&mov	(&DWP(0,"esi"),"ecx");
2946		&mov	(&DWP(4,"esi"),"edx");
2947		&mov	("eax",&DWP(8,"esi"));
2948		&mov	("ebx",&DWP(12,"esi"));
2949		&mov	("ecx",&DWP(8,"edi"));
2950		&mov	("edx",&DWP(12,"edi"));
2951		&mov	(&DWP(8,"edi"),"eax");
2952		&mov	(&DWP(12,"edi"),"ebx");
2953		&mov	(&DWP(8,"esi"),"ecx");
2954		&mov	(&DWP(12,"esi"),"edx");
2955		&add	("esi",16);
2956		&sub	("edi",16);
2957		&cmp	("esi","edi");
2958	&jne	(&label("invert"));
2959
2960	&mov	($key,&wparam(2));
2961	&mov	($acc,&DWP(240,$key));		# pull number of rounds
2962	&lea	($acc,&DWP(-2,$acc,$acc));
2963	&lea	($acc,&DWP(0,$key,$acc,8));
2964	&mov	(&wparam(2),$acc);
2965
2966	&mov	($s0,&DWP(16,$key));		# modulo-scheduled load
2967	&set_label("permute",4);		# permute the key schedule
2968		&add	($key,16);
2969		&deckey	(0,$key,$s0,$s1,$s2,$s3);
2970		&deckey	(1,$key,$s1,$s2,$s3,$s0);
2971		&deckey	(2,$key,$s2,$s3,$s0,$s1);
2972		&deckey	(3,$key,$s3,$s0,$s1,$s2);
2973		&cmp	($key,&wparam(2));
2974	&jb	(&label("permute"));
2975
2976	&xor	("eax","eax");			# return success
2977&function_end("private_AES_set_decrypt_key");
2978&asciz("AES for x86, CRYPTOGAMS by <appro\@openssl.org>");
2979
2980&asm_finish();
2981